Marton Elek created RATIS-1042:
----------------------------------
Summary: Watch for commit calls are blocked for a long if no other
message
Key: RATIS-1042
URL: https://issues.apache.org/jira/browse/RATIS-1042
Project: Ratis
Issue Type: Improvement
Reporter: Marton Elek
I tested Ozone with freon key generator and teragen (2 mappers, 10G) and found
that the during the teragen execution the Ozone client spends a lot of time
waiting for the watchForCommit call.
In ozone the minimum HB timeout is 5 seconds (max is 5.2) and the HB thread in
GrpcLogAppender can sleep (min.hb.timeout / 2) if there is no outstanding
requests. But we need faster response from the followers when we wait the
results of watch for commit.
Lokesh created a patch for me (thanks) which checks if the lastCommitIndex is
the same on the follower and the leader and sends out additional heartbeats if
they don't match.
Ozone teragen 10 (mappers=2) performance is improved with 15% with this
approach.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
