2020-07-21 13:20:51 +00:00
|
|
|
/**
|
|
|
|
* JS controller for the pypy sandbox.
|
|
|
|
*/
|
2022-05-18 16:05:37 +00:00
|
|
|
import {arrayToString} from 'app/common/arrayToString';
|
2020-07-21 13:20:51 +00:00
|
|
|
import * as marshal from 'app/common/marshal';
|
|
|
|
import {ISandbox, ISandboxCreationOptions, ISandboxCreator} from 'app/server/lib/ISandbox';
|
2022-07-04 14:14:55 +00:00
|
|
|
import log from 'app/server/lib/log';
|
2023-02-13 20:52:17 +00:00
|
|
|
import {getAppRoot, getAppRootFor, getUnpackedAppRoot} from 'app/server/lib/places';
|
2022-08-11 11:03:45 +00:00
|
|
|
import {
|
|
|
|
DirectProcessControl,
|
|
|
|
ISandboxControl,
|
|
|
|
NoProcessControl,
|
|
|
|
ProcessInfo,
|
|
|
|
SubprocessControl
|
|
|
|
} from 'app/server/lib/SandboxControl';
|
2020-07-21 13:20:51 +00:00
|
|
|
import * as sandboxUtil from 'app/server/lib/sandboxUtil';
|
|
|
|
import * as shutdown from 'app/server/lib/shutdown';
|
2023-03-06 21:56:25 +00:00
|
|
|
import {ChildProcess, fork, spawn} from 'child_process';
|
2022-08-11 11:03:45 +00:00
|
|
|
import * as fs from 'fs';
|
|
|
|
import * as _ from 'lodash';
|
2020-07-21 13:20:51 +00:00
|
|
|
import * as path from 'path';
|
|
|
|
import {Stream, Writable} from 'stream';
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
import * as which from 'which';
|
2020-07-21 13:20:51 +00:00
|
|
|
|
|
|
|
type SandboxMethod = (...args: any[]) => any;
|
|
|
|
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
/**
|
|
|
|
*
|
|
|
|
* A collection of options for weird and wonderful ways to run Grist.
|
|
|
|
* The sandbox at heart is just python, but run in different ways
|
|
|
|
* (sandbox 'flavors': pynbox, docker, gvisor, and unsandboxed).
|
|
|
|
*
|
|
|
|
* The "command" is an external program/container to call to run the
|
|
|
|
* sandbox, and it depends on sandbox flavor. Pynbox is built into
|
|
|
|
* Grist and has a hard-wired command, so the command option should be
|
|
|
|
* empty. For gvisor and unsandboxed, command is the path to an
|
|
|
|
* external program to run. For docker, it is the name of an image.
|
|
|
|
*
|
|
|
|
* Once python is running, ordinarily some Grist code should be
|
|
|
|
* started by setting `useGristEntrypoint` (the only exception is
|
2021-08-09 14:51:43 +00:00
|
|
|
* in tests) which runs grist/main.py.
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
*/
|
2022-02-04 11:13:03 +00:00
|
|
|
export interface ISandboxOptions {
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
command?: string; // External program or container to call to run the sandbox.
|
2020-07-21 13:20:51 +00:00
|
|
|
args: string[]; // The arguments to pass to the python process.
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
|
2021-08-26 03:12:34 +00:00
|
|
|
preferredPythonVersion?: string; // Mandatory for gvisor; ignored by other methods.
|
|
|
|
|
2021-08-09 14:51:43 +00:00
|
|
|
// TODO: update
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
// ISandboxCreationOptions to talk about directories instead of
|
|
|
|
// mounts, since it may not be possible to remap directories as
|
|
|
|
// mounts (e.g. for unsandboxed operation).
|
2021-08-09 14:51:43 +00:00
|
|
|
importDir?: string; // a directory containing data file(s) to import by plugins
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
|
|
|
|
minimalPipeMode?: boolean; // Whether to use newer 3-pipe operation
|
|
|
|
deterministicMode?: boolean; // Whether to override time + randomness
|
|
|
|
|
2020-07-21 13:20:51 +00:00
|
|
|
exports?: {[name: string]: SandboxMethod}; // Functions made available to the sandboxed process.
|
|
|
|
logCalls?: boolean; // (Not implemented) Whether to log all system calls from the python sandbox.
|
|
|
|
logTimes?: boolean; // Whether to log time taken by calls to python sandbox.
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
unsilenceLog?: boolean; // Don't silence the sel_ldr logging (pynbox only).
|
2020-07-21 13:20:51 +00:00
|
|
|
logMeta?: log.ILogMeta; // Log metadata (e.g. including docId) to report in all log messages.
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
|
|
|
|
useGristEntrypoint?: boolean; // Should be set for everything except tests, which
|
|
|
|
// may want to pass arguments to python directly.
|
2020-07-21 13:20:51 +00:00
|
|
|
}
|
|
|
|
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
/**
|
|
|
|
* We interact with sandboxes as a separate child process. Data engine work is done
|
|
|
|
* across standard input and output streams from and to this process. We also monitor
|
|
|
|
* and control resource utilization via a distinct control interface.
|
|
|
|
*/
|
|
|
|
interface SandboxProcess {
|
|
|
|
child: ChildProcess;
|
|
|
|
control: ISandboxControl;
|
2023-03-06 21:56:25 +00:00
|
|
|
dataToSandboxDescriptor?: number; // override sandbox's 'stdin' for data
|
|
|
|
dataFromSandboxDescriptor?: number; // override sandbox's 'stdout' for data
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
}
|
|
|
|
|
2020-07-21 13:20:51 +00:00
|
|
|
type ResolveRejectPair = [(value?: any) => void, (reason?: unknown) => void];
|
|
|
|
|
|
|
|
// Type for basic message identifiers, available as constants in sandboxUtil.
|
|
|
|
type MsgCode = null | true | false;
|
|
|
|
|
2021-06-30 13:35:05 +00:00
|
|
|
// Optional root folder to store binary data sent to and from the sandbox
|
|
|
|
// See test_replay.py
|
|
|
|
const recordBuffersRoot = process.env.RECORD_SANDBOX_BUFFERS_DIR;
|
|
|
|
|
2020-07-21 13:20:51 +00:00
|
|
|
export class NSandbox implements ISandbox {
|
2021-07-08 21:52:00 +00:00
|
|
|
|
2020-07-21 13:20:51 +00:00
|
|
|
public readonly childProc: ChildProcess;
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
private _control: ISandboxControl;
|
2020-07-21 13:20:51 +00:00
|
|
|
private _logTimes: boolean;
|
|
|
|
private _exportedFunctions: {[name: string]: SandboxMethod};
|
2021-06-24 12:23:33 +00:00
|
|
|
private _marshaller = new marshal.Marshaller({stringToBuffer: false, version: 2});
|
2020-07-21 13:20:51 +00:00
|
|
|
private _unmarshaller = new marshal.Unmarshaller({ bufferToString: false });
|
|
|
|
|
|
|
|
// Members used for reading from the sandbox process.
|
|
|
|
private _pendingReads: ResolveRejectPair[] = [];
|
|
|
|
private _isReadClosed = false;
|
|
|
|
private _isWriteClosed = false;
|
|
|
|
|
|
|
|
private _logMeta: log.ILogMeta;
|
|
|
|
private _streamToSandbox: Writable;
|
|
|
|
private _streamFromSandbox: Stream;
|
2022-05-18 16:05:37 +00:00
|
|
|
private _lastStderr: Uint8Array; // Record last error line seen.
|
2020-07-21 13:20:51 +00:00
|
|
|
|
2021-06-30 13:35:05 +00:00
|
|
|
// Create a unique subdirectory for each sandbox process so they can be replayed separately
|
|
|
|
private _recordBuffersDir = recordBuffersRoot ? path.resolve(recordBuffersRoot, new Date().toISOString()) : null;
|
|
|
|
|
2020-07-21 13:20:51 +00:00
|
|
|
/*
|
|
|
|
* Callers may listen to events from sandbox.childProc (a ChildProcess), e.g. 'close' and 'error'.
|
|
|
|
* The sandbox listens for 'aboutToExit' event on the process, to properly shut down.
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
*
|
|
|
|
* Grist interacts with the sandbox via message passing through pipes to an isolated
|
|
|
|
* process. Some read-only shared code is made available to the sandbox.
|
|
|
|
* For plugins, read-only data files are made available.
|
|
|
|
*
|
|
|
|
* At the time of writing, Grist has been using an NaCl sandbox with python2.7 compiled
|
|
|
|
* for it for several years (pynbox), and we are now experimenting with other sandboxing
|
|
|
|
* options. Variants can be activated by passing in a non-default "spawner" function.
|
|
|
|
*
|
2020-07-21 13:20:51 +00:00
|
|
|
*/
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
constructor(options: ISandboxOptions, spawner: SpawnFn = pynbox) {
|
2020-07-21 13:20:51 +00:00
|
|
|
this._logTimes = Boolean(options.logTimes || options.logCalls);
|
|
|
|
this._exportedFunctions = options.exports || {};
|
|
|
|
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
const sandboxProcess = spawner(options);
|
|
|
|
this._control = sandboxProcess.control;
|
|
|
|
this.childProc = sandboxProcess.child;
|
2020-07-21 13:20:51 +00:00
|
|
|
|
|
|
|
this._logMeta = {sandboxPid: this.childProc.pid, ...options.logMeta};
|
|
|
|
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
if (options.minimalPipeMode) {
|
2021-07-08 21:52:00 +00:00
|
|
|
log.rawDebug("3-pipe Sandbox started", this._logMeta);
|
2023-03-06 21:56:25 +00:00
|
|
|
if (sandboxProcess.dataToSandboxDescriptor) {
|
|
|
|
this._streamToSandbox =
|
|
|
|
(this.childProc.stdio as Stream[])[sandboxProcess.dataToSandboxDescriptor] as Writable;
|
|
|
|
} else {
|
|
|
|
this._streamToSandbox = this.childProc.stdin!;
|
|
|
|
}
|
|
|
|
if (sandboxProcess.dataFromSandboxDescriptor) {
|
|
|
|
this._streamFromSandbox =
|
|
|
|
(this.childProc.stdio as Stream[])[sandboxProcess.dataFromSandboxDescriptor];
|
|
|
|
} else {
|
|
|
|
this._streamFromSandbox = this.childProc.stdout!;
|
|
|
|
}
|
2021-07-08 21:52:00 +00:00
|
|
|
} else {
|
|
|
|
log.rawDebug("5-pipe Sandbox started", this._logMeta);
|
2023-03-06 21:56:25 +00:00
|
|
|
if (sandboxProcess.dataFromSandboxDescriptor || sandboxProcess.dataToSandboxDescriptor) {
|
|
|
|
throw new Error('cannot override file descriptors in 5 pipe mode');
|
|
|
|
}
|
2021-07-08 21:52:00 +00:00
|
|
|
this._streamToSandbox = (this.childProc.stdio as Stream[])[3] as Writable;
|
|
|
|
this._streamFromSandbox = (this.childProc.stdio as Stream[])[4];
|
(core) Speed up and upgrade build.
