Shubham Chaurasia created HIVE-23034:
----------------------------------------
Summary: Arrow serializer should not keep the reference of arrow
offset and validity buffers
Key: HIVE-23034
URL: https://issues.apache.org/jira/browse/HIVE-23034
Project: Hive
Issue Type: Bug
Components: llap, Serializers/Deserializers
Reporter: Shubham Chaurasia
Assignee: Shubham Chaurasia
Currently, a part of writeList() method in arrow serializer is implemented like
-
{code:java}
final ArrowBuf offsetBuffer = arrowVector.getOffsetBuffer();
int nextOffset = 0;
for (int rowIndex = 0; rowIndex < size; rowIndex++) {
int selectedIndex = rowIndex;
if (vectorizedRowBatch.selectedInUse) {
selectedIndex = vectorizedRowBatch.selected[rowIndex];
}
if (hiveVector.isNull[selectedIndex]) {
offsetBuffer.setInt(rowIndex * OFFSET_WIDTH, nextOffset);
} else {
offsetBuffer.setInt(rowIndex * OFFSET_WIDTH, nextOffset);
nextOffset += (int) hiveVector.lengths[selectedIndex];
arrowVector.setNotNull(rowIndex);
}
}
offsetBuffer.setInt(size * OFFSET_WIDTH, nextOffset);
{code}
1) Here we obtain a reference to {{final ArrowBuf offsetBuffer =
arrowVector.getOffsetBuffer();}} and keep updating the arrow vector and offset
vector.
Problem -
{{arrowVector.setNotNull(rowIndex)}} keeps checking the index and reallocates
the offset and validity buffers when a threshold is crossed, updates the
references internally and also releases the old buffers (which decrements the
buffer reference count). Now the reference which we obtained in 1) becomes
obsolete. Furthermore if try to read or write old buffer, we see -
{code:java}
Caused by: io.netty.util.IllegalReferenceCountException: refCnt: 0
at
io.netty.buffer.AbstractByteBuf.ensureAccessible(AbstractByteBuf.java:1413)
at io.netty.buffer.ArrowBuf.checkIndexD(ArrowBuf.java:131)
at io.netty.buffer.ArrowBuf.chk(ArrowBuf.java:162)
at io.netty.buffer.ArrowBuf.setInt(ArrowBuf.java:656)
at
org.apache.hadoop.hive.ql.io.arrow.Serializer.writeList(Serializer.java:432)
at
org.apache.hadoop.hive.ql.io.arrow.Serializer.write(Serializer.java:285)
at
org.apache.hadoop.hive.ql.io.arrow.Serializer.writeStruct(Serializer.java:352)
at
org.apache.hadoop.hive.ql.io.arrow.Serializer.write(Serializer.java:288)
at
org.apache.hadoop.hive.ql.io.arrow.Serializer.writeList(Serializer.java:419)
at
org.apache.hadoop.hive.ql.io.arrow.Serializer.write(Serializer.java:285)
at
org.apache.hadoop.hive.ql.io.arrow.Serializer.serializeBatch(Serializer.java:205)
{code}
Solution -
This can be fixed by getting the buffers each time (
{{arrowVector.getOffsetBuffer()}} ) we want to update them.
In our internal tests, this is very frequently seen on arrow 0.8.0 but not on
0.10.0 but should be handled the same way for 0.10.0 too as it does the same
thing.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)