ACI compute-worker feature

[igormusinov]

• aci_compute_worker (ACI https://azure.microsoft.com/en-us/services/container-instances/)
• logging with config
• alternative queue bug fixed

Tested https://codalab.coresearch.club/my/competition/76/submissions/?page=3&phase=173

[igormusinov]

Hi @ckcollab ! I would be happy to commit to this project. Here is our team's instance of the codalab https://codalab.coresearch.club/ . Could you please guide me what I should do first to make a commit ?

[ckcollab]

This seems like it should become a fork instead of get merged upstream? We don't use this on Codalab production.

[ckcollab]

Just seeing your comment now, I think this is fine stuff but a different approach than we use. As separate repos seems to make sense. I could add an "alternative workers" section? Version 2 of Codalab is here with completely different workers:
https://github.com/codalab/competitions-v2/tree/develop/docker/compute_worker

[igormusinov]

I think this is fine stuff but a different approach than we use

something like that , but at the same time there is some fixes with old code (alternative queue bug, logs). And I didn't change a lot of code for local worker

"I could add an "alternative workers" section"

it would be cool

Version 2 of Codalab is here with completely different workers:
https://github.com/codalab/competitions-v2/tree/develop/docker/compute_worker

cool , but is compute_worker-v2 backward compatible with v1 ?

[ckcollab]

I could add an "alternative workers" section

it would be cool

Added!

cool , but is compute_worker-v2 backward compatible with v1 ?

The worker expects to be used quite differently, but v1 competitions should port over to v2 fine. I'm sure there are still many small discrepancies, v2 is still very early in actual usage, but we've made quite a few competitions work properly.

Main benefits of the new worker: API back to Codalab during runs + websockets to talk back to Codalab to give real time stdout and whatever features we can dream up. I.e. realtime error rate updates, status updates (submitted, running, finished, etc.) so you don't have to refresh the page, sending back chart data to the frontend, python 3, etc.