Summary:
Recent python3 changes perturbed timing again, and a few more tests started failing.
Contains an unrelated correction for gvisor running under docker (a useful configuration on macs for debugging gvisor problems, but not supported by throttling code).
Test Plan: updated tests
Reviewers: dsagal, alexmojaki
Reviewed By: dsagal, alexmojaki
Differential Revision: https://phab.getgrist.com/D3129
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
