2020-07-21 13:20:51 +00:00
|
|
|
import {MapWithTTL} from 'app/common/AsyncCreate';
|
|
|
|
import * as version from 'app/common/version';
|
|
|
|
import {DocStatus, DocWorkerInfo, IDocWorkerMap} from 'app/server/lib/DocWorkerMap';
|
|
|
|
import * as log from 'app/server/lib/log';
|
|
|
|
import {checkPermitKey, formatPermitKey, Permit} from 'app/server/lib/Permit';
|
|
|
|
import {promisifyAll} from 'bluebird';
|
|
|
|
import mapValues = require('lodash/mapValues');
|
|
|
|
import {createClient, Multi, RedisClient} from 'redis';
|
|
|
|
import * as Redlock from 'redlock';
|
|
|
|
import * as uuidv4 from 'uuid/v4';
|
|
|
|
|
|
|
|
promisifyAll(RedisClient.prototype);
|
|
|
|
promisifyAll(Multi.prototype);
|
|
|
|
|
|
|
|
// Max time for which we will hold a lock, by default. In milliseconds.
|
|
|
|
const LOCK_TIMEOUT = 3000;
|
|
|
|
|
|
|
|
// How long do checksums stored in redis last. In milliseconds.
|
|
|
|
// Should be long enough to allow S3 to reach consistency with very high probability.
|
|
|
|
// Consistency failures shorter than this interval will be detectable, failures longer
|
|
|
|
// than this interval will not be detectable.
|
|
|
|
const CHECKSUM_TTL_MSEC = 24 * 60 * 60 * 1000; // 24 hours
|
|
|
|
|
|
|
|
// How long do permits stored in redis last, in milliseconds.
|
|
|
|
const PERMIT_TTL_MSEC = 1 * 60 * 1000; // 1 minute
|
|
|
|
|
|
|
|
class DummyDocWorkerMap implements IDocWorkerMap {
|
|
|
|
private _worker?: DocWorkerInfo;
|
|
|
|
private _available: boolean = false;
|
|
|
|
private _permits = new MapWithTTL<string, string>(PERMIT_TTL_MSEC);
|
|
|
|
private _elections = new MapWithTTL<string, string>(1); // default ttl never used
|
|
|
|
|
|
|
|
public async getDocWorker(docId: string) {
|
|
|
|
if (!this._worker) { throw new Error('no workers'); }
|
|
|
|
return {docMD5: 'unknown', docWorker: this._worker, isActive: true};
|
|
|
|
}
|
|
|
|
|
|
|
|
public async assignDocWorker(docId: string) {
|
|
|
|
if (!this._worker || !this._available) { throw new Error('no workers'); }
|
|
|
|
return {docMD5: 'unknown', docWorker: this._worker, isActive: true};
|
|
|
|
}
|
|
|
|
|
|
|
|
public async getDocWorkerOrAssign(docId: string, workerId: string): Promise<DocStatus> {
|
|
|
|
if (!this._worker || !this._available) { throw new Error('no workers'); }
|
|
|
|
if (this._worker.id !== workerId) { throw new Error('worker not known'); }
|
|
|
|
return {docMD5: 'unknown', docWorker: this._worker, isActive: true};
|
|
|
|
}
|
|
|
|
|
|
|
|
public async updateDocStatus(docId: string, checksum: string) {
|
|
|
|
// nothing to do
|
|
|
|
}
|
|
|
|
|
|
|
|
public async addWorker(info: DocWorkerInfo): Promise<void> {
|
|
|
|
this._worker = info;
|
|
|
|
}
|
|
|
|
|
|
|
|
public async removeWorker(workerId: string): Promise<void> {
|
|
|
|
this._worker = undefined;
|
|
|
|
}
|
|
|
|
|
|
|
|
public async setWorkerAvailability(workerId: string, available: boolean): Promise<void> {
|
|
|
|
this._available = available;
|
|
|
|
}
|
|
|
|
|
|
|
|
public async releaseAssignment(workerId: string, docId: string): Promise<void> {
|
|
|
|
// nothing to do
|
|
|
|
}
|
|
|
|
|
|
|
|
public async getAssignments(workerId: string): Promise<string[]> {
|
|
|
|
return [];
|
|
|
|
}
|
|
|
|
|
|
|
|
public async setPermit(permit: Permit): Promise<string> {
|
|
|
|
const key = formatPermitKey(uuidv4());
|
|
|
|
this._permits.set(key, JSON.stringify(permit));
|
|
|
|
return key;
|
|
|
|
}
|
|
|
|
|
|
|
|
public async getPermit(key: string): Promise<Permit> {
|
|
|
|
const result = this._permits.get(key);
|
|
|
|
return result ? JSON.parse(result) : null;
|
|
|
|
}
|
|
|
|
|
|
|
|
public async removePermit(key: string): Promise<void> {
|
|
|
|
this._permits.delete(key);
|
|
|
|
}
|
|
|
|
|
|
|
|
public async close(): Promise<void> {
|
|
|
|
this._permits.clear();
|
|
|
|
this._elections.clear();
|
|
|
|
}
|
|
|
|
|
|
|
|
public async getElection(name: string, durationInMs: number): Promise<string|null> {
|
|
|
|
if (this._elections.get(name)) { return null; }
|
|
|
|
const key = uuidv4();
|
|
|
|
this._elections.setWithCustomTTL(name, key, durationInMs);
|
|
|
|
return key;
|
|
|
|
}
|
|
|
|
|
|
|
|
public async removeElection(name: string, electionKey: string): Promise<void> {
|
|
|
|
if (this._elections.get(name) === electionKey) {
|
|
|
|
this._elections.delete(name);
|
|
|
|
}
|
|
|
|
}
|
2020-10-30 16:53:23 +00:00
|
|
|
|
|
|
|
public async updateChecksum(family: string, key: string, checksum: string) {
|
|
|
|
// nothing to do
|
|
|
|
}
|
|
|
|
|
|
|
|
public async getChecksum(family: string, key: string) {
|
|
|
|
return null;
|
|
|
|
}
|
(core) support GRIST_WORKER_GROUP to place worker into an exclusive group
Summary:
In an emergency, we may want to serve certain documents with "old" workers as we fix problems. This diff adds some support for that.
