Daniel Barclay (Drill) created DRILL-4001:
---------------------------------------------
Summary: Empty vectors from previous batch left by MapVector.load(...)/RecordBatchLoader.load(...)
Key: DRILL-4001
URL: https://issues.apache.org/jira/browse/DRILL-4001
Project: Apache Drill
Issue Type: Bug
Reporter: Daniel Barclay (Drill)
In certain cases, {{MapVector.load(...)}} (called by {{RecordBatchLoader.load(...)}}) returns
with some map child vectors having a length of zero instead of having a length matching the
length of sibling vectors and the number of records in the batch. (This caused some of the
{{IndexOutOfBoundException}} errors seen in fixing DRILL-2288.)
The condition seems to be that a child field (e.g., an HBase column in a HBase column family)
appears in an earlier batch and does not appear in a later batch.
(The HBase column's child vector gets created (in the MapVector for the HBase column family)
during loading of the earlier batch. During loading of the later batch, all vectors get reset
to zero length, and then only vectors for fields _appearing in the batch message being loaded_
get loaded and set to the length of the batch-\-other vectors created from earlier messages/{{load}}
calls are left with a length of zero (instead of, say, being filled with nulls to the length
of their siblings and the current record batch).)
See the TODO(DRILL-xxxx) mark and workaround in {{MapVector.getObject(int)}}.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
|