[
https://issues.apache.org/jira/browse/SPARK-59832?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
L. C. Hsieh updated SPARK-59832:
--------------------------------
Docs Text:
The vectorized Parquet reader now fails with a {{ParquetDecodingException}} on
corrupt pages whose RLE/bit-packed data (definition and repetition levels,
dictionary ids, or RLE-encoded booleans) ends early, or has an invalid length,
bit width or bit-packed run header. Previously such pages could silently return
wrong values, make the task hang, or allocate a very large buffer. Valid
Parquet files are not affected.
A few corrupt or non-compliant pages that used to read without an error, and
gave the right result by chance, now fail:
* Dictionary ids or RLE boolean values that end early, when every row after the
end of the data is skipped by column index filtering. This was possible since
Spark 3.3.0; Spark 3.2 also failed on these pages. Selecting rows after the end
of the data returned wrong values before this change.
* A dictionary-encoded page whose data section is empty, without the bit width
byte that the Parquet spec requires, when all its non-null rows are skipped, or
when the dictionary has a single entry. Before this change, every value of such
a page was read as dictionary id 0.
These pages are now treated as corrupt. By default the query fails. With
{{spark.sql.files.ignoreCorruptFiles=true}}, Spark skips the rest of such a
file, starting from the batch that contains the page, and logs a warning.
Labels: pull-request-available release-notes (was:
pull-request-available)
> Fail on truncated input in the vectorized RLE reader
> ----------------------------------------------------
>
> Key: SPARK-59832
> URL: https://issues.apache.org/jira/browse/SPARK-59832
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 3.5.8, 4.2.0, 4.3.0, 4.1.3, 4.0.4, 4.4.0
> Reporter: L. C. Hsieh
> Assignee: L. C. Hsieh
> Priority: Minor
> Labels: pull-request-available, release-notes
> Fix For: 4.4.0, 4.2.2, 4.1.5, 4.0.6, 4.3.1
>
>
> Follow-up of SPARK-59812. The vectorized RLE/bit-packing hybrid reader
> (VectorizedRleValuesReader), used for definition/repetition levels,
> dictionary ids and booleans, does not fail when the encoded data ends before
> the requested values are read:
> * When readNextGroup() returns false at the end of the stream, the batch read
> and skip loops silently break, so dictionary ids and levels are only
> partially read or skipped.
> * readBooleans ignores the return value of readNextGroup(). At the end of the
> stream currentCount stays 0, so the loop makes no progress and never
> terminates.
> * readInteger() also ignores the return value of readNextGroup().
> * The length passed to sliceStream(length) for levels, and the PACKED run
> size `numGroups * bitWidth`, are read from the file without validation.
> The reader should fail with a ParquetDecodingException in these cases, like
> the other vectorized readers now do.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
