Make RPCTimeout be as close to "max time to respond to client" as possible
--------------------------------------------------------------------------
Key: CASSANDRA-3018
URL: https://issues.apache.org/jira/browse/CASSANDRA-3018
Project: Cassandra
Issue Type: Improvement
Reporter: Sylvain Lebresne
Priority: Minor
Quoting a comment of CASSANDRA-2494:
{quote}
rpctimeout is working exactly as intended, i.e., to prevent waiting
indefinitely for a node that died after we sent it a request. Treating it as
"max time to respond to client" has never really been correct. (E.g., in the CL
> ONE case we can already wait up to rpctimeout twice, one for the original
digest read set, and again for the data read after mismatch.)
{quote}
Let's make treating it as "max time to respond to client" more correct. It
should be too hard (we just need to grab the time at the start and make every
wait depend on how long we still have), and it will be more useful. It probably
cannot be perfect, but the goal is not to aim for sub-milisecond precision, but
to avoid the "waiting up twice the value 'asked' for".
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