Summary:
- Upgrades to build-related packages:
- Upgrade typescript, related libraries and typings.
- Upgrade webpack, eslint; add tsc-watch, node-dev, eslint_d.
- Build organization changes:
- Build webpack from original typescript, transpiling only; with errors still
reported by a background tsc watching process.
- Typescript-related changes:
- Reduce imports of AWS dependencies (very noticeable speedup)
- Avoid auto-loading global @types
- Client code is now built with isolatedModules flag (for safe transpilation)
- Use allowJs to avoid copying JS files manually.
- Linting changes
- Enhance Arcanist ESLintLinter to run before/after commands, and set up to use eslint_d
- Update eslint config, and include .eslintignore to avoid linting generated files.
- Include a bunch of eslint-prompted and eslint-generated fixes
- Add no-unused-expression rule to eslint, and fix a few warnings about it
- Other items:
- Refactor cssInput to avoid circular dependency
- Remove a bit of unused code, libraries, dependencies
Test Plan: No behavior changes, all existing tests pass. There are 30 tests fewer reported because `test_gpath.py` was removed (it's been unused for years)
Reviewers: paulfitz
Reviewed By: paulfitz
Subscribers: paulfitz
Differential Revision: https://phab.getgrist.com/D3498
2022-06-27 20:09:41 +00:00
|
|
|
this.childProc.stdout!.on('data', sandboxUtil.makeLinePrefixer('Sandbox stdout: ', this._logMeta));
|
2021-07-08 21:52:00 +00:00
|
|
|
}
|
2022-05-18 16:05:37 +00:00
|
|
|
const sandboxStderrLogger = sandboxUtil.makeLinePrefixer('Sandbox stderr: ', this._logMeta);
|
(core) Speed up and upgrade build.
Summary:
- Upgrades to build-related packages:
- Upgrade typescript, related libraries and typings.
- Upgrade webpack, eslint; add tsc-watch, node-dev, eslint_d.
- Build organization changes:
- Build webpack from original typescript, transpiling only; with errors still
reported by a background tsc watching process.
- Typescript-related changes:
- Reduce imports of AWS dependencies (very noticeable speedup)
- Avoid auto-loading global @types
- Client code is now built with isolatedModules flag (for safe transpilation)
- Use allowJs to avoid copying JS files manually.
- Linting changes
- Enhance Arcanist ESLintLinter to run before/after commands, and set up to use eslint_d
- Update eslint config, and include .eslintignore to avoid linting generated files.
- Include a bunch of eslint-prompted and eslint-generated fixes
- Add no-unused-expression rule to eslint, and fix a few warnings about it
- Other items:
- Refactor cssInput to avoid circular dependency
- Remove a bit of unused code, libraries, dependencies
Test Plan: No behavior changes, all existing tests pass. There are 30 tests fewer reported because `test_gpath.py` was removed (it's been unused for years)
Reviewers: paulfitz
Reviewed By: paulfitz
Subscribers: paulfitz
Differential Revision: https://phab.getgrist.com/D3498
2022-06-27 20:09:41 +00:00
|
|
|
this.childProc.stderr!.on('data', data => {
|
2022-05-18 16:05:37 +00:00
|
|
|
this._lastStderr = data;
|
|
|
|
sandboxStderrLogger(data);
|
|
|
|
});
|
2020-07-21 13:20:51 +00:00
|
|
|
|
|
|
|
this.childProc.on('close', this._onExit.bind(this));
|
|
|
|
this.childProc.on('error', this._onError.bind(this));
|
|
|
|
|
|
|
|
this._streamFromSandbox.on('data', (data) => this._onSandboxData(data));
|
|
|
|
this._streamFromSandbox.on('end', () => this._onSandboxClose());
|
|
|
|
this._streamFromSandbox.on('error', (err) => {
|
|
|
|
log.rawError(`Sandbox error reading: ${err}`, this._logMeta);
|
|
|
|
this._onSandboxClose();
|
|
|
|
});
|
|
|
|
|
|
|
|
this._streamToSandbox.on('error', (err) => {
|
|
|
|
if (!this._isWriteClosed) {
|
|
|
|
log.rawError(`Sandbox error writing: ${err}`, this._logMeta);
|
|
|
|
}
|
|
|
|
});
|
|
|
|
|
|
|
|
// On shutdown, shutdown the child process cleanly, and wait for it to exit.
|
|
|
|
shutdown.addCleanupHandler(this, this.shutdown);
|
|
|
|
|
2021-06-30 13:35:05 +00:00
|
|
|
if (this._recordBuffersDir) {
|
|
|
|
log.rawDebug(`Recording sandbox buffers in ${this._recordBuffersDir}`, this._logMeta);
|
|
|
|
fs.mkdirSync(this._recordBuffersDir, {recursive: true});
|
|
|
|
}
|
2020-07-21 13:20:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Shuts down the sandbox process cleanly, and wait for it to exit.
|
|
|
|
* @return {Promise} Promise that's resolved with [code, signal] when the sandbox exits.
|
|
|
|
*/
|
|
|
|
public async shutdown() {
|
|
|
|
log.rawDebug("Sandbox shutdown starting", this._logMeta);
|
|
|
|
shutdown.removeCleanupHandlers(this);
|
|
|
|
|
|
|
|
// The signal ensures the sandbox process exits even if it's hanging in an infinite loop or
|
|
|
|
// long computation. It doesn't get a chance to clean up, but since it is sandboxed, there is
|
|
|
|
// nothing it needs to clean up anyway.
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
const timeoutID = setTimeout(async () => {
|
2020-07-21 13:20:51 +00:00
|
|
|
log.rawWarn("Sandbox sending SIGKILL", this._logMeta);
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
await this._control.kill();
|
2020-07-21 13:20:51 +00:00
|
|
|
}, 1000);
|
|
|
|
|
2022-03-24 20:27:34 +00:00
|
|
|
const result = await new Promise<void>((resolve, reject) => {
|
2020-07-21 13:20:51 +00:00
|
|
|
if (this._isWriteClosed) { resolve(); }
|
|
|
|
this.childProc.on('error', reject);
|
|
|
|
this.childProc.on('close', resolve);
|
2021-05-13 02:06:19 +00:00
|
|
|
this.childProc.on('exit', resolve);
|
2020-07-21 13:20:51 +00:00
|
|
|
this._close();
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
}).finally(() => this._control.close());
|
2020-07-21 13:20:51 +00:00
|
|
|
|
|
|
|
// In the normal case, the kill timer is pending when the process exits, and we can clear it. If
|
|
|
|
// the process got killed, the timer is invalid, and clearTimeout() does nothing.
|
|
|
|
clearTimeout(timeoutID);
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Makes a call to the python process implementing our calling convention on stdin/stdout.
|
|
|
|
* @param funcName The name of the python RPC function to call.
|
|
|
|
* @param args Arguments to pass to the given function.
|
|
|
|
* @returns A promise for the return value from the Python function.
|
|
|
|
*/
|
2023-03-06 21:56:25 +00:00
|
|
|
public async pyCall(funcName: string, ...varArgs: unknown[]): Promise<any> {
|
2020-07-21 13:20:51 +00:00
|
|
|
const startTime = Date.now();
|
|
|
|
this._sendData(sandboxUtil.CALL, Array.from(arguments));
|
2023-03-06 21:56:25 +00:00
|
|
|
const slowCallCheck = setTimeout(() => {
|
|
|
|
// Log calls that take some time, can be a useful symptom of misconfiguration
|
|
|
|
// (or just benign if the doc is big).
|
|
|
|
log.rawWarn('Slow pyCall', {...this._logMeta, funcName});
|
|
|
|
}, 10000);
|
|
|
|
const result = await this._pyCallWait(funcName, startTime);
|
|
|
|
clearTimeout(slowCallCheck);
|
|
|
|
return result;
|
2020-07-21 13:20:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Returns the RSS (resident set size) of the sandbox process, in bytes.
|
|
|
|
*/
|
|
|
|
public async reportMemoryUsage() {
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
const {memory} = await this._control.getUsage();
|
2020-07-21 13:20:51 +00:00
|
|
|
log.rawDebug('Sandbox memory', {memory, ...this._logMeta});
|
|
|
|
}
|
|
|
|
|
|
|
|
private async _pyCallWait(funcName: string, startTime: number): Promise<any> {
|
|
|
|
try {
|
|
|
|
return await new Promise((resolve, reject) => {
|
|
|
|
this._pendingReads.push([resolve, reject]);
|
|
|
|
});
|
2022-03-18 19:59:01 +00:00
|
|
|
} catch (e) {
|
|
|
|
throw new sandboxUtil.SandboxError(e.message);
|
2020-07-21 13:20:51 +00:00
|
|
|
} finally {
|
|
|
|
if (this._logTimes) {
|
|
|
|
log.rawDebug(`Sandbox pyCall[${funcName}] took ${Date.now() - startTime} ms`, this._logMeta);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
private _close() {
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
this._control.prepareToClose();
|
2020-07-21 13:20:51 +00:00
|
|
|
if (!this._isWriteClosed) {
|
|
|
|
// Close the pipe to the sandbox, which should cause the sandbox to exit cleanly.
