Kris Mok created SPARK-39839:
--------------------------------
Summary: Handle special case of null variable-length Decimal with
non-zero offsetAndSize in UnsafeRow structural integrity check
Key: SPARK-39839
URL: https://issues.apache.org/jira/browse/SPARK-39839
Project: Spark
Issue Type: Bug
Components: SQL
Affects Versions: 3.3.0, 3.2.0, 3.1.0
Reporter: Kris Mok
The {{UnsafeRow}} structural integrity check in
{{UnsafeRowUtils.validateStructuralIntegrity}} is added in Spark 3.1.0. It’s
supposed to validate that a given {{UnsafeRow}} conforms to the format that the
{{UnsafeRowWriter}} would have produced.
Currently the check expects all fields that are marked as null should also have
its field (i.e. the fixed-length part) set to all zeros. It needs to be updated
to handle a special case for variable-length {{{}Decimal{}}}s, where the
{{UnsafeRowWriter}} may mark a field as null but also leave the fixed-length
part of the field as {{OffsetAndSize(offset=current_offset, size=0)}}. This may
happen when the {{Decimal}} being written is either a real {{null}} or has
overflowed the specified precision.
Logic in {{UnsafeRowWriter}}:
in general:
{code:scala}
public void setNullAt(int ordinal) {
BitSetMethods.set(getBuffer(), startingOffset, ordinal); // set null bit
write(ordinal, 0L); // also zero out
the fixed-length field
} {code}
special case for {{DecimalType}}:
{code:scala}
// Make sure Decimal object has the same scale as DecimalType.
// Note that we may pass in null Decimal object to set null for it.
if (input == null || !input.changePrecision(precision, scale)) {
BitSetMethods.set(getBuffer(), startingOffset, ordinal); // set null bit
// keep the offset for future update
setOffsetAndSize(ordinal, 0); // doesn't
zero out the fixed-length field
} {code}
The special case is introduced to allow all {{DecimalType}}s (including both
fixed-length and variable-length ones) to be mutable – thus need to leave space
for the variable-length field even if it’s currently null.
Note that this special case in {{UnsafeRowWriter}} has been there since Spark
1.6.0, where as the integrity check was added in Spark 3.1.0. The check was
originally added for Structured Streaming’s checkpoint evolution validation, so
that a newer version of Spark can check whether or not an older checkpoint file
for Structured Streaming queries can be supported, and/or if the contents of
the checkpoint file is corrupted.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
