Description
I found two bugs
- org.apache.orc.impl.RecordReaderUtils#zeroCopyReadRanges
ByteBuffer copy = currentBuffer.duplicate(); copy.position((int) (current.getOffset() - currentOffset));
if currentBuffer position is not 0, copy.position() will set a uncorrect position.
I think we should use slice() replace duplicate().
- org.apache.orc.impl.ZeroCopyShims.ZeroCopyAdapter#releaseBuffer
The parameter `buffer` is not the original buffer from hdfs reading. It is the slice or duplicate of the original buffer.
This bug is reported at ORC-700.
Attachments
Issue Links
- duplicates
-
ORC-700 StripePlanner buffer management issues with zero copy
- Open
- links to