[
https://issues.apache.org/jira/browse/KUDU-3820?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Gabriella Lotz updated KUDU-3820:
---------------------------------
Description:
Both the per-table pass (RunLeaderRebalancer) and the global pass
(RunGlobalLeaderRebalance) go through the tables in the same order every round
and stop once the budget is used up. The table list comes
from an unordered_map in the catalog manager, so the order isn't meaningful,
but as far as I can tell it also doesn't change from one round to the next
unless the map is resized.
That means tables at the front of the list can use the whole budget every
round, for example if they keep getting skewed by leader elections after
restarts. Tables at the end may never be reached, so their leaders stay
unbalanced for as long as the earlier tables need moves.
Example: with tables A, B, C, D, E and a cap of 10, if A and B need 10 or more
moves every round, C, D and E are skipped every round.
was:
Both the per-table pass (RunLeaderRebalancer) and the global pass
(RunGlobalLeaderRebalance) go through the tables in the same order
every round and stop once the budget is used up. The table list comes
from an unordered_map in the catalog manager, so the order isn't
meaningful, but as far as I can tell it also doesn't change from one
round to the next unless the map is resized.
That means tables at the front of the list can use the whole budget
every round, for example if they keep getting skewed by leader
elections after restarts. Tables at the end may never be reached, so
their leaders stay unbalanced for as long as the earlier tables need
moves.
Example: with tables A, B, C, D, E and a cap of 10, if A and B need
10 or more moves every round, C, D and E are skipped every round.
> Rotate table order in auto leader rebalancer rounds
> ---------------------------------------------------
>
> Key: KUDU-3820
> URL: https://issues.apache.org/jira/browse/KUDU-3820
> Project: Kudu
> Issue Type: Improvement
> Reporter: Gabriella Lotz
> Priority: Major
>
> Both the per-table pass (RunLeaderRebalancer) and the global pass
> (RunGlobalLeaderRebalance) go through the tables in the same order every
> round and stop once the budget is used up. The table list comes
> from an unordered_map in the catalog manager, so the order isn't meaningful,
> but as far as I can tell it also doesn't change from one round to the next
> unless the map is resized.
> That means tables at the front of the list can use the whole budget every
> round, for example if they keep getting skewed by leader elections after
> restarts. Tables at the end may never be reached, so their leaders stay
> unbalanced for as long as the earlier tables need moves.
> Example: with tables A, B, C, D, E and a cap of 10, if A and B need 10 or
> more moves every round, C, D and E are skipped every round.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)