fix: Add retry on locks

[Fabianoshz]

what

I'm opening this as a draft to receive feedback early, I don't expect this to break anything but I believe it could be hidden behind a flag and with better default values for timeout and retries (maybe exponential retry?).

This adds a retry logic to the lock mechanism to mitigate the issue described in here and also in this ADR.

The locking issue itself is more complex and requires much more work, this is just a small step so users don't have to see the error anymore effectively making the code wait instead of asking the user to retry.

why

Currently the user has to rerun any operations that fail because a certain workspace path is locked, this tries just to automate the process.

tests

Will add if this approach receive support.

references

Did my best to try to understand which issues this would affect.

Relates to #3345
Relates to #2921
Relates to #2882

Relates to #4489
Relates to #305
Relates to #4829
Relates to #1847
Relates to #4566

Closes #1618
Closes #2200
Closes #3785
Closes #4489
Closes #4368

[Fabianoshz]

After some testing, this is what I'm seeing if I run 2 plan at the same time.

This doesn't really fix the issue, but at least reduces the amount of times users have to deal with the error messages.

[X-Guardian]

I don't see that this change is adding much value or helping on any of the issues you have linked. The only scenario where it will help is in the example you have given, i.e. running two plan commands consecutively on the same commit, which is not what most people have issues with.

[jamengual]

I don't see that this change is adding much value or helping on any of the issues you have linked. The only scenario where it will help is in the example you have given, i.e. running two plan commands consecutively on the same commit, which is not what most people have issues with.

There are a lot of people using terragrunt, atmos and other tools that can run many projects/plans at once, so I see how this help those users.

[X-Guardian]

I use Terragrunt extensively with Atlantis running parallel plans/applies in a PR, and don't have any locking issues that this change would make any difference to.

[Fabianoshz]

I don't see that this change is adding much value or helping on any of the issues you have linked. The only scenario where it will help is in the example you have given, i.e. running two plan commands consecutively on the same commit, which is not what most people have issues with.

I truly believe we still have much to do to make sure that the locking issue is dealt with. Although I'm not so sure that this code only is applicable for simultaneous commands, I've only used it to showcase the issue since is the same mechanism (TryLock). Maybe I'm overly optimistic about this...

Buy I may argue that the UX will be even better for most cases. The current behavior is:

• John tries to plan common-dir and dir-1
• While John still has the locks for common-dir and dir-1 Mary tries to plan common-dir and dir-2
• Mary receives The default workspace is currently locked by another command that is running for this pull request.

This PR will change this to:

• John tries to plan common-dir and dir-1
• While John has the locks for common-dir and dir-1 Mary tries to plan common-dir and dir-2
• When John finishes using the lock, Mary acquires the lock (given it didn't reached the timeout)

As I stated here most users are suffering from seeing the message that just says "try again later", I'm automating this step so users only see this message if we are more or less sure that there is a real issue with the plan taking too long (which can be configured by the timeout setting).

[Fabianoshz]

I've used to work in a company that had a pretty big Atlantis install, unfortunately I don't anymore so I can't really test this in large scale, I invite anyone who can test this PR to give it a try.

Two areas of improvement I already can see:

• We could evaluate if a user is trying to trigger the same command while another one is already being executed, this would allow us to avoid this scenario while proving a better UX, as I believe it's not good to let duplicated commands reach this locking mechanism.
• We could queue executions instead of just blindly retrying it as I'm doing it here now if we want to have FIFO, queueing is a bit harder though because we might have issues if the queue gets to big because external issues.