Fix flaky transparent_decompress_chunk

[mkindahl]

It can pick merge joins in favor of hash joins when testing outer joins. Since these tests are only testing outer joins and currently uses hash joins, we ensure that it will consistently pick those by disabling merge joins for these queries only. Merge join is used for other queries, so we do not disable merge join for the entire test.

Disable-check: force-changelog-file

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 92.29%. Comparing base (59f50f2) to head (88e0765).
Report is 355 commits behind head on main.

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #7288       +/-   ##
===========================================
+ Coverage   80.06%   92.29%   +12.22%     
===========================================
  Files         190      205       +15     
  Lines       37181    38538     +1357     
  Branches     9450     9997      +547     
===========================================
+ Hits        29770    35568     +5798     
+ Misses       2997     2970       -27     
+ Partials     4414        0     -4414     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[erimatnor]

This is fine to stabilize the test.

However, this seems like a symptom of a larger problem we've seen with compression where it doesn't have stable plans because (column) stats don't really work.

Normally, an analyze should stabilize tests like this after stats are updated.

Stats is not something we can fix now, however, so disabling certain joins is the only way. But I think it is still important to highlight the larger issue we need to address that can lead to bad query plans.

[akuzm]

Normally, an analyze should stabilize tests like this after stats are updated.

Stats is not something we can fix now, however, so disabling certain joins is the only way. But I think it is still important to highlight the larger issue we need to address that can lead to bad query plans.

It does stabilize, you have to run it after compression. This uncovers so many problems though that it's gonna take a month to fix, and the problems are not easy, many things are broken in DecompressChunk costing or even in upstream costs. I tried before #6550

[mkindahl]

This is fine to stabilize the test.

However, this seems like a symptom of a larger problem we've seen with compression where it doesn't have stable plans because (column) stats don't really work.

Normally, an analyze should stabilize tests like this after stats are updated.

Yeah, but it require a more elaborate update of the test, so this is just an immediate fix to avoid the current problem.

Stats is not something we can fix now, however, so disabling certain joins is the only way. But I think it is still important to highlight the larger issue we need to address that can lead to bad query plans.

I think we need to go over these tests and actually try to test DecompressChunk with all kinds of joins since they have slightly different behavior and requirements on what fields are set, so there can be bugs hiding here, but that is a bigger effort.

[mkindahl]

Normally, an analyze should stabilize tests like this after stats are updated.
Stats is not something we can fix now, however, so disabling certain joins is the only way. But I think it is still important to highlight the larger issue we need to address that can lead to bad query plans.

It does stabilize, you have to run it after compression. This uncovers so many problems though that it's gonna take a month to fix, and the problems are not easy, many things are broken in DecompressChunk costing or even in upstream costs. I tried before #6550

I suspected this, but did not dig too much into it.