|
|
|
|
this._streamToSandbox.end();
|
|
|
|
this._isWriteClosed = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
private _onExit(code: number, signal: string) {
|
2022-04-11 13:52:34 +00:00
|
|
|
const expected = this._isWriteClosed;
|
2020-07-21 13:20:51 +00:00
|
|
|
this._close();
|
2022-04-11 13:52:34 +00:00
|
|
|
if (expected) {
|
|
|
|
log.rawDebug(`Sandbox exited with code ${code} signal ${signal}`, this._logMeta);
|
|
|
|
} else {
|
|
|
|
log.rawWarn(`Sandbox unexpectedly exited with code ${code} signal ${signal}`, this._logMeta);
|
|
|
|
}
|
2020-07-21 13:20:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
private _onError(err: Error) {
|
|
|
|
this._close();
|
|
|
|
log.rawWarn(`Sandbox could not be spawned: ${err}`, this._logMeta);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Send a message to the sandbox process with the given message code and data.
|
|
|
|
*/
|
|
|
|
private _sendData(msgCode: MsgCode, data: any) {
|
|
|
|
if (this._isReadClosed) {
|
2022-05-18 16:05:37 +00:00
|
|
|
throw this._sandboxClosedError('PipeToSandbox');
|
2020-07-21 13:20:51 +00:00
|
|
|
}
|
|
|
|
this._marshaller.marshal(msgCode);
|
|
|
|
this._marshaller.marshal(data);
|
2021-06-30 13:35:05 +00:00
|
|
|
const buf = this._marshaller.dumpAsBuffer();
|
|
|
|
if (this._recordBuffersDir) {
|
|
|
|
fs.appendFileSync(path.resolve(this._recordBuffersDir, "input"), buf);
|
|
|
|
}
|
|
|
|
return this._streamToSandbox.write(buf);
|
2020-07-21 13:20:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Process a buffer of data received from the sandbox process.
|
|
|
|
*/
|
|
|
|
private _onSandboxData(data: any) {
|
|
|
|
this._unmarshaller.parse(data, buf => {
|
|
|
|
const value = marshal.loads(buf, { bufferToString: true });
|
2021-06-30 13:35:05 +00:00
|
|
|
if (this._recordBuffersDir) {
|
|
|
|
fs.appendFileSync(path.resolve(this._recordBuffersDir, "output"), buf);
|
|
|
|
}
|
2020-07-21 13:20:51 +00:00
|
|
|
this._onSandboxMsg(value[0], value[1]);
|
|
|
|
});
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Process the closing of the pipe by the sandboxed process.
|
|
|
|
*/
|
|
|
|
private _onSandboxClose() {
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
this._control.prepareToClose();
|
2020-07-21 13:20:51 +00:00
|
|
|
this._isReadClosed = true;
|
|
|
|
// Clear out all reads pending on PipeFromSandbox, rejecting them with the given error.
|
2022-05-18 16:05:37 +00:00
|
|
|
const err = this._sandboxClosedError('PipeFromSandbox');
|
2022-03-24 20:27:34 +00:00
|
|
|
|
2020-07-21 13:20:51 +00:00
|
|
|
this._pendingReads.forEach(resolvePair => resolvePair[1](err));
|
|
|
|
this._pendingReads = [];
|
|
|
|
}
|
|
|
|
|
2022-05-18 16:05:37 +00:00
|
|
|
/**
|
|
|
|
* Generate an error message for a pipe to the sandbox. Include the
|
|
|
|
* last stderr line seen from the sandbox - more reliable than
|
|
|
|
* error results send via the standard protocol.
|
|
|
|
*/
|
|
|
|
private _sandboxClosedError(label: string) {
|
|
|
|
const parts = [`${label} is closed`];
|
|
|
|
if (this._lastStderr) {
|
|
|
|
parts.push(arrayToString(this._lastStderr));
|
|
|
|
}
|
|
|
|
return new sandboxUtil.SandboxError(parts.join(': '));
|
|
|
|
}
|
2020-07-21 13:20:51 +00:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Process a parsed message from the sandboxed process.
|
|
|
|
*/
|
|
|
|
private _onSandboxMsg(msgCode: MsgCode, data: any) {
|
|
|
|
if (msgCode === sandboxUtil.CALL) {
|
|
|
|
// Handle calls FROM the sandbox.
|
|
|
|
if (!Array.isArray(data) || data.length === 0) {
|
|
|
|
log.rawWarn("Sandbox invalid call from the sandbox", this._logMeta);
|
|
|
|
} else {
|
|
|
|
const fname = data[0];
|
|
|
|
const args = data.slice(1);
|
|
|
|
log.rawDebug(`Sandbox got call to ${fname} (${args.length} args)`, this._logMeta);
|
|
|
|
Promise.resolve()
|
|
|
|
.then(() => {
|
|
|
|
const func = this._exportedFunctions[fname];
|
|
|
|
if (!func) { throw new Error("No such exported function: " + fname); }
|
|
|
|
return func(...args);
|
|
|
|
})
|
|
|
|
.then((ret) => {
|
|
|
|
this._sendData(sandboxUtil.DATA, ret);
|
|
|
|
}, (err) => {
|
|
|
|
this._sendData(sandboxUtil.EXC, err.toString());
|
|
|
|
})
|
|
|
|
.catch((err) => {
|
|
|
|
log.rawDebug(`Sandbox sending response failed: ${err}`, this._logMeta);
|
|
|
|
});
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// Handle return values for calls made to the sandbox.
|
|
|
|
const resolvePair = this._pendingReads.shift();
|
|
|
|
if (resolvePair) {
|
|
|
|
if (msgCode === sandboxUtil.EXC) {
|
2022-03-18 19:59:01 +00:00
|
|
|
resolvePair[1](new Error(data));
|
2020-07-21 13:20:51 +00:00
|
|
|
} else if (msgCode === sandboxUtil.DATA) {
|
|
|
|
resolvePair[0](data);
|
|
|
|
} else {
|
|
|
|
log.rawWarn("Sandbox invalid message from sandbox", this._logMeta);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
/**
|
|
|
|
* Functions for spawning all of the currently supported sandboxes.
|
|
|
|
*/
|
|
|
|
const spawners = {
|
|
|
|
pynbox, // Grist's "classic" sandbox - python2 within NaCl.
|
|
|
|
unsandboxed, // No sandboxing, straight to host python.
|
|
|
|
// This offers no protection to the host.
|
|
|
|
docker, // Run sandboxes in distinct docker containers.
|
|
|
|
gvisor, // Gvisor's runsc sandbox.
|
2021-10-07 17:49:23 +00:00
|
|
|
macSandboxExec, // Use "sandbox-exec" on Mac.
|
2023-03-06 21:56:25 +00:00
|
|
|
pyodide, // Run data engine using pyodide.
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
};
|
|
|
|
|
2022-03-24 20:27:34 +00:00
|
|
|
function isFlavor(flavor: string): flavor is keyof typeof spawners {
|
|
|
|
return flavor in spawners;
|
|
|
|
}
|
|
|
|
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
/**
|
|
|
|
* A sandbox factory. This doesn't do very much beyond remembering a default
|
|
|
|
* flavor of sandbox (which at the time of writing differs between hosted grist and
|
|
|
|
* grist-core), and trying to regularize creation options a bit.
|
|
|
|
*
|
2021-08-03 21:21:11 +00:00
|
|
|
* The flavor of sandbox to use can be overridden by some environment variables:
|
2022-09-28 21:29:32 +00:00
|
|
|
* - GRIST_SANDBOX_FLAVOR: should be one of the spawners (gvisor, unsandboxed, docker,
|
|
|
|
* macSandboxExec, pynbox)
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
* - GRIST_SANDBOX: a program or image name to run as the sandbox. Not needed for
|
2022-02-19 09:46:49 +00:00
|
|
|
* pynbox (it is either built in or not available). For unsandboxed, should be an
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
* absolute path to python within a virtualenv with all requirements installed.
|
|
|
|
* For docker, it should be `grist-docker-sandbox` (an image built via makefile
|
|
|
|
* in `sandbox/docker`) or a derived image. For gvisor, it should be the full path
|
|
|
|
* to `sandbox/gvisor/run.py` (if runsc available locally) or to
|
|
|
|
* `sandbox/gvisor/wrap_in_docker.sh` (if runsc should be run using the docker
|
2022-09-28 21:29:32 +00:00
|
|
|
* image built in that directory).
|
2021-08-03 21:21:11 +00:00
|
|
|
* - PYTHON_VERSION: for gvisor, this is mandatory, and must be set to "2" or "3".
|
|
|
|
* It is ignored by other flavors.