* Creates duplicate task definitions and services for staging and production doc workers (called grist-docs-staging2 and grist-docs-prod2), pulling from distinct docker tags (staging2 and prod2). The services are set to have zero workers until we need them.
* These new workers are started with a new env variable `GRIST_WORKER_GROUP` set to `secondary`.
* The `GRIST_WORKER_GROUP` variable, if set, makes the worker available to documents in the named group, and only that group.
* An unauthenticated `/assign` endpoint is added to documents which, when POSTed to, checks that the doc is served by a worker in the desired group for that doc (as set manually in redis), and if not frees the doc up for reassignment. This makes it possible to move individual docs between workers without redeployments.
The bash scripts added are a record of how the task definitions + services were created. The services could just have been copied manually, but the task definitions will need to be updated whenever the definitions for the main doc workers are updated, so it is worth scripting that.
For example, if a certain document were to fail on a new deployment of Grist, but rolling back the full deployment wasn't practical:
* Set prod2 tag in docker to desired codebase for that document
* Set desired_count for grist-docs-prod2 service to non-zero
* Set doc-<docid>-group for that doc in redis to secondary
* Hit /api/docs/<docid>/assign to move the doc to grist-docs-prod2
(If the document needs to be reverted to a previous snapshot, that currently would need doing manually - could be made simpler, but not in scope of this diff).
Test Plan: added tests
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D2649
2020-11-02 19:24:46 +00:00
|
|
|
|
|
|
|
public async getWorkerGroup(workerId: string): Promise<string|null> {
|
|
|
|
return null;
|
|
|
|
}
|
|
|
|
|
|
|
|
public async getDocGroup(docId: string): Promise<string|null> {
|
|
|
|
return null;
|
|
|
|
}
|
2020-07-21 13:20:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Manage the relationship between document and workers. Backed by Redis.
|
|
|
|
* Can also assign workers to "groups" for serving particular documents.
|
|
|
|
* Keys used:
|
|
|
|
* workers - the set of known active workers, identified by workerId
|
|
|
|
* workers-available - the set of workers available for assignment (a subset of the workers set)
|
|
|
|
* workers-available-{group} - the set of workers available for a given group
|
|
|
|
* worker-{workerId} - a hash of contact information for a worker
|
|
|
|
* worker-{workerId}-docs - a set of docs assigned to a worker, identified by docId
|
|
|
|
* worker-{workerId}-group - if set, marks the worker as serving a particular group
|
|
|
|
* doc-${docId} - a hash containing (JSON serialized) DocStatus fields, other than docMD5.
|
|
|
|
* doc-${docId}-checksum - the docs docMD5, or 'null' if docMD5 is null
|
|
|
|
* doc-${docId}-group - if set, marks the doc as to be served by workers in a given group
|
|
|
|
* workers-lock - a lock used when working with the list of workers
|
|
|
|
* groups - a hash from groupIds (arbitrary strings) to desired number of workers in group
|
|
|
|
* elections-${deployment} - a hash, from groupId to a (serialized json) list of worker ids
|
|
|
|
*
|
|
|
|
* Assignments of documents to workers can end abruptly at any time. Clients
|
|
|
|
* should be prepared to retry if a worker is not responding or denies that a document
|
|
|
|
* is assigned to it.
|
|
|
|
*
|
|
|
|
* If the groups key is set, workers assign themselves to groupIds to
|
|
|
|
* fill the counts specified in groups (in order of groupIds), and
|
|
|
|
* once those are exhausted, get assigned to the special group
|
|
|
|
* "default".
|
|
|
|
*/
|
|
|
|
export class DocWorkerMap implements IDocWorkerMap {
|
|
|
|
private _client: RedisClient;
|
(core) support GRIST_WORKER_GROUP to place worker into an exclusive group
Summary:
In an emergency, we may want to serve certain documents with "old" workers as we fix problems. This diff adds some support for that.
* Creates duplicate task definitions and services for staging and production doc workers (called grist-docs-staging2 and grist-docs-prod2), pulling from distinct docker tags (staging2 and prod2). The services are set to have zero workers until we need them.
* These new workers are started with a new env variable `GRIST_WORKER_GROUP` set to `secondary`.
* The `GRIST_WORKER_GROUP` variable, if set, makes the worker available to documents in the named group, and only that group.
* An unauthenticated `/assign` endpoint is added to documents which, when POSTed to, checks that the doc is served by a worker in the desired group for that doc (as set manually in redis), and if not frees the doc up for reassignment. This makes it possible to move individual docs between workers without redeployments.
The bash scripts added are a record of how the task definitions + services were created. The services could just have been copied manually, but the task definitions will need to be updated whenever the definitions for the main doc workers are updated, so it is worth scripting that.
