andygrove opened a new issue, #5776:
URL: https://github.com/apache/datafusion-comet/issues/5776
### Describe the bug
With `spark.comet.write.iceberg.splitOperator.enabled` and
`spark.comet.iceberg.write.enabled` both on (#5644, experiment PR #5677), the
upstream Iceberg test
`TestMetadataTablesWithPartitionEvolution.testPartitionColumnNamedPartition`
fails because an unordered `SELECT *` returns its rows in a different order
than iceberg-java produces:
```
org.opentest4j.AssertionFailedError: [Should return all expected rows: row 1
col 1 contents should match]
expected: 1
but was: 3
at
org.apache.iceberg.spark.SparkTestHelperBase.assertEquals(SparkTestHelperBase.java:85)
at
org.apache.iceberg.spark.source.TestMetadataTablesWithPartitionEvolution.testPartitionColumnNamedPartition(TestMetadataTablesWithPartitionEvolution.java:652)
```
The test writes four rows across two partitions and compares the result of
`SELECT *` positionally:
```java
sql("CREATE TABLE %s (id int, partition int) USING iceberg PARTITIONED BY
(partition)", tableName);
sql("INSERT INTO %s VALUES (1, 1), (2, 1), (3, 2), (2, 2)", tableName);
List<Object[]> expected = ImmutableList.of(row(1, 1), row(2, 1), row(3, 2),
row(2, 2));
assertEquals("Should return all expected rows", expected, sql("SELECT * FROM
%s", tableName));
```
The first row read back is `(3, ...)` rather than `(1, 1)`, so the
`partition = 2` data file is coming back first.
### Steps to reproduce
Run the Iceberg Spark SQL Tests jobs with both write flags defaulted on.
Observed in [run
33995765073](https://github.com/apache/datafusion-comet/actions/runs/33995765073)
at commit `d86d3b5`:
- `iceberg-spark/iceberg-1.11.0/spark-4.1.3`: 5 parameter combinations fail
(`fileFormat = ORC/formatVersion 1`, `PARQUET/1`, `PARQUET/2`, `AVRO/1`,
`AVRO/2`, across the hadoop, hive and hive-with-`parquet-enabled=false`
catalogs)
- `iceberg-spark/iceberg-1.9.1/spark-3.5.9`: 2 combinations fail
(`fileFormat = ORC`, formatVersion 1 and 2, testhive catalog)
- The 1.10.0/3.5 and 1.8.1/3.4 jobs do not fail this test
### Expected behavior
Either the write path produces the same data-file layout and read-back order
as iceberg-java so the upstream assertion holds, or we establish that the
assertion is order-dependent and adapt it in `dev/diffs/iceberg/*.diff` (never
by disabling Comet).
### Additional context
Needs triage before it can be fixed: the failing set includes `ORC` and
`AVRO` file formats, where the native Parquet writer cannot be the one
producing the files, and it covers only a subset of the parameter matrix while
other versions pass the same test. That pattern is as consistent with an
order-dependent assertion that Comet makes unstable (read-side task or file
ordering) as with a writer file-layout divergence, so the first step is to
determine which side changes the order.
Part of the native Iceberg writes epic, #5649, and a blocker for #5644.
Surfaced by the failure inventory on #5677.
--
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]