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
*/
|
2020-07-21 13:20:51 +00:00
|
|
|
export class NSandboxCreator implements ISandboxCreator {
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
private _flavor: keyof typeof spawners;
|
|
|
|
private _command?: string;
|
2021-08-26 03:12:34 +00:00
|
|
|
private _preferredPythonVersion?: string;
|
|
|
|
|
|
|
|
public constructor(options: {
|
|
|
|
defaultFlavor: keyof typeof spawners,
|
|
|
|
command?: string,
|
|
|
|
preferredPythonVersion?: string,
|
|
|
|
}) {
|
2022-03-24 20:27:34 +00:00
|
|
|
const flavor = options.defaultFlavor;
|
|
|
|
if (!isFlavor(flavor)) {
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
throw new Error(`Unrecognized sandbox flavor: ${flavor}`);
|
|
|
|
}
|
2022-03-24 20:27:34 +00:00
|
|
|
this._flavor = flavor;
|
|
|
|
this._command = options.command;
|
|
|
|
this._preferredPythonVersion = options.preferredPythonVersion;
|
2020-07-21 13:20:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
public create(options: ISandboxCreationOptions): ISandbox {
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
const args: string[] = [];
|
2020-07-21 13:20:51 +00:00
|
|
|
if (!options.entryPoint && options.comment) {
|
|
|
|
// When using default entry point, we can add on a comment as an argument - it isn't
|
|
|
|
// used, but will show up in `ps` output for the sandbox process. Comment is intended
|
|
|
|
// to be a document name/id.
|
|
|
|
args.push(options.comment);
|
|
|
|
}
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
const translatedOptions: ISandboxOptions = {
|
|
|
|
minimalPipeMode: true,
|
|
|
|
deterministicMode: Boolean(process.env.LIBFAKETIME_PATH),
|
2020-07-21 13:20:51 +00:00
|
|
|
args,
|
|
|
|
logCalls: options.logCalls,
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
logMeta: {flavor: this._flavor, command: this._command,
|
|
|
|
entryPoint: options.entryPoint || '(default)',
|
|
|
|
...options.logMeta},
|
2020-07-21 13:20:51 +00:00
|
|
|
logTimes: options.logTimes,
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
command: this._command,
|
2022-01-07 17:06:04 +00:00
|
|
|
preferredPythonVersion: this._preferredPythonVersion || options.preferredPythonVersion,
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
useGristEntrypoint: true,
|
2021-08-09 14:51:43 +00:00
|
|
|
importDir: options.importMount,
|
2022-02-04 11:13:03 +00:00
|
|
|
...options.sandboxOptions,
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
};
|
|
|
|
return new NSandbox(translatedOptions, spawners[this._flavor]);
|
2020-07-21 13:20:51 +00:00
|
|
|
}
|
|
|
|
}
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
|
|
|
|
// A function that takes sandbox options and starts a sandbox process.
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
type SpawnFn = (options: ISandboxOptions) => SandboxProcess;
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Helper function to run a nacl sandbox. It takes care of most arguments, similarly to
|
|
|
|
* nacl/bin/run script, but without the reliance on bash. We can't use bash when -r/-w options
|
|
|
|
* because on Windows it doesn't pass along the open file descriptors. Bash is also unavailable
|
|
|
|
* when installing a standalone version on Windows.
|
|
|
|
*
|
|
|
|
* This is quite old code, with attention to Windows support that is no longer tested.
|
|
|
|
* I've done my best to avoid changing behavior by not touching it too much.
|
|
|
|
*/
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
function pynbox(options: ISandboxOptions): SandboxProcess {
|
2021-08-09 14:51:43 +00:00
|
|
|
const {command, args: pythonArgs, unsilenceLog, importDir} = options;
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
if (command) {
|
|
|
|
throw new Error("NaCl can only run the specific python2.7 package built for it");
|
|
|
|
}
|
|
|
|
if (options.useGristEntrypoint) {
|
2021-08-09 14:51:43 +00:00
|
|
|
pythonArgs.unshift('grist/main.pyc');
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
}
|
|
|
|
const spawnOptions = {
|
|
|
|
stdio: ['pipe', 'pipe', 'pipe'] as 'pipe'[],
|
|
|
|
env: getWrappingEnv(options)
|
|
|
|
};
|
|
|
|
const wrapperArgs = new FlagBag({env: '-E', mount: '-m'});
|
2021-08-09 14:51:43 +00:00
|
|
|
if (importDir) {
|
|
|
|
wrapperArgs.addMount(`${importDir}:/importdir:ro`);
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
if (!options.minimalPipeMode) {
|
|
|
|
// add two more pipes
|
|
|
|
spawnOptions.stdio.push('pipe', 'pipe');
|
|
|
|
// We use these options to set up communication with the sandbox:
|
|
|
|
// -r 3:3 to associate a file descriptor 3 on the outside of the sandbox with FD 3 on the
|
|
|
|
// inside, for reading from the inside. This becomes `this._streamToSandbox`.
|
|
|
|
// -w 4:4 to associate FD 4 on the outside with FD 4 on the inside for writing from the inside.
|
|
|
|
// This becomes `this._streamFromSandbox`
|
|
|
|
wrapperArgs.push('-r', '3:3', '-w', '4:4');
|
|
|
|
}
|
|
|
|
wrapperArgs.addAllEnv(getInsertedEnv(options));
|
|
|
|
wrapperArgs.addEnv('PYTHONPATH', 'grist:thirdparty');
|
|
|
|
|
|
|
|
const noLog = unsilenceLog ? [] :
|
|
|
|
(process.env.OS === 'Windows_NT' ? ['-l', 'NUL'] : ['-l', '/dev/null']);
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
const child = spawn('sandbox/nacl/bin/sel_ldr', [
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
'-B', './sandbox/nacl/lib/irt_core.nexe', '-m', './sandbox/nacl/root:/:ro',
|
|
|
|
...noLog,
|
|
|
|
...wrapperArgs.get(),
|
|
|
|
'./sandbox/nacl/lib/runnable-ld.so',
|
|
|
|
'--library-path', '/slib', '/python/bin/python2.7.nexe',
|
|
|
|
...pythonArgs
|
|
|
|
], spawnOptions);
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
return {child, control: new DirectProcessControl(child, options.logMeta)};
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Helper function to run python without sandboxing. GRIST_SANDBOX should have
|
|
|
|
* been set with an absolute path to a version of python within a virtualenv that
|
|
|
|
* has all the dependencies installed (e.g. the sandbox_venv3 virtualenv created
|
|
|
|
* by `./build python3`. Using system python works too, if all dependencies have
|
|
|
|
* been installed globally.
|
|
|
|
*/
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
function unsandboxed(options: ISandboxOptions): SandboxProcess {
|
2021-08-09 14:51:43 +00:00
|
|
|
const {args: pythonArgs, importDir} = options;
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
const paths = getAbsolutePaths(options);
|
|
|
|
if (options.useGristEntrypoint) {
|
2021-08-09 14:51:43 +00:00
|
|
|
pythonArgs.unshift(paths.main);
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
}
|
|
|
|
const spawnOptions = {
|
|
|
|
stdio: ['pipe', 'pipe', 'pipe'] as 'pipe'[],
|
|
|
|
env: {
|
|
|
|
PYTHONPATH: paths.engine,
|
2021-08-09 14:51:43 +00:00
|
|
|
IMPORTDIR: importDir,
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
...getInsertedEnv(options),
|
|
|
|
...getWrappingEnv(options),
|
|
|
|
}
|
|
|
|
};
|
|
|
|
if (!options.minimalPipeMode) {
|
|
|
|
spawnOptions.stdio.push('pipe', 'pipe');
|
|
|
|
}
|
2022-01-07 17:06:04 +00:00
|
|
|
const command = findPython(options.command, options.preferredPythonVersion);
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
const child = spawn(command, pythonArgs,
|
|
|
|
{cwd: path.join(process.cwd(), 'sandbox'), ...spawnOptions});
|
|
|
|
return {child, control: new DirectProcessControl(child, options.logMeta)};
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
}
|
|
|
|
|
2023-03-06 21:56:25 +00:00
|
|
|
function pyodide(options: ISandboxOptions): SandboxProcess {
|
|
|
|
const paths = getAbsolutePaths(options);
|
|
|
|
// We will fork with three regular pipes (stdin, stdout, stderr), then
|
|
|
|
// ipc (mandatory for calling fork), and a replacement pipe for stdin
|
|
|
|
// and for stdout.
|
|
|
|
// The regular stdin always opens non-blocking in node, which is a pain
|
|
|
|
// in this case, so we just use a different pipe. There's a different
|
|
|
|
// problem with stdout, with the same solution.
|
|
|
|
const spawnOptions = {
|
|
|
|
stdio: ['ignore', 'ignore', 'pipe', 'ipc', 'pipe', 'pipe'] as Array<'pipe'|'ipc'>,
|
|
|
|
env: {
|
|
|
|
PYTHONPATH: paths.engine,
|
|
|
|
IMPORTDIR: options.importDir,
|
|
|
|
...getInsertedEnv(options),
|
|
|
|
...getWrappingEnv(options),
|
|
|
|
}
|
|
|
|
};
|
|
|
|
const base = getUnpackedAppRoot();
|
|
|
|
const child = fork(path.join(base, 'sandbox', 'pyodide', 'pipe.js'),
|
|
|
|
{cwd: path.join(process.cwd(), 'sandbox'), ...spawnOptions});
|
|
|
|
return {
|
|
|
|
child,
|
|
|
|
control: new DirectProcessControl(child, options.logMeta),
|
|
|
|
dataToSandboxDescriptor: 4, // Cannot use normal descriptor, node
|
|
|
|
// makes it non-blocking. Can be worked around in linux and osx, but
|
|
|
|
// for windows just using a different file descriptor seems simplest.
|
|
|
|
// In the sandbox, calling async methods from emscripten code is
|
|
|
|
// possible but would require more changes to the data engine code
|
|
|
|
// than seems reasonable at this time. The top level sandbox.run
|
|
|
|
// can be tweaked to step operations, which actually works for a
|
|
|
|
// lot of things, but not for cases where the sandbox calls back
|
|
|
|
// into node (e.g. for column type guessing). TLDR: just switching
|
|
|
|
// to FD 4 and reading synchronously is more practical solution.
|
|
|
|
dataFromSandboxDescriptor: 5, // There's an equally long but different
|
|
|
|
// story about why stdout is a bit messed up under pyodide right now.
