Old job with Computing status in Jobs List. Invalid Ready state of activities.

[pwalski]

The problem:

• old job can stay in a Computing status
• the said Computing job does not follow list ordering (by date)

Reproduction scenario:

• start MockUI, run a job (in my case Automatic example)
• kill ya-provider process (note ya-runtime-ai does not stop when provider is killed)
• stop yagna, and click Start again
• Click List Jobs. There should be a job in a Computing state

Cause:
After killing ya-provider activity's state stays as Ready, and at no point later it gets changed.

Attachment contains Golem module's data dir.
When imported and started it should display in Jobs List an activity
6ca83ac1079122b87f460587e2c344f1909c090fbc13c74a2a7474fc12140cb1 with Computing status
golem-data.tar.gz

Solution proposal:

• After yagna restart spawn monitor_activity for all not terminated Activities: https://github.com/golemfactory/yagna/blob/60d103cfcd9194a38afd285dc20f0651a3d9e22d/core/activity/src/provider/service.rs#L197-L203
• After some period of time Activities state will be changed to Unresponsive
• After second timeout Activities will be destroyed

This solution allow to destroy activities, but doesn't close possibility of bringing activities back to live after restart.

[pwalski]

ya-provider does not support any process groups / jobs on Windows
https://github.com/golemfactory/yagna/blob/f189daee32819d0b44d3a6809e2ef4f617106042/utils/process/src/lib.rs#L66

Maybe we should move Windows process Job related code from ya-runtime-ai to some lib and use it in ya-provider?
https://github.com/golemfactory/ya-runtime-ai/blob/d5683703359165543c116e7c55e679a71a7e2c74/src/main.rs#L290

It could be moved into dedicated crate in lets say exe-unit/components/proc together with this module https://github.com/golemfactory/yagna/tree/staszek/gamerhash-combined/exe-unit/components/counters/src/os/process because it always looked odd there (and ya-counters will use ya-proc when os feature id enabled).

[pwalski]

@nieznanysprawiciel market.db from data dir I use to reproduce the bug contains an agreement in Approved state with an expired valid_to field.
Shouldn't it get changed too?
State Terminated seems most suitable, because Cancelled has a comment /// Cancelled by a Requestor, but other agreements also have Terminated state, so it would not indicate this one got interrupted.

[pwalski]

NVM