For example, if a certain document were to fail on a new deployment of Grist, but rolling back the full deployment wasn't practical:
* Set prod2 tag in docker to desired codebase for that document
* Set desired_count for grist-docs-prod2 service to non-zero
* Set doc-<docid>-group for that doc in redis to secondary
* Hit /api/docs/<docid>/assign to move the doc to grist-docs-prod2
(If the document needs to be reverted to a previous snapshot, that currently would need doing manually - could be made simpler, but not in scope of this diff).
Test Plan: added tests
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D2649
2020-11-02 19:24:46 +00:00
|
|
|
private _clients: RedisClient[];
|
2020-07-21 13:20:51 +00:00
|
|
|
private _redlock: Redlock;
|
|
|
|
|
|
|
|
// Optional deploymentKey argument supplies a key unique to the deployment (this is important
|
|
|
|
// for maintaining groups across redeployments only)
|
|
|
|
constructor(_clients?: RedisClient[], private _deploymentKey?: string, private _options?: {
|
|
|
|
permitMsec?: number
|
|
|
|
}) {
|
|
|
|
this._deploymentKey = this._deploymentKey || version.version;
|
(core) support GRIST_WORKER_GROUP to place worker into an exclusive group
Summary:
In an emergency, we may want to serve certain documents with "old" workers as we fix problems. This diff adds some support for that.
* Creates duplicate task definitions and services for staging and production doc workers (called grist-docs-staging2 and grist-docs-prod2), pulling from distinct docker tags (staging2 and prod2). The services are set to have zero workers until we need them.
* These new workers are started with a new env variable `GRIST_WORKER_GROUP` set to `secondary`.
* The `GRIST_WORKER_GROUP` variable, if set, makes the worker available to documents in the named group, and only that group.
* An unauthenticated `/assign` endpoint is added to documents which, when POSTed to, checks that the doc is served by a worker in the desired group for that doc (as set manually in redis), and if not frees the doc up for reassignment. This makes it possible to move individual docs between workers without redeployments.
The bash scripts added are a record of how the task definitions + services were created. The services could just have been copied manually, but the task definitions will need to be updated whenever the definitions for the main doc workers are updated, so it is worth scripting that.
For example, if a certain document were to fail on a new deployment of Grist, but rolling back the full deployment wasn't practical:
* Set prod2 tag in docker to desired codebase for that document
* Set desired_count for grist-docs-prod2 service to non-zero
* Set doc-<docid>-group for that doc in redis to secondary
* Hit /api/docs/<docid>/assign to move the doc to grist-docs-prod2
(If the document needs to be reverted to a previous snapshot, that currently would need doing manually - could be made simpler, but not in scope of this diff).
Test Plan: added tests
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D2649
2020-11-02 19:24:46 +00:00
|
|
|
this._clients = _clients || [createClient(process.env.REDIS_URL)];
|
|
|
|
this._redlock = new Redlock(this._clients);
|
|
|
|
this._client = this._clients[0]!;
|
2020-07-21 13:20:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
public async addWorker(info: DocWorkerInfo): Promise<void> {
|
|
|
|
log.info(`DocWorkerMap.addWorker ${info.id}`);
|
|
|
|
const lock = await this._redlock.lock('workers-lock', LOCK_TIMEOUT);
|
|
|
|
try {
|
(core) support GRIST_WORKER_GROUP to place worker into an exclusive group
Summary:
In an emergency, we may want to serve certain documents with "old" workers as we fix problems. This diff adds some support for that.
* Creates duplicate task definitions and services for staging and production doc workers (called grist-docs-staging2 and grist-docs-prod2), pulling from distinct docker tags (staging2 and prod2). The services are set to have zero workers until we need them.
* These new workers are started with a new env variable `GRIST_WORKER_GROUP` set to `secondary`.
* The `GRIST_WORKER_GROUP` variable, if set, makes the worker available to documents in the named group, and only that group.
* An unauthenticated `/assign` endpoint is added to documents which, when POSTed to, checks that the doc is served by a worker in the desired group for that doc (as set manually in redis), and if not frees the doc up for reassignment. This makes it possible to move individual docs between workers without redeployments.
The bash scripts added are a record of how the task definitions + services were created. The services could just have been copied manually, but the task definitions will need to be updated whenever the definitions for the main doc workers are updated, so it is worth scripting that.
For example, if a certain document were to fail on a new deployment of Grist, but rolling back the full deployment wasn't practical:
* Set prod2 tag in docker to desired codebase for that document
* Set desired_count for grist-docs-prod2 service to non-zero
* Set doc-<docid>-group for that doc in redis to secondary
* Hit /api/docs/<docid>/assign to move the doc to grist-docs-prod2
(If the document needs to be reverted to a previous snapshot, that currently would need doing manually - could be made simpler, but not in scope of this diff).
Test Plan: added tests
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D2649
2020-11-02 19:24:46 +00:00
|
|
|
// Make a worker-{workerId} key with contact info.
|
2020-07-21 13:20:51 +00:00
|
|
|
await this._client.hmsetAsync(`worker-${info.id}`, info);
|
(core) support GRIST_WORKER_GROUP to place worker into an exclusive group
Summary:
In an emergency, we may want to serve certain documents with "old" workers as we fix problems. This diff adds some support for that.
* Creates duplicate task definitions and services for staging and production doc workers (called grist-docs-staging2 and grist-docs-prod2), pulling from distinct docker tags (staging2 and prod2). The services are set to have zero workers until we need them.
* These new workers are started with a new env variable `GRIST_WORKER_GROUP` set to `secondary`.
* The `GRIST_WORKER_GROUP` variable, if set, makes the worker available to documents in the named group, and only that group.
