Ariel Weisberg created CASSANDRA-19419:
------------------------------------------
Summary: Non-transactional schema updates can interfere with
Accord transaction execuion
Key: CASSANDRA-19419
URL: https://issues.apache.org/jira/browse/CASSANDRA-19419
Project: Cassandra
Issue Type: Bug
Reporter: Ariel Weisberg
While Accord can handle topology changes correctly it can’t handle
non-transaction schema updates because those execute outside of Accord. When
Accord tries to execute a transaction against the schema in the epoch the
transaction is supposed to execute in then it is possible for different nodes
to see different schemas when reading or writing data as part of a transaction.
Dropping a needed a column or table is the most likely issue as we don't
support altering column types.
Because commit is async it is possible for a table or to be dropped before the
writes can be propagated after it was acknowledged instead of signaling an
error. While the table was dropped it's possible the client needed the error to
know that the request was processed improperly or that it needed to take some
other action client side.
Or add table where the original coordinator can't read the table, but the
recovery coordinator can and might apply different results to different
replicas.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