|
|
|
|
};
|
|
|
|
}
|
|
|
|
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
/**
|
|
|
|
* Helper function to run python in gvisor's runsc, with multiple
|
|
|
|
* sandboxes run within the same container. GRIST_SANDBOX should
|
|
|
|
* point to `sandbox/gvisor/run.py` (to map call onto gvisor's runsc
|
|
|
|
* directly) or `wrap_in_docker.sh` (to use runsc within a container).
|
|
|
|
* Be sure to read setup instructions in that directory.
|
|
|
|
*/
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
function gvisor(options: ISandboxOptions): SandboxProcess {
|
2022-03-24 20:27:34 +00:00
|
|
|
const {args: pythonArgs} = options;
|
|
|
|
let command = options.command;
|
|
|
|
if (!command) {
|
|
|
|
try {
|
|
|
|
// If runsc is available directly on the host, use the wrapper
|
|
|
|
// utility in sandbox/gvisor/run.py to run it.
|
|
|
|
which.sync('runsc');
|
|
|
|
command = 'sandbox/gvisor/run.py';
|
|
|
|
} catch(e) {
|
|
|
|
// Otherwise, don't try any heroics, user will need to
|
|
|
|
// explicitly set the command.
|
|
|
|
throw new Error('runsc not found');
|
|
|
|
}
|
|
|
|
}
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
if (!options.minimalPipeMode) {
|
|
|
|
throw new Error("gvisor only supports 3-pipe operation");
|
|
|
|
}
|
|
|
|
const paths = getAbsolutePaths(options);
|
|
|
|
const wrapperArgs = new FlagBag({env: '-E', mount: '-m'});
|
|
|
|
wrapperArgs.addEnv('PYTHONPATH', paths.engine);
|
|
|
|
wrapperArgs.addAllEnv(getInsertedEnv(options));
|
|
|
|
wrapperArgs.addMount(paths.sandboxDir);
|
2021-08-09 14:51:43 +00:00
|
|
|
if (paths.importDir) {
|
|
|
|
wrapperArgs.addMount(paths.importDir);
|
|
|
|
wrapperArgs.addEnv('IMPORTDIR', paths.importDir);
|
|
|
|
}
|
|
|
|
if (options.useGristEntrypoint) {
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
pythonArgs.unshift(paths.main);
|
|
|
|
}
|
|
|
|
if (options.deterministicMode) {
|
|
|
|
wrapperArgs.push('--faketime', FAKETIME);
|
|
|
|
}
|
2021-08-26 03:12:34 +00:00
|
|
|
const pythonVersion = options.preferredPythonVersion;
|
2021-08-03 21:21:11 +00:00
|
|
|
if (pythonVersion !== '2' && pythonVersion !== '3') {
|
|
|
|
throw new Error("PYTHON_VERSION must be set to 2 or 3");
|
|
|
|
}
|
2022-03-24 20:27:34 +00:00
|
|
|
|
|
|
|
// Check for local virtual environments created with core's
|
|
|
|
// install:python2 or install:python3 targets. They'll need
|
|
|
|
// some extra sharing to make available in the sandbox.
|
2023-02-13 20:52:17 +00:00
|
|
|
const venv = path.join(getAppRootFor(getAppRoot(), 'sandbox'),
|
2022-03-24 20:27:34 +00:00
|
|
|
pythonVersion === '2' ? 'venv' : 'sandbox_venv3');
|
2022-08-11 11:03:45 +00:00
|
|
|
if (fs.existsSync(venv)) {
|
2022-03-24 20:27:34 +00:00
|
|
|
wrapperArgs.addMount(venv);
|
|
|
|
wrapperArgs.push('-s', path.join(venv, 'bin', 'python'));
|
|
|
|
}
|
|
|
|
|
(core) get all tests working under python3/gvisor
Summary:
This verifies that all existing tests are capable of running under python3/gvisor, and fixes the small issues that came up. It does not yet activate python3 tests on all diffs, only diffs that specifically request them.
* Adds a suffix in test names and output directories for tests run with PYTHON_VERSION=3, so that results of the same test run with and without the flag can be aggregated cleanly.
* Adds support for checkpointing to the gvisor sandbox adapter.
* Prepares a checkpoint made after grist python code has loaded in the gvisor sandbox.
* Changes how `DOC_URL` is passed to the sandbox, since it can no longer be passed in as an environment variable when using checkpoints.
* Uses the checkpoint to speed up tests using the gvisor sandbox, otherwise a lot of tests need more time (especially on mac under docker).
* Directs jenkins to run all tests with python2 and python3 when a new file `buildtools/changelogs/python.txt` is touched (this diff counts as touching that file).
* Tweaks miscellaneous tests
- some needed fixes exposed by slightly different timing
- a small number actually give different results in py3 (removal of `u` prefixes).
- some needed a little more time
The DOC_URL change is not the ultimate solution we want for DOC_URL. Eventually it should be a variable that gets updated, like the date perhaps. This is just a small pragmatic change to preserve existing behavior.
Tests are run mindlessly as py3, and for some tests it won't change anything (e.g. if they do not use NSandbox). Tests are not run in parallel, doubling overall test time.
Checkpoints could be useful in deployment, though this diff doesn't use them there.
The application of checkpoints doesn't check for other configuration like 3-versus-5-pipe that we don't actually use.
Python2 tests run using pynbox as always for now.
The diff got sufficiently bulky that I didn't tackle running py3 on "regular" diffs in it. My preference, given that most tests don't appear to stress the python side of things, would be to make a selection of the tests that do and a few wild cards, and run those tests on both pythons rather then all of them. For diffs making a significant python change, I'd propose touching buildtools/changelogs/python.txt for full tests. But this is a conversation in progress.
A total of 6886 tests ran on this diff.
Test Plan: this is a step in preparing tests for py3 transition
Reviewers: dsagal
Reviewed By: dsagal
Subscribers: dsagal
Differential Revision: https://phab.getgrist.com/D3066
2021-10-18 17:37:51 +00:00
|
|
|
// For a regular sandbox not being used for importing, if GRIST_CHECKPOINT is set
|
|
|
|
// try to restore from it. If GRIST_CHECKPOINT_MAKE is set, try to recreate the
|
|
|
|
// checkpoint (this is an awkward place to do it, but avoids mismatches
|
|
|
|
// between the checkpoint and how it gets used later).
|
|
|
|
// If a sandbox is being used for import, it will have a special mount we can't
|
|
|
|
// deal with easily right now. Should be possible to do in future if desired.
|
2022-08-11 11:03:45 +00:00
|
|
|
if (options.useGristEntrypoint && pythonVersion === '3' && process.env.GRIST_CHECKPOINT && !paths.importDir) {
|
(core) get all tests working under python3/gvisor
Summary:
This verifies that all existing tests are capable of running under python3/gvisor, and fixes the small issues that came up. It does not yet activate python3 tests on all diffs, only diffs that specifically request them.
* Adds a suffix in test names and output directories for tests run with PYTHON_VERSION=3, so that results of the same test run with and without the flag can be aggregated cleanly.
* Adds support for checkpointing to the gvisor sandbox adapter.
* Prepares a checkpoint made after grist python code has loaded in the gvisor sandbox.
* Changes how `DOC_URL` is passed to the sandbox, since it can no longer be passed in as an environment variable when using checkpoints.
* Uses the checkpoint to speed up tests using the gvisor sandbox, otherwise a lot of tests need more time (especially on mac under docker).
* Directs jenkins to run all tests with python2 and python3 when a new file `buildtools/changelogs/python.txt` is touched (this diff counts as touching that file).
* Tweaks miscellaneous tests
- some needed fixes exposed by slightly different timing
- a small number actually give different results in py3 (removal of `u` prefixes).
- some needed a little more time
The DOC_URL change is not the ultimate solution we want for DOC_URL. Eventually it should be a variable that gets updated, like the date perhaps. This is just a small pragmatic change to preserve existing behavior.
Tests are run mindlessly as py3, and for some tests it won't change anything (e.g. if they do not use NSandbox). Tests are not run in parallel, doubling overall test time.
Checkpoints could be useful in deployment, though this diff doesn't use them there.
The application of checkpoints doesn't check for other configuration like 3-versus-5-pipe that we don't actually use.
Python2 tests run using pynbox as always for now.
The diff got sufficiently bulky that I didn't tackle running py3 on "regular" diffs in it. My preference, given that most tests don't appear to stress the python side of things, would be to make a selection of the tests that do and a few wild cards, and run those tests on both pythons rather then all of them. For diffs making a significant python change, I'd propose touching buildtools/changelogs/python.txt for full tests. But this is a conversation in progress.
A total of 6886 tests ran on this diff.
Test Plan: this is a step in preparing tests for py3 transition
Reviewers: dsagal
Reviewed By: dsagal
Subscribers: dsagal
Differential Revision: https://phab.getgrist.com/D3066
2021-10-18 17:37:51 +00:00
|
|
|
if (process.env.GRIST_CHECKPOINT_MAKE) {
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
const child =
|
2022-05-06 15:26:45 +00:00
|
|
|
spawn(command, [...wrapperArgs.get(), '--checkpoint', process.env.GRIST_CHECKPOINT!,
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
`python${pythonVersion}`, '--', ...pythonArgs]);
|
|
|
|
// We don't want process control for this.
|
|
|
|
return {child, control: new NoProcessControl(child)};
|
(core) get all tests working under python3/gvisor
Summary:
This verifies that all existing tests are capable of running under python3/gvisor, and fixes the small issues that came up. It does not yet activate python3 tests on all diffs, only diffs that specifically request them.
* Adds a suffix in test names and output directories for tests run with PYTHON_VERSION=3, so that results of the same test run with and without the flag can be aggregated cleanly.
* Adds support for checkpointing to the gvisor sandbox adapter.
* Prepares a checkpoint made after grist python code has loaded in the gvisor sandbox.