* An unauthenticated `/assign` endpoint is added to documents which, when POSTed to, checks that the doc is served by a worker in the desired group for that doc (as set manually in redis), and if not frees the doc up for reassignment. This makes it possible to move individual docs between workers without redeployments.
The bash scripts added are a record of how the task definitions + services were created. The services could just have been copied manually, but the task definitions will need to be updated whenever the definitions for the main doc workers are updated, so it is worth scripting that.
For example, if a certain document were to fail on a new deployment of Grist, but rolling back the full deployment wasn't practical:
* Set prod2 tag in docker to desired codebase for that document
* Set desired_count for grist-docs-prod2 service to non-zero
* Set doc-<docid>-group for that doc in redis to secondary
* Hit /api/docs/<docid>/assign to move the doc to grist-docs-prod2
(If the document needs to be reverted to a previous snapshot, that currently would need doing manually - could be made simpler, but not in scope of this diff).
Test Plan: added tests
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D2649
2020-11-02 19:24:46 +00:00
|
|
|
// Add this worker to set of workers (but don't make it available for work yet).
|
2020-07-21 13:20:51 +00:00
|
|
|
await this._client.saddAsync('workers', info.id);
|
(core) support GRIST_WORKER_GROUP to place worker into an exclusive group
Summary:
In an emergency, we may want to serve certain documents with "old" workers as we fix problems. This diff adds some support for that.
* Creates duplicate task definitions and services for staging and production doc workers (called grist-docs-staging2 and grist-docs-prod2), pulling from distinct docker tags (staging2 and prod2). The services are set to have zero workers until we need them.
* These new workers are started with a new env variable `GRIST_WORKER_GROUP` set to `secondary`.
* The `GRIST_WORKER_GROUP` variable, if set, makes the worker available to documents in the named group, and only that group.
* An unauthenticated `/assign` endpoint is added to documents which, when POSTed to, checks that the doc is served by a worker in the desired group for that doc (as set manually in redis), and if not frees the doc up for reassignment. This makes it possible to move individual docs between workers without redeployments.
The bash scripts added are a record of how the task definitions + services were created. The services could just have been copied manually, but the task definitions will need to be updated whenever the definitions for the main doc workers are updated, so it is worth scripting that.
For example, if a certain document were to fail on a new deployment of Grist, but rolling back the full deployment wasn't practical:
* Set prod2 tag in docker to desired codebase for that document
* Set desired_count for grist-docs-prod2 service to non-zero
* Set doc-<docid>-group for that doc in redis to secondary
* Hit /api/docs/<docid>/assign to move the doc to grist-docs-prod2
(If the document needs to be reverted to a previous snapshot, that currently would need doing manually - could be made simpler, but not in scope of this diff).
Test Plan: added tests
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D2649
2020-11-02 19:24:46 +00:00
|
|
|
|
|
|
|
if (info.group) {
|
|
|
|
// Accept work only for a specific group.
|
|
|
|
// Do not accept work not associated with the specified group.
|
|
|
|
await this._client.setAsync(`worker-${info.id}-group`, info.group);
|
|
|
|
} else {
|
|
|
|
// Figure out if worker should belong to a group via elections.
|
|
|
|
// Be careful: elections happen within a single deployment, so are somewhat
|
|
|
|
// unintuitive in behavior. For example, if a document is assigned to a group
|
|
|
|
// but there is no worker available for that group, it may open on any worker.
|
|
|
|
// And if a worker is assigned to a group, it may still end up assigned work
|
|
|
|
// not associated with that group if it is the only worker available.
|
|
|
|
const groups = await this._client.hgetallAsync('groups');
|
|
|
|
if (groups) {
|
|
|
|
const elections = await this._client.hgetallAsync(`elections-${this._deploymentKey}`) || {};
|
|
|
|
for (const group of Object.keys(groups).sort()) {
|
|
|
|
const count = parseInt(groups[group], 10) || 0;
|
|
|
|
if (count < 1) { continue; }
|
|
|
|
const elected: string[] = JSON.parse(elections[group] || '[]');
|
|
|
|
if (elected.length >= count) { continue; }
|
|
|
|
elected.push(info.id);
|
|
|
|
await this._client.setAsync(`worker-${info.id}-group`, group);
|
|
|
|
await this._client.hsetAsync(`elections-${this._deploymentKey}`, group, JSON.stringify(elected));
|
|
|
|
break;
|
|
|
|
}
|
2020-07-21 13:20:51 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
} finally {
|
|
|
|
await lock.unlock();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
public async removeWorker(workerId: string): Promise<void> {
|
|
|
|
log.info(`DocWorkerMap.removeWorker ${workerId}`);
|
|
|
|
const lock = await this._redlock.lock('workers-lock', LOCK_TIMEOUT);
|
|
|
|
try {
|
|
|
|
// Drop out of available set first.
|
|
|
|
await this._client.sremAsync('workers-available', workerId);
|
|
|
|
const group = await this._client.getAsync(`worker-${workerId}-group`) || 'default';
|
|
|
|
await this._client.sremAsync(`workers-available-${group}`, workerId);
|
|
|
|
// At this point, this worker should no longer be receiving new doc assignments, though
|
|
|
|
// clients may still be directed to the worker.
|
|
|
|
|
|
|
|
// If we were elected for anything, back out.
