L. C. Hsieh created SPARK-59831:
-----------------------------------
Summary: Validate bounds in the vectorized BYTE_STREAM_SPLIT reader
Key: SPARK-59831
URL: https://issues.apache.org/jira/browse/SPARK-59831
Project: Spark
Issue Type: Bug
Components: SQL
Affects Versions: 5.0.0
Reporter: L. C. Hsieh
Follow-up of SPARK-59812. The vectorized BYTE_STREAM_SPLIT reader
(VectorizedByteStreamSplitValuesReader, added by SPARK-56894) does not validate
its position against the page, so a corrupt page can be decoded silently:
* The skip methods (skipIntegers, skipLongs, skipFloats, skipDoubles,
skipBytes, skipShorts, skipBinary, skipFixedLenByteArray) only do `offset +=
total`, without checking that the page has that many values. The read paths
index pageData without a bound check either, so reading past the end reads
bytes from the next stream or fails with an ArrayIndexOutOfBoundsException
instead of a ParquetDecodingException.
* initFromPage derives the value count as `totalBytes / typeWidth` and does not
reject `totalBytes % typeWidth != 0`. For such a page, every stream starts at
the wrong offset, so every value is silently decoded from the wrong bytes.
SPARK-56894 is in branch-4.3 and v4.3.0-rc1, so this should be fixed before
4.3.0 is released.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