* Changes how `DOC_URL` is passed to the sandbox, since it can no longer be passed in as an environment variable when using checkpoints.
* Uses the checkpoint to speed up tests using the gvisor sandbox, otherwise a lot of tests need more time (especially on mac under docker).
* Directs jenkins to run all tests with python2 and python3 when a new file `buildtools/changelogs/python.txt` is touched (this diff counts as touching that file).
* Tweaks miscellaneous tests
- some needed fixes exposed by slightly different timing
- a small number actually give different results in py3 (removal of `u` prefixes).
- some needed a little more time
The DOC_URL change is not the ultimate solution we want for DOC_URL. Eventually it should be a variable that gets updated, like the date perhaps. This is just a small pragmatic change to preserve existing behavior.
Tests are run mindlessly as py3, and for some tests it won't change anything (e.g. if they do not use NSandbox). Tests are not run in parallel, doubling overall test time.
Checkpoints could be useful in deployment, though this diff doesn't use them there.
The application of checkpoints doesn't check for other configuration like 3-versus-5-pipe that we don't actually use.
Python2 tests run using pynbox as always for now.
The diff got sufficiently bulky that I didn't tackle running py3 on "regular" diffs in it. My preference, given that most tests don't appear to stress the python side of things, would be to make a selection of the tests that do and a few wild cards, and run those tests on both pythons rather then all of them. For diffs making a significant python change, I'd propose touching buildtools/changelogs/python.txt for full tests. But this is a conversation in progress.
A total of 6886 tests ran on this diff.
Test Plan: this is a step in preparing tests for py3 transition
Reviewers: dsagal
Reviewed By: dsagal
Subscribers: dsagal
Differential Revision: https://phab.getgrist.com/D3066
2021-10-18 17:37:51 +00:00
|
|
|
}
|
|
|
|
wrapperArgs.push('--restore');
|
2022-05-06 15:26:45 +00:00
|
|
|
wrapperArgs.push(process.env.GRIST_CHECKPOINT!);
|
(core) get all tests working under python3/gvisor
Summary:
This verifies that all existing tests are capable of running under python3/gvisor, and fixes the small issues that came up. It does not yet activate python3 tests on all diffs, only diffs that specifically request them.
* Adds a suffix in test names and output directories for tests run with PYTHON_VERSION=3, so that results of the same test run with and without the flag can be aggregated cleanly.
* Adds support for checkpointing to the gvisor sandbox adapter.
* Prepares a checkpoint made after grist python code has loaded in the gvisor sandbox.
* Changes how `DOC_URL` is passed to the sandbox, since it can no longer be passed in as an environment variable when using checkpoints.
* Uses the checkpoint to speed up tests using the gvisor sandbox, otherwise a lot of tests need more time (especially on mac under docker).
* Directs jenkins to run all tests with python2 and python3 when a new file `buildtools/changelogs/python.txt` is touched (this diff counts as touching that file).
* Tweaks miscellaneous tests
- some needed fixes exposed by slightly different timing
- a small number actually give different results in py3 (removal of `u` prefixes).
- some needed a little more time
The DOC_URL change is not the ultimate solution we want for DOC_URL. Eventually it should be a variable that gets updated, like the date perhaps. This is just a small pragmatic change to preserve existing behavior.
Tests are run mindlessly as py3, and for some tests it won't change anything (e.g. if they do not use NSandbox). Tests are not run in parallel, doubling overall test time.
Checkpoints could be useful in deployment, though this diff doesn't use them there.
The application of checkpoints doesn't check for other configuration like 3-versus-5-pipe that we don't actually use.
Python2 tests run using pynbox as always for now.
The diff got sufficiently bulky that I didn't tackle running py3 on "regular" diffs in it. My preference, given that most tests don't appear to stress the python side of things, would be to make a selection of the tests that do and a few wild cards, and run those tests on both pythons rather then all of them. For diffs making a significant python change, I'd propose touching buildtools/changelogs/python.txt for full tests. But this is a conversation in progress.
A total of 6886 tests ran on this diff.
Test Plan: this is a step in preparing tests for py3 transition
Reviewers: dsagal
Reviewed By: dsagal
Subscribers: dsagal
Differential Revision: https://phab.getgrist.com/D3066
2021-10-18 17:37:51 +00:00
|
|
|
}
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
const child = spawn(command, [...wrapperArgs.get(), `python${pythonVersion}`, '--', ...pythonArgs]);
|
|
|
|
// For gvisor under ptrace, main work is done by a traced process identifiable as
|
|
|
|
// being labeled "exe" and having a parent also labeled "exe".
|
|
|
|
const recognizeTracedProcess = (p: ProcessInfo) => {
|
|
|
|
return p.label.includes('exe') && p.parentLabel.includes('exe');
|
|
|
|
};
|
|
|
|
// The traced process is managed by a regular process called "runsc-sandbox"
|
|
|
|
const recognizeSandboxProcess = (p: ProcessInfo) => {
|
|
|
|
return p.label.includes('runsc-sandbox');
|
|
|
|
};
|
|
|
|
// If docker is in use, this process control will log a warning message and do nothing.
|
|
|
|
return {child, control: new SubprocessControl({
|
|
|
|
pid: child.pid,
|
|
|
|
recognizers: {
|
|
|
|
sandbox: recognizeSandboxProcess, // this process we start and stop
|
|
|
|
memory: recognizeTracedProcess, // measure memory for the ptraced process
|
|
|
|
cpu: recognizeTracedProcess, // measure cpu for the ptraced process
|
|
|
|
traced: recognizeTracedProcess, // the ptraced process
|
|
|
|
},
|
|
|
|
logMeta: options.logMeta
|
|
|
|
})};
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Helper function to run python in a container. Each sandbox run in a
|
|
|
|
* distinct container. GRIST_SANDBOX should be the name of an image where
|
|
|
|
* `python` can be run and all Grist dependencies are installed. See
|
|
|
|
* `sandbox/docker` for more.
|
|
|
|
*/
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
function docker(options: ISandboxOptions): SandboxProcess {
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
const {args: pythonArgs, command} = options;
|
|
|
|
if (options.useGristEntrypoint) {
|
2021-08-09 14:51:43 +00:00
|
|
|
pythonArgs.unshift('grist/main.py');
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
}
|
|
|
|
if (!options.minimalPipeMode) {
|
|
|
|
throw new Error("docker only supports 3-pipe operation (although runc has --preserve-file-descriptors)");
|
|
|
|
}
|
|
|
|
const paths = getAbsolutePaths(options);
|
|
|
|
const wrapperArgs = new FlagBag({env: '--env', mount: '-v'});
|
2021-08-09 14:51:43 +00:00
|
|
|
if (paths.importDir) {
|
|
|
|
wrapperArgs.addMount(`${paths.importDir}:/importdir:ro`);
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
}
|
|
|
|
wrapperArgs.addMount(`${paths.engine}:/grist:ro`);
|
|
|
|
wrapperArgs.addAllEnv(getInsertedEnv(options));
|
|
|
|
wrapperArgs.addEnv('PYTHONPATH', 'grist:thirdparty');
|
|
|
|
const commandParts: string[] = ['python'];
|
|
|
|
if (options.deterministicMode) {
|
|
|
|
// DETERMINISTIC_MODE is already set by getInsertedEnv(). We also take
|
|
|
|
// responsibility here for running faketime around python.
|
|
|
|
commandParts.unshift('faketime', '-f', FAKETIME);
|
|
|
|
}
|
|
|
|
const dockerPath = which.sync('docker');
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
const child = spawn(dockerPath, [
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
'run', '--rm', '-i', '--network', 'none',
|
|
|
|
...wrapperArgs.get(),
|
|
|
|
command || 'grist-docker-sandbox', // this is the docker image to use
|
|
|
|
...commandParts,
|
|
|
|
...pythonArgs,
|
|
|
|
]);
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
log.rawDebug("cannot do process control via docker yet", {...options.logMeta});
|
|
|
|
return {child, control: new NoProcessControl(child)};
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
}
|
|
|
|
|
2021-10-07 17:49:23 +00:00
|
|
|
/**
|
|
|
|
* Helper function to run python using the sandbox-exec command
|
|
|
|
* available on MacOS. This command is a bit shady - not much public
|
|
|
|
* documentation for it, and what there is has been marked deprecated
|
|
|
|
* for a few releases. But mac sandboxing seems to rely heavily on
|
|
|
|
* the infrastructure this command is a thin wrapper around, and there's
|
|
|
|
* no obvious native sandboxing alternative.
|
|
|
|
*/
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
function macSandboxExec(options: ISandboxOptions): SandboxProcess {
|
2021-10-07 17:49:23 +00:00
|
|
|
const {args: pythonArgs} = options;
|
|
|
|
if (!options.minimalPipeMode) {
|
|
|
|
throw new Error("macSandboxExec flavor only supports 3-pipe operation");
|
|
|
|
}
|
|
|
|
const paths = getAbsolutePaths(options);
|
|
|
|
if (options.useGristEntrypoint) {
|
|
|
|
pythonArgs.unshift(paths.main);
|
|
|
|
}
|
|
|
|
const env = {
|
|
|
|
PYTHONPATH: paths.engine,
|
|
|
|
IMPORTDIR: paths.importDir,
|
|
|
|
...getInsertedEnv(options),
|
|
|
|
...getWrappingEnv(options),
|
|
|
|
};
|
2022-01-07 17:06:04 +00:00
|
|
|
const command = findPython(options.command, options.preferredPythonVersion);
|
2021-10-07 17:49:23 +00:00
|
|
|
const realPath = fs.realpathSync(command);
|
(core) get all tests working under python3/gvisor
Summary:
This verifies that all existing tests are capable of running under python3/gvisor, and fixes the small issues that came up. It does not yet activate python3 tests on all diffs, only diffs that specifically request them.
* Adds a suffix in test names and output directories for tests run with PYTHON_VERSION=3, so that results of the same test run with and without the flag can be aggregated cleanly.