|
|
|
|
const elections = await this._client.hgetallAsync(`elections-${this._deploymentKey}`);
|
|
|
|
if (elections) {
|
|
|
|
if (group in elections) {
|
|
|
|
const elected: string[] = JSON.parse(elections[group]);
|
|
|
|
const newElected = elected.filter(worker => worker !== workerId);
|
|
|
|
if (elected.length !== newElected.length) {
|
|
|
|
if (newElected.length > 0) {
|
|
|
|
await this._client.hsetAsync(`elections-${this._deploymentKey}`, group,
|
|
|
|
JSON.stringify(newElected));
|
|
|
|
} else {
|
|
|
|
await this._client.hdelAsync(`elections-${this._deploymentKey}`, group);
|
|
|
|
delete elections[group];
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// We're the last one involved in elections - remove the key entirely.
|
|
|
|
if (Object.keys(elected).length === 0) {
|
|
|
|
await this._client.delAsync(`elections-${this._deploymentKey}`);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Now, we start removing the assignments.
|
|
|
|
const assignments = await this._client.smembersAsync(`worker-${workerId}-docs`);
|
|
|
|
if (assignments) {
|
|
|
|
const op = this._client.multi();
|
|
|
|
for (const doc of assignments) { op.del(`doc-${doc}`); }
|
|
|
|
await op.execAsync();
|
|
|
|
}
|
|
|
|
|
|
|
|
// Now remove worker-{workerId}* keys.
|
|
|
|
await this._client.delAsync(`worker-${workerId}-docs`);
|
|
|
|
await this._client.delAsync(`worker-${workerId}-group`);
|
|
|
|
await this._client.delAsync(`worker-${workerId}`);
|
|
|
|
|
|
|
|
// Forget about this worker completely.
|
|
|
|
await this._client.sremAsync('workers', workerId);
|
|
|
|
} finally {
|
|
|
|
await lock.unlock();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
public async setWorkerAvailability(workerId: string, available: boolean): Promise<void> {
|
|
|
|
log.info(`DocWorkerMap.setWorkerAvailability ${workerId} ${available}`);
|
|
|
|
const group = await this._client.getAsync(`worker-${workerId}-group`) || 'default';
|
|
|
|
if (available) {
|
(core) support GRIST_WORKER_GROUP to place worker into an exclusive group
Summary:
In an emergency, we may want to serve certain documents with "old" workers as we fix problems. This diff adds some support for that.
* Creates duplicate task definitions and services for staging and production doc workers (called grist-docs-staging2 and grist-docs-prod2), pulling from distinct docker tags (staging2 and prod2). The services are set to have zero workers until we need them.
* These new workers are started with a new env variable `GRIST_WORKER_GROUP` set to `secondary`.
* The `GRIST_WORKER_GROUP` variable, if set, makes the worker available to documents in the named group, and only that group.
* An unauthenticated `/assign` endpoint is added to documents which, when POSTed to, checks that the doc is served by a worker in the desired group for that doc (as set manually in redis), and if not frees the doc up for reassignment. This makes it possible to move individual docs between workers without redeployments.
The bash scripts added are a record of how the task definitions + services were created. The services could just have been copied manually, but the task definitions will need to be updated whenever the definitions for the main doc workers are updated, so it is worth scripting that.
For example, if a certain document were to fail on a new deployment of Grist, but rolling back the full deployment wasn't practical:
* Set prod2 tag in docker to desired codebase for that document
* Set desired_count for grist-docs-prod2 service to non-zero
* Set doc-<docid>-group for that doc in redis to secondary
* Hit /api/docs/<docid>/assign to move the doc to grist-docs-prod2
(If the document needs to be reverted to a previous snapshot, that currently would need doing manually - could be made simpler, but not in scope of this diff).
Test Plan: added tests
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D2649
2020-11-02 19:24:46 +00:00
|
|
|
const docWorker = await this._client.hgetallAsync(`worker-${workerId}`) as DocWorkerInfo|null;
|
|
|
|
if (!docWorker) { throw new Error('no doc worker contact info available'); }
|
2020-07-21 13:20:51 +00:00
|
|
|
await this._client.saddAsync(`workers-available-${group}`, workerId);
|
(core) support GRIST_WORKER_GROUP to place worker into an exclusive group
Summary:
In an emergency, we may want to serve certain documents with "old" workers as we fix problems. This diff adds some support for that.
* Creates duplicate task definitions and services for staging and production doc workers (called grist-docs-staging2 and grist-docs-prod2), pulling from distinct docker tags (staging2 and prod2). The services are set to have zero workers until we need them.
* These new workers are started with a new env variable `GRIST_WORKER_GROUP` set to `secondary`.
* The `GRIST_WORKER_GROUP` variable, if set, makes the worker available to documents in the named group, and only that group.
* An unauthenticated `/assign` endpoint is added to documents which, when POSTed to, checks that the doc is served by a worker in the desired group for that doc (as set manually in redis), and if not frees the doc up for reassignment. This makes it possible to move individual docs between workers without redeployments.
The bash scripts added are a record of how the task definitions + services were created. The services could just have been copied manually, but the task definitions will need to be updated whenever the definitions for the main doc workers are updated, so it is worth scripting that.
For example, if a certain document were to fail on a new deployment of Grist, but rolling back the full deployment wasn't practical:
* Set prod2 tag in docker to desired codebase for that document
* Set desired_count for grist-docs-prod2 service to non-zero
* Set doc-<docid>-group for that doc in redis to secondary
* Hit /api/docs/<docid>/assign to move the doc to grist-docs-prod2
(If the document needs to be reverted to a previous snapshot, that currently would need doing manually - could be made simpler, but not in scope of this diff).
