andygrove opened a new issue, #4477:
URL: https://github.com/apache/datafusion-comet/issues/4477
## Describe the bug
Spark 4.1.1 added a `legacySplitTruncate` flag to `StringToMap` (and several
other split-based expressions) that is driven by
`spark.sql.legacy.truncateForEmptyRegexSplit`. When the flag is enabled (legacy
mode), `CollationAwareUTF8String.splitSQL` truncates trailing empty matches
from the split result. Comet's native `str_to_map` does not honour this flag
and always behaves as if it were `false` (the non-legacy default).
Surfaced by the map-expressions audit (collection PR queue).
## Steps to reproduce
```sql
SET spark.sql.legacy.truncateForEmptyRegexSplit = true;
SELECT str_to_map('a:1,,b:2,', ',', ':');
```
Spark 4.1.1 with the legacy flag enabled would truncate the trailing empty
`("","null")` entry; the non-legacy default and Comet keep it.
## Expected behavior
Either propagate the flag through the proto so the native impl can branch on
it, or downgrade `CometStrToMap` to `Incompatible` when
`spark.sql.legacy.truncateForEmptyRegexSplit=true` on Spark 4.1.1+.
## Additional context
- Comet wiring: `QueryPlanSerde.scala` -> `classOf[StringToMap] ->
CometStrToMap` (a one-liner over `CometScalarFunction("str_to_map")`).
- Spark 4.1.1 change: `complexTypeCreator.scala:604-606`:
```scala
private lazy val legacySplitTruncate =
SQLConf.get.getConf(SQLConf.LEGACY_TRUNCATE_FOR_EMPTY_REGEX_SPLIT)
```
- The same flag also affects `StringSplit` (already-tracked behaviour
difference there) and other split-based ops in 4.1.1.
--
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]
