Details
-
Improvement
-
Status: Accepted
-
Major
-
Resolution: Unresolved
-
None
-
None
Description
Right now, the catchup protocol does a full Paxos round for every log entry missing from a recovering replica. This should be unnecessary for "learned" log entries, because there is no uncertainty about the status of such log entries. We'll probably only need to do a full Paxos round for entries near the tail of the log.
To motivate changing this, it would be good to first measure the performance of the catchup protocol to understand if/when performance improvements would be useful. There's also MESOS-770, which mentions the idea of doing catchup in parallel.
Attachments
Issue Links
- is related to
-
MESOS-770 Rate control and randomization of Replicated Log catching-up
- Accepted