Test Plan: added tests
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D2649
2020-11-02 19:24:46 +00:00
|
|
|
// If we're not assigned exclusively to a group, add this worker also to the general
|
|
|
|
// pool of workers.
|
|
|
|
if (!docWorker.group) {
|
|
|
|
await this._client.saddAsync('workers-available', workerId);
|
|
|
|
}
|
2020-07-21 13:20:51 +00:00
|
|
|
} else {
|
|
|
|
await this._client.sremAsync('workers-available', workerId);
|
|
|
|
await this._client.sremAsync(`workers-available-${group}`, workerId);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
public async releaseAssignment(workerId: string, docId: string): Promise<void> {
|
|
|
|
const op = this._client.multi();
|
|
|
|
op.del(`doc-${docId}`);
|
|
|
|
op.srem(`worker-${workerId}-docs`, docId);
|
|
|
|
await op.execAsync();
|
|
|
|
}
|
|
|
|
|
|
|
|
public async getAssignments(workerId: string): Promise<string[]> {
|
|
|
|
return this._client.smembersAsync(`worker-${workerId}-docs`);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Defined by IDocWorkerMap.
|
|
|
|
*
|
|
|
|
* Looks up which DocWorker is responsible for this docId.
|
|
|
|
* Responsibility could change at any time after this call, so it
|
|
|
|
* should be treated as a hint, and clients should be prepared to be
|
|
|
|
* refused and need to retry.
|
|
|
|
*/
|
|
|
|
public async getDocWorker(docId: string): Promise<DocStatus|null> {
|
|
|
|
// Fetch the various elements that go into making a DocStatus
|
|
|
|
const props = await this._client.multi()
|
|
|
|
.hgetall(`doc-${docId}`)
|
|
|
|
.get(`doc-${docId}-checksum`)
|
|
|
|
.execAsync() as [{[key: string]: any}|null, string|null]|null;
|
|
|
|
if (!props) { return null; }
|
|
|
|
|
|
|
|
// If there is no worker, return null. An alternative would be to modify
|
|
|
|
// DocStatus so that it is possible for it to not have a worker assignment.
|
|
|
|
if (!props[0]) { return null; }
|
|
|
|
|
|
|
|
// Fields are JSON encoded since redis cannot store them directly.
|
|
|
|
const doc = mapValues(props[0], (val) => JSON.parse(val));
|
|
|
|
|
|
|
|
// Redis cannot store a null value, so we encode it as 'null', which does
|
|
|
|
// not match any possible MD5.
|
|
|
|
doc.docMD5 = props[1] === 'null' ? null : props[1];
|
|
|
|
|
|
|
|
// Ok, we have a valid DocStatus at this point.
|
|
|
|
return doc as DocStatus;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
*
|
|
|
|
* Defined by IDocWorkerMap.
|
|
|
|
*
|
|
|
|
* Assigns a DocWorker to this docId if one is not yet assigned.
|
|
|
|
* Note that the assignment could be unmade at any time after this
|
|
|
|
* call if the worker dies, is brought down, or for other potential
|
|
|
|
* reasons in the future such as migration of individual documents
|
|
|
|
* between workers.
|
|
|
|
*
|
|
|
|
* A preferred doc worker can be specified, which will be assigned
|
|
|
|
* if no assignment is already made.
|
|
|
|
*
|
2021-01-12 15:48:40 +00:00
|
|
|
* For the special docId "import", return a potential assignment.
|
|
|
|
* It will be up to the doc worker to assign the eventually
|
|
|
|
* created document, if desired.
|
|
|
|
*
|
2020-07-21 13:20:51 +00:00
|
|
|
*/
|
|
|
|
public async assignDocWorker(docId: string, workerId?: string): Promise<DocStatus> {
|
2021-01-12 15:48:40 +00:00
|
|
|
if (docId === 'import') {
|
|
|
|
const lock = await this._redlock.lock(`workers-lock`, LOCK_TIMEOUT);
|
|
|
|
try {
|
2021-05-23 17:43:11 +00:00
|
|
|
const _workerId = await this._client.srandmemberAsync(`workers-available-default`);
|
|
|
|
if (!_workerId) { throw new Error('no doc worker available'); }
|
|
|
|
const docWorker = await this._client.hgetallAsync(`worker-${_workerId}`) as DocWorkerInfo|null;
|
2021-01-12 15:48:40 +00:00
|
|
|
if (!docWorker) { throw new Error('no doc worker contact info available'); }
|
|
|
|
return {
|
|
|
|
docMD5: null,
|
|
|
|
docWorker,
|
|
|
|
isActive: false
|
|
|
|
};
|
|
|
|
} finally {
|
|
|
|
await lock.unlock();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2020-07-21 13:20:51 +00:00
|
|
|
// Check if a DocWorker is already assigned; if so return result immediately
|
|
|
|
// without locking.
|
|
|
|
let docStatus = await this.getDocWorker(docId);
|
|
|
|
if (docStatus) { return docStatus; }
|
|
|
|
|
|
|
|
// No assignment yet, so let's lock and set an assignment up.
|
|
|
|
const lock = await this._redlock.lock(`workers-lock`, LOCK_TIMEOUT);
|
|
|
|
|
|
|
|
try {
|
|
|
|
// Now that we've locked, recheck that the worker hasn't been reassigned
|
|
|
|
// in the meantime. Return immediately if it has.
|
|
|
|
docStatus = await this.getDocWorker(docId);
|
|
|
|
if (docStatus) { return docStatus; }
|
|
|
|
|
|
|
|
if (!workerId) {
|
|
|
|
// Check if document has a preferred worker group set.
