Johnny Miller created CASSANDRA-21719:
-----------------------------------------
Summary: ANN query responses fail to serialize when the table has
static columns outside the selection
Key: CASSANDRA-21719
URL: https://issues.apache.org/jira/browse/CASSANDRA-21719
Project: Apache Cassandra
Issue Type: Bug
Components: Feature/SAI
Reporter: Johnny Miller
On a replica serving an ANN query, StorageAttachedIndexSearcher wraps the top-k
results in InMemoryUnfilteredPartitionIterator
(StorageAttachedIndexSearcher.java line 149 on trunk). That iterator reports
its columns as command.metadata().regularAndStaticColumns() instead of the
columns the command actually fetched (InMemoryUnfilteredPartitionIterator.java
line 115, InMemoryPartitionIterator.java line 107, same on cassandra-5.0). The
response serializer builds its header from iterator.columns()
(UnfilteredRowIteratorSerializer.java lines 101 to 104), and
SerializationHeader.serializeForMessaging requires those columns to be a subset
of what the command fetched. When the table has a static column the query did
not select, that check fails and the read errors with
IllegalStateException("... is not a subset of ...") from Columns.java line 618.
For example, SELECT s1 with an ANN ORDER BY on a table that also has a static
column s2.
The DataStax fork fixed this by returning the base iterator's columns,
datastax/cassandra@ddbb38014c (their CNDB-14997). PartitionInfo.java here is
byte for byte the fork's pre-fix version. Their regression test covers
fork-only features (generic ORDER BY, BM25) so a port would need a trimmed
test. I could not find an equivalent fix or an existing ticket here.
Verified on cassandra-5.0 and trunk.
Happy to port it the same way as CASSANDRA-21615 if wanted.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