* Adds support for checkpointing to the gvisor sandbox adapter.
* Prepares a checkpoint made after grist python code has loaded in the gvisor sandbox.
* Changes how `DOC_URL` is passed to the sandbox, since it can no longer be passed in as an environment variable when using checkpoints.
* Uses the checkpoint to speed up tests using the gvisor sandbox, otherwise a lot of tests need more time (especially on mac under docker).
* Directs jenkins to run all tests with python2 and python3 when a new file `buildtools/changelogs/python.txt` is touched (this diff counts as touching that file).
* Tweaks miscellaneous tests
- some needed fixes exposed by slightly different timing
- a small number actually give different results in py3 (removal of `u` prefixes).
- some needed a little more time
The DOC_URL change is not the ultimate solution we want for DOC_URL. Eventually it should be a variable that gets updated, like the date perhaps. This is just a small pragmatic change to preserve existing behavior.
Tests are run mindlessly as py3, and for some tests it won't change anything (e.g. if they do not use NSandbox). Tests are not run in parallel, doubling overall test time.
Checkpoints could be useful in deployment, though this diff doesn't use them there.
The application of checkpoints doesn't check for other configuration like 3-versus-5-pipe that we don't actually use.
Python2 tests run using pynbox as always for now.
The diff got sufficiently bulky that I didn't tackle running py3 on "regular" diffs in it. My preference, given that most tests don't appear to stress the python side of things, would be to make a selection of the tests that do and a few wild cards, and run those tests on both pythons rather then all of them. For diffs making a significant python change, I'd propose touching buildtools/changelogs/python.txt for full tests. But this is a conversation in progress.
A total of 6886 tests ran on this diff.
Test Plan: this is a step in preparing tests for py3 transition
Reviewers: dsagal
Reviewed By: dsagal
Subscribers: dsagal
Differential Revision: https://phab.getgrist.com/D3066
2021-10-18 17:37:51 +00:00
|
|
|
log.rawDebug("macSandboxExec found a python", {...options.logMeta, command: realPath});
|
2021-10-07 17:49:23 +00:00
|
|
|
|
|
|
|
// Prepare sandbox profile
|
|
|
|
const profile: string[] = [];
|
|
|
|
|
|
|
|
// Deny everything by default, including network
|
|
|
|
profile.push('(version 1)', '(deny default)');
|
|
|
|
|
|
|
|
// Allow execution of the command, either by name provided or ultimate symlink if different
|
|
|
|
profile.push(`(allow process-exec (literal ${JSON.stringify(command)}))`);
|
|
|
|
profile.push(`(allow process-exec (literal ${JSON.stringify(realPath)}))`);
|
|
|
|
|
|
|
|
// There are now a series of extra read and execute permissions added, to deal with the
|
|
|
|
// twisted maze of symlinks around python on a mac.
|
|
|
|
|
|
|
|
// For python symlinks to work, we need to allow reading all the intermediate directories
|
|
|
|
// (this is determined experimentally, perhaps it can be more precise).
|
|
|
|
const intermediatePaths = new Set<string>();
|
|
|
|
for (const target of [command, realPath]) {
|
|
|
|
const parts = path.dirname(target).split(path.sep);
|
|
|
|
for (let i = 1; i < parts.length; i++) {
|
|
|
|
const p = path.join('/', ...parts.slice(0, i));
|
|
|
|
intermediatePaths.add(p);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
for (const p of intermediatePaths) {
|
|
|
|
profile.push(`(allow file-read* (literal ${JSON.stringify(p)}))`);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Grant read access to everything within an enclosing bin directory of original command.
|
|
|
|
if (path.dirname(command).split(path.sep).pop() === 'bin') {
|
|
|
|
const p = path.join(path.dirname(command), '..');
|
|
|
|
profile.push(`(allow file-read* (subpath ${JSON.stringify(p)}))`);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Grant read+execute access to everything within an enclosing bin directory of final target.
|
|
|
|
if (path.dirname(realPath).split(path.sep).pop() === 'bin') {
|
|
|
|
const p = path.join(path.dirname(realPath), '..');
|
|
|
|
profile.push(`(allow file-read* (subpath ${JSON.stringify(p)}))`);
|
|
|
|
profile.push(`(allow process-exec (subpath ${JSON.stringify(p)}))`);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Sundry extra permissions that proved necessary. These work at the time of writing for
|
|
|
|
// python versions installed by brew. Other arrangements could need tweaking.
|
|
|
|
profile.push(`(allow file-read* (subpath "/usr/local/"))`);
|
|
|
|
profile.push('(allow sysctl-read)'); // needed for os.uname()
|
|
|
|
// From another python installation variant.
|
|
|
|
profile.push(`(allow file-read* (subpath "/usr/lib/"))`);
|
|
|
|
profile.push(`(allow file-read* (subpath "/System/Library/Frameworks/"))`);
|
2022-12-19 17:44:57 +00:00
|
|
|
profile.push(`(allow file-read* (subpath "/Library/Apple/usr/libexec/oah/"))`);
|
2021-10-07 17:49:23 +00:00
|
|
|
|
|
|
|
// Give access to Grist material.
|
|
|
|
const cwd = path.join(process.cwd(), 'sandbox');
|
|
|
|
profile.push(`(allow file-read* (subpath ${JSON.stringify(paths.sandboxDir)}))`);
|
|
|
|
profile.push(`(allow file-read* (subpath ${JSON.stringify(cwd)}))`);
|
|
|
|
if (options.importDir) {
|
|
|
|
profile.push(`(allow file-read* (subpath ${JSON.stringify(paths.importDir)}))`);
|
|
|
|
}
|
|
|
|
|
|
|
|
const profileString = profile.join('\n');
|
(core) tweak throttling to work for gvisor/runsc
Summary:
Grist has, up to now, used a throttling mechanism that allows a sandbox free rein until it starts using above some threshold percentage of a cpu for some time - at that point, we start sending STOP and CONT signals on a duty cycle, with longer and longer STOPped periods until cpu usage is at a threshold. The general idea is to do short jobs quickly, while throttling long jobs (thus unfortunately making them even longer) in order to continue doing other short jobs quickly.
The runsc sandbox is not a single process, there are in fact 5 per sandbox in our setup. Runsc can work with kvm or ptrace. Kvm is not available to us, so we use ptrace. With ptrace, there is one process that is the appropriate one to duty cycle, and another that needs to receive a signal in order to yield. This diff adds the necessary machinery.
This is a conservative change, where I stick with our existing throttling mechanism and adapt it to the new sandbox. It would be reasonable to consider switching throttling. There's a lot the OS allows. We can set a quota for how much cpu a process can use within a given period, for example. However the overall behavior with that would be quite different to what we have, so feels like this would need more discussion.
The implementation contains use of a linux utility `pgrep` since portability is not important (runsc is only available on linux) and there's no node api for enumerating children of a process.
The diff contains some tweaks to `buildtools/contain.sh` to streamline experimenting with Grist and runsc on a mac. It is important for throttling that node and the sandbox processes are in the same process name space, if docker is in between them then some extra machinery is needed (a proxy throttler and a way to communicate with it) which I chose not to implement.
Test Plan: added test; a lot of manual testing
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D3113
2021-11-04 20:44:59 +00:00
|
|
|
const child = spawn('/usr/bin/sandbox-exec', ['-p', profileString, command, ...pythonArgs],
|
|
|
|
{cwd, env});
|
|
|
|
return {child, control: new DirectProcessControl(child, options.logMeta)};
|
2021-10-07 17:49:23 +00:00
|
|
|
}
|
|
|
|
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
/**
|
|
|
|
* Collect environment variables that should end up set within the sandbox.
|
|
|
|
*/
|
2021-09-28 18:46:19 +00:00
|
|
|
export function getInsertedEnv(options: ISandboxOptions) {
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
const env: NodeJS.ProcessEnv = {
|
|
|
|
// use stdin/stdout/stderr only.
|
|
|
|
PIPE_MODE: options.minimalPipeMode ? 'minimal' : 'classic',
|
|
|
|
};
|
|
|
|
|
|
|
|
if (options.deterministicMode) {
|
|
|
|
// Making time and randomness act deterministically for testing purposes.
|
|
|
|
// See test/utils/recordPyCalls.ts
|
|
|
|
// tells python to seed the random module
|
|
|
|
env.DETERMINISTIC_MODE = '1';
|
|
|
|
}
|
|
|
|
return env;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Collect environment variables to activate faketime if needed. The paths
|
|
|
|
* here only make sense for unsandboxed operation, or for pynbox. For gvisor,
|
|
|
|
* faketime doesn't work, and must be done inside the sandbox. For docker,
|
|
|
|
* likewise wrapping doesn't make sense. In those cases, LIBFAKETIME_PATH can
|
|
|
|
* just be set to ON to activate faketime in a sandbox dependent manner.
|
|
|
|
*/
|
|
|
|
function getWrappingEnv(options: ISandboxOptions) {
|
|
|
|
const env: NodeJS.ProcessEnv = options.deterministicMode ? {
|
|
|
|
// Making time and randomness act deterministically for testing purposes.
|
|
|
|
// See test/utils/recordPyCalls.ts
|
|
|
|
FAKETIME, // setting for libfaketime
|
|
|
|
// For Linux
|
|
|
|
LD_PRELOAD: process.env.LIBFAKETIME_PATH,
|
|
|
|
|
|
|
|
// For Mac (https://github.com/wolfcw/libfaketime/blob/master/README.OSX)
|
|
|
|
DYLD_INSERT_LIBRARIES: process.env.LIBFAKETIME_PATH,
|
|
|
|
DYLD_FORCE_FLAT_NAMESPACE: '1',
|
|
|
|
} : {};
|
|
|
|
return env;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Extract absolute paths from options. By sticking with the directory
|
|
|
|
* structure on the host rather than remapping, we can simplify nesting
|
|
|
|
* wrappers, or cases where remapping isn't possible. It does leak the names
|
|
|
|
* of the host directories though, and there could be silly complications if the
|
2022-02-19 09:46:49 +00:00
|
|
|
* directories have spaces or other idiosyncrasies. When committing to a sandbox
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
* technology, for stand-alone Grist, it would be worth rethinking this.
|
|
|
|
*/
|
|
|
|
function getAbsolutePaths(options: ISandboxOptions) {
|
|
|
|
// Get path to sandbox directory - this is a little idiosyncratic to work well
|
|
|
|
// in grist-core. It is important to use real paths since we may be viewing
|
|
|
|
// the file system through a narrow window in a container.
|
|
|
|
const sandboxDir = path.join(fs.realpathSync(path.join(process.cwd(), 'sandbox', 'grist')),
|
|
|
|
'..');
|
|
|
|
// Copy plugin options, and then make them absolute.
|
2021-08-09 14:51:43 +00:00
|
|
|
if (options.importDir) {
|
|
|
|
options.importDir = fs.realpathSync(options.importDir);
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
}
|
|
|
|
return {
|
|
|
|
sandboxDir,
|
2021-08-09 14:51:43 +00:00
|
|
|
importDir: options.importDir,
|
(core) support python3 in grist-core, and running engine via docker and/or gvisor
Summary:
* Moves essential plugins to grist-core, so that basic imports (e.g. csv) work.