|
|
|
|
const group = await this._client.getAsync(`doc-${docId}-group`) || 'default';
|
|
|
|
|
|
|
|
// Let's start off by assigning documents to available workers randomly.
|
|
|
|
// TODO: use a smarter algorithm.
|
|
|
|
workerId = await this._client.srandmemberAsync(`workers-available-${group}`) || undefined;
|
|
|
|
if (!workerId) {
|
|
|
|
// No workers available in the desired worker group. Rather than refusing to
|
|
|
|
// open the document, we fall back on assigning a worker from any of the workers
|
|
|
|
// available, regardless of grouping.
|
|
|
|
// This limits the impact of operational misconfiguration (bad redis setup,
|
|
|
|
// or not starting enough workers). It has the downside of potentially disguising
|
|
|
|
// problems, so we log a warning.
|
|
|
|
log.warn(`DocWorkerMap.assignDocWorker ${docId} found no workers for group ${group}`);
|
|
|
|
workerId = await this._client.srandmemberAsync('workers-available') || undefined;
|
|
|
|
}
|
|
|
|
if (!workerId) { throw new Error('no doc workers available'); }
|
|
|
|
} else {
|
|
|
|
if (!await this._client.sismemberAsync('workers-available', workerId)) {
|
|
|
|
throw new Error(`worker ${workerId} not known or not available`);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Look up how to contact the worker.
|
|
|
|
const docWorker = await this._client.hgetallAsync(`worker-${workerId}`) as DocWorkerInfo|null;
|
|
|
|
if (!docWorker) { throw new Error('no doc worker contact info available'); }
|
|
|
|
|
|
|
|
// We can now construct a DocStatus.
|
|
|
|
const newDocStatus = {docMD5: null, docWorker, isActive: true};
|
|
|
|
|
|
|
|
// We add the assignment to worker-{workerId}-docs and save doc-{docId}.
|
|
|
|
const result = await this._client.multi()
|
|
|
|
.sadd(`worker-${workerId}-docs`, docId)
|
|
|
|
.hmset(`doc-${docId}`, {
|
|
|
|
docWorker: JSON.stringify(docWorker), // redis can't store nested objects, strings only
|
|
|
|
isActive: JSON.stringify(true) // redis can't store booleans, strings only
|
|
|
|
})
|
|
|
|
.setex(`doc-${docId}-checksum`, CHECKSUM_TTL_MSEC / 1000.0, 'null')
|
|
|
|
.execAsync();
|
|
|
|
if (!result) { throw new Error('failed to store new assignment'); }
|
|
|
|
return newDocStatus;
|
|
|
|
} finally {
|
|
|
|
await lock.unlock();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
*
|
|
|
|
* Defined by IDocWorkerMap.
|
|
|
|
*
|
|
|
|
* Assigns a specific DocWorker to this docId if one is not yet assigned.
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
public async getDocWorkerOrAssign(docId: string, workerId: string): Promise<DocStatus> {
|
|
|
|
return this.assignDocWorker(docId, workerId);
|
|
|
|
}
|
|
|
|
|
|
|
|
public async updateDocStatus(docId: string, checksum: string): Promise<void> {
|
2021-04-26 21:54:09 +00:00
|
|
|
return this.updateChecksum('doc', docId, checksum);
|
2020-10-30 16:53:23 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
public async updateChecksum(family: string, key: string, checksum: string) {
|
|
|
|
await this._client.setexAsync(`${family}-${key}-checksum`, CHECKSUM_TTL_MSEC / 1000.0, checksum);
|
|
|
|
}
|
|
|
|
|
|
|
|
public async getChecksum(family: string, key: string) {
|
|
|
|
const checksum = await this._client.getAsync(`${family}-${key}-checksum`);
|
|
|
|
return checksum === 'null' ? null : checksum;
|
2020-07-21 13:20:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
public async setPermit(permit: Permit): Promise<string> {
|
|
|
|
const key = formatPermitKey(uuidv4());
|
|
|
|
const duration = (this._options && this._options.permitMsec) || PERMIT_TTL_MSEC;
|
|
|
|
// seems like only integer seconds are supported?
|
|
|
|
await this._client.setexAsync(key, Math.ceil(duration / 1000.0),
|
|
|
|
JSON.stringify(permit));
|
|
|
|
return key;
|
|
|
|
}
|
|
|
|
|
|
|
|
public async getPermit(key: string): Promise<Permit|null> {
|
|
|
|
if (!checkPermitKey(key)) { throw new Error('permit could not be read'); }
|
|
|
|
const result = await this._client.getAsync(key);
|
|
|
|
return result && JSON.parse(result);
|
|
|
|
}
|
|
|
|
|
|
|
|
public async removePermit(key: string): Promise<void> {
|
|
|
|
if (!checkPermitKey(key)) { throw new Error('permit could not be read'); }
|
|
|
|
await this._client.delAsync(key);
|
|
|
|
}
|
|
|
|
|
|
|
|
public async close(): Promise<void> {
|
(core) support GRIST_WORKER_GROUP to place worker into an exclusive group
Summary:
In an emergency, we may want to serve certain documents with "old" workers as we fix problems. This diff adds some support for that.
* Creates duplicate task definitions and services for staging and production doc workers (called grist-docs-staging2 and grist-docs-prod2), pulling from distinct docker tags (staging2 and prod2). The services are set to have zero workers until we need them.
* These new workers are started with a new env variable `GRIST_WORKER_GROUP` set to `secondary`.
* The `GRIST_WORKER_GROUP` variable, if set, makes the worker available to documents in the named group, and only that group.
