SQL queries return unicode strings while CMSSW only accepts byte strings about t0 HOT 7 CLOSED

T0 uses the same database access code as WMCore. With convert_unicode deprecated, how is this done now? Are we silently using unicode for any string return from a database call now and we only notice in the Tier0 since it passed some of these to CMSSW?

from t0.

As I understand, the convert_unicode was deprecated in the latest SQLAlchemy upgrade in April
cms-sw/cmsdist@2187529#diff-c359e72e06d238c2f7e059ec6f8c3c3cL1

Therefore, all old T0 releases used byte strings returned from database call and everything worked well (since the old SQLAlchemy release was used).
Only starting from the new Tier0 release (the one currently being tested) we are using unicode strings from any database call.
And yes, we only noticed that in a release test replay as some of these unicode strings got passed to CMSSW and caused all the jobs to fail.

Overall, I am wondering whether we should review these database calls more carefully or should we just go with the fix I am testing now, if it works without errors?

from t0.

I think converting every single DAO where we pull strings out of a database individually is the wrong approach. Maybe some central place would be better. I would coordinate this with @amaltaro , since this really is a more general WMCore problem.

from t0.

I think the generic fix is what DBFormatter format functions provide. Since you have to format every single DAO output, you just call one of those methods.
We might still have to improve something there - like casting the key as well, if needed - but I'd rather have this conversion only done where we need; I fear having that conversion for every single DAO in the fetch* method, for instance, might degrade the agent performance by a non-negligible factor.

from t0.

I agree with Alan that the DBFormatter is a more central/generic place for these checks.
I adjusted my changes accordingly, to use the formatDict method to convert results:
master...vytjan:deployment-updates
It seems the GetStreamDatasetTriggers DAO is the only one where we hit this unicode string incompatibility.
Maybe it could be enough to apply the patch I created and leave all the other DAOs as they are? @amaltaro Alan, what do you think?
In such case we wouldn't need to convert every single DAO result of unicode->byte string as Python handles them fine and there seems to be no other issues with T0 WMAgent of this kind.

from t0.

Your change looks fine to me, Vytas.

from t0.

Thanks Alan. Tested and merged #4503

from t0.