Ariel Weisberg created CASSANDRA-20589:
------------------------------------------
Summary: Improve consistency of non-SERIAL mutations on Accord
Key: CASSANDRA-20589
URL: https://issues.apache.org/jira/browse/CASSANDRA-20589
Project: Apache Cassandra
Issue Type: Improvement
Components: Accord
Reporter: Ariel Weisberg
Right now all non-SERIAL mutations on Accord use the coordinator generated
timestamp not the Accord generated timestamp making them not linearizable. This
isn’t a regression, but it could maybe be improved to be linearizable in more
cases.
One complication is that `USING TIMESTAMP` needs to be detected and respected
by Accord and `USING TIMESTAMP` can be per statement in both a logged and
unlogged batch. Additionally this information is not available anymore at the
`StorageProxy` layer where coordinating mutations takes place.
Hints and batch log will repeatedly apply operations and they need to be
idempotent so those should always re-use the same timestamp otherwise you risk
data resurrection. Additionally the underlying update might be applied as a
transaction, and then it might be applied again as a hint/batchlog replay so
the correctness of this approach needs to be thoroughly evaluated since we are
now mixing Accord and coordinator generated timestamps in the non-SERIAL write
path.
--
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