* An unauthenticated `/assign` endpoint is added to documents which, when POSTed to, checks that the doc is served by a worker in the desired group for that doc (as set manually in redis), and if not frees the doc up for reassignment. This makes it possible to move individual docs between workers without redeployments.
The bash scripts added are a record of how the task definitions + services were created. The services could just have been copied manually, but the task definitions will need to be updated whenever the definitions for the main doc workers are updated, so it is worth scripting that.
For example, if a certain document were to fail on a new deployment of Grist, but rolling back the full deployment wasn't practical:
* Set prod2 tag in docker to desired codebase for that document
* Set desired_count for grist-docs-prod2 service to non-zero
* Set doc-<docid>-group for that doc in redis to secondary
* Hit /api/docs/<docid>/assign to move the doc to grist-docs-prod2
(If the document needs to be reverted to a previous snapshot, that currently would need doing manually - could be made simpler, but not in scope of this diff).
Test Plan: added tests
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D2649
2020-11-02 19:24:46 +00:00
|
|
|
for (const cli of this._clients || []) {
|
|
|
|
await cli.quitAsync();
|
|
|
|
}
|
2020-07-21 13:20:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
public async getElection(name: string, durationInMs: number): Promise<string|null> {
|
|
|
|
// Could use "set nx" for election, but redis docs don't encourage that any more,
|
|
|
|
// favoring redlock:
|
|
|
|
// https://redis.io/commands/setnx#design-pattern-locking-with-codesetnxcode
|
|
|
|
const redisKey = `nomination-${name}`;
|
|
|
|
const lock = await this._redlock.lock(`${redisKey}-lock`, LOCK_TIMEOUT);
|
|
|
|
try {
|
|
|
|
if (await this._client.getAsync(redisKey) !== null) { return null; }
|
|
|
|
const electionKey = uuidv4();
|
|
|
|
// seems like only integer seconds are supported?
|
|
|
|
await this._client.setexAsync(redisKey, Math.ceil(durationInMs / 1000.0), electionKey);
|
|
|
|
return electionKey;
|
|
|
|
} finally {
|
|
|
|
await lock.unlock();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
public async removeElection(name: string, electionKey: string): Promise<void> {
|
|
|
|
const redisKey = `nomination-${name}`;
|
|
|
|
const lock = await this._redlock.lock(`${redisKey}-lock`, LOCK_TIMEOUT);
|
|
|
|
try {
|
|
|
|
const current = await this._client.getAsync(redisKey);
|
|
|
|
if (current === electionKey) {
|
|
|
|
await this._client.delAsync(redisKey);
|
|
|
|
} else if (current !== null) {
|
|
|
|
throw new Error('could not remove election');
|
|
|
|
}
|
|
|
|
} finally {
|
|
|
|
await lock.unlock();
|
|
|
|
}
|
|
|
|
}
|
(core) support GRIST_WORKER_GROUP to place worker into an exclusive group
Summary:
In an emergency, we may want to serve certain documents with "old" workers as we fix problems. This diff adds some support for that.
* Creates duplicate task definitions and services for staging and production doc workers (called grist-docs-staging2 and grist-docs-prod2), pulling from distinct docker tags (staging2 and prod2). The services are set to have zero workers until we need them.
* These new workers are started with a new env variable `GRIST_WORKER_GROUP` set to `secondary`.
* The `GRIST_WORKER_GROUP` variable, if set, makes the worker available to documents in the named group, and only that group.
* An unauthenticated `/assign` endpoint is added to documents which, when POSTed to, checks that the doc is served by a worker in the desired group for that doc (as set manually in redis), and if not frees the doc up for reassignment. This makes it possible to move individual docs between workers without redeployments.
The bash scripts added are a record of how the task definitions + services were created. The services could just have been copied manually, but the task definitions will need to be updated whenever the definitions for the main doc workers are updated, so it is worth scripting that.
For example, if a certain document were to fail on a new deployment of Grist, but rolling back the full deployment wasn't practical:
* Set prod2 tag in docker to desired codebase for that document
* Set desired_count for grist-docs-prod2 service to non-zero
* Set doc-<docid>-group for that doc in redis to secondary
* Hit /api/docs/<docid>/assign to move the doc to grist-docs-prod2
(If the document needs to be reverted to a previous snapshot, that currently would need doing manually - could be made simpler, but not in scope of this diff).
Test Plan: added tests
Reviewers: dsagal
Reviewed By: dsagal
Differential Revision: https://phab.getgrist.com/D2649
2020-11-02 19:24:46 +00:00
|
|
|
|
|
|
|
public async getWorkerGroup(workerId: string): Promise<string|null> {
|
|
|
|
return this._client.getAsync(`worker-${workerId}-group`);
|
|
|
|
}
|
|
|
|
|
|
|
|
public async getDocGroup(docId: string): Promise<string|null> {
|
|
|
|
return this._client.getAsync(`doc-${docId}-group`);
|
|
|
|
}
|
2020-07-21 13:20:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// If we don't have redis available and use a DummyDocWorker, it should be a singleton.
|
|
|
|
let dummyDocWorkerMap: DummyDocWorkerMap|null = null;
|
|
|
|
|
|
|
|
export function getDocWorkerMap(): IDocWorkerMap {
|
|
|
|
if (process.env.REDIS_URL) {
|
|
|
|
return new DocWorkerMap();
|
|
|
|
} else {
|
|
|
|
dummyDocWorkerMap = dummyDocWorkerMap || new DummyDocWorkerMap();
|
|
|
|
return dummyDocWorkerMap;
|
|
|
|
}
|
|
|
|
}
|