Log job failure even when there are retries configured

[wxtim]

Closes #6151

Check List

• I have read CONTRIBUTING.md and added my name as a Code Contributor.
• Contains logically grouped changes (else tidy your branch by rebase).
• Does not contain off-topic changes (use other PRs for other changes).
• Applied any dependency changes to both setup.cfg (and conda-environment.yml if present).
• Tests are included (or explain why tests are not needed).
• CHANGES.md entry included if this is a change that can affect users
• Cylc-Doc pull request opened if required at cylc/cylc-doc/pull/XXXX.
• If this is a bug fix, PR should be raised against the relevant ?.?.x branch.

[oliver-sanders]

Unfortunately, I don't think it's this simple.

• I think this diff means that polled log messages for task failure will go back to being duplicated.
• This only covers failure, but submission failure also has retries so may require similar treatment.
• I think the failures before retries are exhausted will now get logged at CRITICAL level rather than INFO.

[wxtim]

I think this diff means that polled log messages for task failure will go back to being duplicated.

I think that you have a particular closed issue in mind, but I can't find it... Can you point it out to me?

This only covers failure, but submission failure also has retries so may require similar treatment.

I think that submission failure is already handled correctly - it certainly is in the simplistic case where you feed it platform=garbage you get a submission failed message at level critical:

CRITICAL - [1/bar/01:preparing] submission failed
CRITICAL - [1/bar/02:preparing] submission failed
CRITICAL - [1/bar/03:preparing] submission failed

These are logged at critical - and I think they should be?

I think the failures before retries are exhausted will now get logged at CRITICAL level rather than INFO.

This would be consistent with submit failure...

[oliver-sanders]

I think this diff means that polled log messages for task failure will go back to being duplicated.

I think that you have a particular closed issue in mind, but I can't find it... Can you point it out to me?

No, I'm not thinking of the other log message duplication issue.

The change made here bypassed logic that was used for suppressing duplicate log messages (the f'{FAIL_MESSAGE_PREFIX}ERR' bit):

8f20ab0#diff-d6de42ef75ecc801c26a6be3a9dc4885b64db89d32bce6f07e319595257b9b2eL930

However, in your more recent "fix" commit, you have put this back the way it was before:

3cedf2f#diff-d6de42ef75ecc801c26a6be3a9dc4885b64db89d32bce6f07e319595257b9b2eR930

because I'm blind

[wxtim]

This does not apply to submit failure, because submit failure will always log a critical warning through the jobs-submit command.

WARNING - platform: mulberry - Could not connect to mymachine.
    * mymachine has been added to the list of unreachable hosts
    * remote-init will retry if another host is available.
ERROR - [jobs-submit cmd] (remote init)
    [jobs-submit ret_code] 1
CRITICAL - [1/bar/02:preparing] submission failed
INFO - [1/bar/02:preparing] => waiting
WARNING - [1/bar:waiting] retrying in PT1S (after 2024-07-22T14:51:37+01:00)

[MetRonnie]

(Test failures)

[wxtim]

I think that the correct level is error.

@hjoliver - does your PR fall victim to any of Oliver's comments from #6169 (review)?

[MetRonnie]

If execution/submission retry delays are configured, then execution/submission failures (respectively) are expected to occur. Therefore it is not a CRITICAL message to log. Only if the retries are exhausted should it be a CRITICAL level message?

[hjoliver]

If execution/submission retry delays are configured, then execution/submission failures (respectively) are expected to occur. Therefore it is not a CRITICAL message to log. Only if the retries are exhausted should it be a CRITICAL level message?

I don't disagree with that, but it was kinda off-topic for this Issue - which is about not hiding the job failure from the log - unless we introduce a jarring inconsistency between the way submission and execution failures are logged.

But OK, if we want to two kill two birds with one stone, let's look at unhiding the job failure AND changing the level of both kinds of failure at once, to maintain consistency...

I agree with @wxtim 's assertion that the correct level (for both) is ERROR.

[hjoliver]

let's look at unhiding the job failure AND changing the level of both kinds of failure at once, to maintain consistency...

In that case, I would go with your approach @wxtim - but with some tweaks:

1. don't log failure messages at all in the early _process_message_check method (where the demotion to DEBUG is currently done)

2. instead log for both retries and not-retries in _process_message_failed where we already have the right code blocks:

if retries:
   LOG.error(message)
   ...
else:
   LOG.critical(message)
   ...

3. and do exactly the same for job submit failure, in process_message_submit_failed

[UPDATE]
Actually, it should be ERROR both ways:

LOG.error(message)
if retries:
   ....
else:
   ...

(a job error is always an error, but whether it's critical or not depends on the graph... if it is, it will cause a stall that does get logged as CRITICAL)

stale

[hjoliver]

@wxtim - is this ready to go again, in light of the above discussion?

[wxtim]

@wxtim - is this ready to go again, in light of the above discussion?

In my head, yes, but I see that there a load of test failures. Will draft now and undraft when ready.

[wxtim]

These test failures were caused by a slight change in the nature of the message caused by moving it: By the time the code reaches _process_message_failed some of the info required for the original error messages (the state) is not available, or has changed.

[hjoliver]

One small typo found.

[hjoliver]

@wxtim this seems to have diverged a bit from what I thought was agreed above, which was:

• log the job (or submit) failure as ERROR whether there are retries or not
• then log the retry message at WARNING level, if there are retries

Now, if there are retries we only get the retry warning. (Which I think is back to the problem we were trying to fix here, although the logging location has changed to the methods that would support the fix).