tree 529ef972d87df72e4d5c484591bcbb33f3e811ab
parent 5beaf85cd5ef2b84a67ebce712e8d73d1e7d41ff
author Vladimir Golubev <vladimir.golubev@databricks.com> 1715359074 +0800
committer Wenchen Fan <wenchen@databricks.com> 1715359074 +0800

[SPARK-48143][SQL] Use lightweight exceptions for control-flow between UnivocityParser and FailureSafeParser

# What changes were proposed in this pull request?
New lightweight exception for control-flow between UnivocityParser and FalureSafeParser to speed-up malformed CSV parsing.

This is a different way to implement these reverted changes: https://github.com/apache/spark/pull/46478

The previous implementation was more invasive - removing `cause` from `BadRecordException` could break upper code, which unwraps errors and checks the types of the causes. This implementation only touches `FailureSafeParser` and `UnivocityParser` since in the codebase they are always used together, unlike `JacksonParser` and `StaxXmlParser`. Removing stacktrace from `BadRecordException` is safe, since the cause itself has an adequate stacktrace (except pure control-flow cases).

### Why are the changes needed?
Parsing in `PermissiveMode` is slow due to heavy exception construction (stacktrace filling + string template substitution in `SparkRuntimeException`)

### Does this PR introduce _any_ user-facing change?
No, since `FailureSafeParser` unwraps `BadRecordException` and correctly rethrows user-facing exceptions in `FailFastMode`

### How was this patch tested?
- `testOnly org.apache.spark.sql.catalyst.csv.UnivocityParserSuite`
- Manually run csv benchmark
- Manually checked correct and malformed csv in sherk-shell (org.apache.spark.SparkException is thrown with the stacktrace)

### Was this patch authored or co-authored using generative AI tooling?
No

Closes #46500 from vladimirg-db/vladimirg-db/use-special-lighweight-exception-for-control-flow-between-univocity-parser-and-failure-safe-parser.

Authored-by: Vladimir Golubev <vladimir.golubev@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>