[
https://issues.apache.org/jira/browse/KUDU-3791?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18119107#comment-18119107
]
Gabriella Lotz commented on KUDU-3791:
--------------------------------------
The metrics half of this is done: f54a74dbd landed on master, adding the six
auto-leader-rebalancer counters and removing the "Add metrics" TODOs.
I'm going to leave the remaining "Make step-down RPCs async" TODO out of scope
for this ticket. I don't think it's a small change: the transfer loop would
need a callback per RPC, a way to wait on the outstanding calls before the
round finishes, and the completed/failed accounting would move into those
callbacks, so the scheduled = completed + failed invariant would have to hold
under concurrency too.
It also doesn't look urgent. leader_rebalancing_max_moves_per_round defaults to
10 and auto_leader_rebalancing_rpc_timeout_seconds to 10, so a round's
step-down phase is bounded at roughly 100s out of a 3600s cycle
(auto_leader_rebalancing_interval_seconds). I believe that makes it a latency
improvement rather than a correctness issue, and the new counters would make it
easy to verify an async version isn't dropping moves.
I've left the TODO in the file with a note on the reasoning. I have filed
KUDU-3808 for it, closing this current ticket.
> Address TODOs in auto_leader_rebalancer.cc
> ------------------------------------------
>
> Key: KUDU-3791
> URL: https://issues.apache.org/jira/browse/KUDU-3791
> Project: Kudu
> Issue Type: Sub-task
> Reporter: Gabriella Lotz
> Assignee: Gabriella Lotz
> Priority: Major
>
> # Add metrics line: 472, 834, 842
> # Make step-down RPCs async line: 410
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
