andygrove commented on issue #6214:
URL:
https://github.com/apache/datafusion-comet/issues/6214#issuecomment-5833349726
Two unrelated failures in this run.
The Spark 4.0 `sql_core-1` failure is `CollationSuite` "hash agg is not used
for non binary collations" (`None was empty (CollationSuite.scala:467)`), the
assertion that the `UTF8_BINARY` `GROUP BY` plan contains a Spark
`HashAggregateExec`. It comes from #6195 and will fail every nightly until the
diff changes. Spark 4.0's `DatasourceV2SQLBase` runs
`spark.sessionState.conf.clear()` after each test, which drops
`spark.shuffle.manager` from the session conf. Before #6195, Comet read the
shuffle manager from there, so Comet shuffle was silently off for every test
after the first one in those suites, and this query's final aggregate stayed on
Spark. #6195 checks the manager that `SparkEnv` holds instead, so Comet shuffle
now plans the exchange and the final aggregate is a `CometHashAggregateExec`.
Spark 4.1's `DatasourceV2SQLBase` only unsets two catalog configs, so 4.1
always saw this, and the 4.1.3 diff already accepts `CometHashAggregateExec` in
that assertion. #6195 ran only the
4.1 suite. The new behaviour is correct, so #6220 regenerates the 4.0.4 diff
to accept `CometHashAggregateExec` the same way, with `run-spark-4.0-tests`
applied.
One side effect: every `DatasourceV2SQLBase` suite on 3.4, 3.5 and 4.0 now
runs Comet shuffle after its first test, where it used to run Spark's. The 3.5
suite and the other 4.0 shards passed in this run. The 3.4 suite runs only with
its label or a manual dispatch, and it hasn't run since #6195 landed.
The Iceberg 1.9 shard-2 failure is
`TestRewriteDataFilesAction.testParallelPartialProgressWithMaxFailedCommitsLargerThanTotalFileGroup`
at `formatVersion = 3` ("1 rewrite commits failed. This is more than the
maximum allowed failures of 0"), a known upstream flake, apache/iceberg#12889.
Three threads commit ten file groups one at a time against a Hadoop table. In
the job's JUnit output, one thread committed eight groups back to back while
group 3's commit backed off four times and then lost a fifth attempt on stale
metadata, past Iceberg's default of four commit retries. That all happens in
Iceberg's commit code, not Comet's. Upstream fixed the test in
apache/iceberg#13208 and apache/iceberg#13598. Both are in 1.10.0 but not
1.9.1, which is why the 1.10 shard-2 passed on the same commit. This is the
first time it has failed in the last ten nightlies. If it keeps recurring, we
could backport those two test edits into the 1.9.1 diff.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