* Adds support for a `GRIST_SANDBOX_FLAVOR` flag that can systematically override how the data engine is run.
- `GRIST_SANDBOX_FLAVOR=pynbox` is "classic" nacl-based sandbox.
- `GRIST_SANDBOX_FLAVOR=docker` runs engines in individual docker containers. It requires an image specified in `sandbox/docker` (alternative images can be named with `GRIST_SANDBOX` flag - need to contain python and engine requirements). It is a simple reference implementation for sandboxing.
- `GRIST_SANDBOX_FLAVOR=unsandboxed` runs whatever local version of python is specified by a `GRIST_SANDBOX` flag directly, with no sandboxing. Engine requirements must be installed, so an absolute path to a python executable in a virtualenv is easiest to manage.
- `GRIST_SANDBOX_FLAVOR=gvisor` runs the data engine via gvisor's runsc. Experimental, with implementation not included in grist-core. Since gvisor runs on Linux only, this flavor supports wrapping the sandboxes in a single shared docker container.
* Tweaks some recent express query parameter code to work in grist-core, which has a slightly different version of express (smoke test doesn't catch this since in Jenkins core is built within a workspace that has node_modules, and wires get crossed - in a dev environment the problem on master can be seen by doing `buildtools/build_core.sh /tmp/any_path_outside_grist`).
The new sandbox options do not have tests yet, nor does this they change the behavior of grist servers today. They are there to clean up and consolidate a collection of patches I've been using that were getting cumbersome, and make it easier to run experiments.
I haven't looked closely at imports beyond core.
Test Plan: tested manually against regular grist and grist-core, including imports
Reviewers: alexmojaki, dsagal
Reviewed By: alexmojaki
Differential Revision: https://phab.getgrist.com/D2942
2021-07-27 23:43:21 +00:00
|
|
|
main: path.join(sandboxDir, 'grist/main.py'),
|
|
|
|
engine: path.join(sandboxDir, 'grist'),
|
|
|
|
};
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* A tiny abstraction to make code setting up command line arguments a bit
|
|
|
|
* easier to read. The sandboxes are quite similar in spirit, but differ
|
|
|
|
* a bit in exact flags used.
|
|
|
|
*/
|
|
|
|
class FlagBag {
|
|
|
|
private _args: string[] = [];
|
|
|
|
|
|
|
|
constructor(private _options: {env: '--env'|'-E', mount: '-m'|'-v'}) {
|
|
|
|
}
|
|
|
|
|
|
|
|
// channel env variables for sandbox via -E / --env
|
|
|
|
public addEnv(key: string, value: string|undefined) {
|
|
|
|
this._args.push(this._options.env, key + '=' + (value || ''));
|
|
|
|
}
|
|
|
|
|
|
|
|
// Channel all of the supplied env variables
|
|
|
|
public addAllEnv(env: NodeJS.ProcessEnv) {
|
|
|
|
for (const [key, value] of _.toPairs(env)) {
|
|
|
|
this.addEnv(key, value);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// channel shared directory for sandbox via -m / -v
|
|
|
|
public addMount(share: string) {
|
|
|
|
this._args.push(this._options.mount, share);
|
|
|
|
}
|
|
|
|
|
|
|
|
// add some ad-hoc arguments
|
|
|
|
public push(...args: string[]) {
|
|
|
|
this._args.push(...args);
|
|
|
|
}
|
|
|
|
|
|
|
|
// get the final list of arguments
|
|
|
|
public get() { return this._args; }
|
|
|
|
}
|
|
|
|
|
|
|
|
// Standard time to default to if faking time.
|
|
|
|
const FAKETIME = '2020-01-01 00:00:00';
|
2021-10-07 17:49:23 +00:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Find a plausible version of python to run, if none provided.
|
2022-01-07 17:06:04 +00:00
|
|
|
* The preferred version is only used if command is not specified.
|
2021-10-07 17:49:23 +00:00
|
|
|
*/
|
2022-01-07 17:06:04 +00:00
|
|
|
function findPython(command: string|undefined, preferredVersion?: string) {
|
2021-10-07 17:49:23 +00:00
|
|
|
if (command) { return command; }
|
|
|
|
// No command specified. In this case, grist-core looks for a "venv"
|
|
|
|
// virtualenv; a python3 virtualenv would be in "sandbox_venv3".
|
|
|
|
// TODO: rationalize this, it is a product of haphazard growth.
|
2022-01-07 17:06:04 +00:00
|
|
|
const prefs = preferredVersion === '2' ? ['venv', 'sandbox_venv3'] : ['sandbox_venv3', 'venv'];
|
|
|
|
for (const venv of prefs) {
|
2023-02-13 20:52:17 +00:00
|
|
|
const base = getUnpackedAppRoot();
|
|
|
|
// Try a battery of possible python executable paths when python is installed
|
|
|
|
// in a standalone directory.
|
|
|
|
// This battery of possibilities comes from Electron packaging, where python
|
|
|
|
// is bundled with Grist. Not all the possibilities are needed (there are
|
|
|
|
// multiple popular python bundles per OS).
|
|
|
|
for (const possiblePath of [['bin', 'python'], ['bin', 'python3'],
|
|
|
|
['Scripts', 'python.exe'], ['python.exe']] as const) {
|
|
|
|
const pythonPath = path.join(base, venv, ...possiblePath);
|
|
|
|
if (fs.existsSync(pythonPath)) {
|
|
|
|
return pythonPath;
|
|
|
|
}
|
2021-10-07 17:49:23 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
// Fall back on system python.
|
2023-02-13 20:52:17 +00:00
|
|
|
return which.sync(preferredVersion === '2' ? 'python2' : 'python3', {nothrow: true})
|
|
|
|
|| which.sync(preferredVersion === '2' ? 'python2.7' : 'python3.9', {nothrow: true})
|
|
|
|
|| which.sync('python');
|
2021-10-07 17:49:23 +00:00
|
|
|
}
|
2022-03-24 20:27:34 +00:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Create a sandbox. The defaultFlavorSpec is a guide to which sandbox
|
|
|
|
* to create, based on the desired python version. Examples:
|
|
|
|
* unsandboxed # no sandboxing
|
|
|
|
* 2:pynbox,gvisor # run python2 in pynbox, anything else in gvisor
|
|
|
|
* 3:macSandboxExec,docker # run python3 with sandbox-exec, anything else in docker
|
|
|
|
* If no particular python version is desired, the first sandbox listed will be used.
|
|
|
|
* The defaultFlavorSpec can be overridden by GRIST_SANDBOX_FLAVOR.
|
|
|
|
* The commands run can be overridden by GRIST_SANDBOX2 (for python2), GRIST_SANDBOX3 (for python3),
|
|
|
|
* or GRIST_SANDBOX (for either, if more specific variable is not specified).
|
|
|
|
* For documents with no preferred python version specified,
|
|
|
|
* PYTHON_VERSION_ON_CREATION or PYTHON_VERSION is used.
|
|
|
|
*/
|
|
|
|
export function createSandbox(defaultFlavorSpec: string, options: ISandboxCreationOptions): ISandbox {
|
|
|
|
const flavors = (process.env.GRIST_SANDBOX_FLAVOR || defaultFlavorSpec).split(',');
|
|
|
|
const preferredPythonVersion = options.preferredPythonVersion ||
|
|
|
|
process.env.PYTHON_VERSION_ON_CREATION ||
|
|
|
|
process.env.PYTHON_VERSION;
|
|
|
|
for (const flavorAndVersion of flavors) {
|
|
|
|
const parts = flavorAndVersion.trim().split(':', 2);
|
|
|
|
const flavor = parts[parts.length - 1];
|
|
|
|
const version = parts.length === 2 ? parts[0] : '*';
|
|
|
|
if (preferredPythonVersion === version || version === '*' || !preferredPythonVersion) {
|
|
|
|
if (!isFlavor(flavor)) {
|
|
|
|
throw new Error(`Unrecognized sandbox flavor: ${flavor}`);
|
|
|
|
}
|
|
|
|
const creator = new NSandboxCreator({
|
|
|
|
defaultFlavor: flavor,
|
|
|
|
command: process.env['GRIST_SANDBOX' + (preferredPythonVersion||'')] ||
|
|
|
|
process.env['GRIST_SANDBOX'],
|
|
|
|
preferredPythonVersion,
|
|
|
|
});
|
|
|
|
return creator.create(options);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
throw new Error('Failed to create a sandbox');
|
|
|
|
}
|