[
https://issues.apache.org/jira/browse/HBASE-24139?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17078260#comment-17078260
]
Sean Busbey commented on HBASE-24139:
-------------------------------------
I am inclined to take the approach of adding a short circuit to
{{needsBalance}} that says to look at plans if there are any region servers
reporting 0 regions.
The down side to this is that if someone sets the "balancer per table" config
to true we'll run balancer plans every possible time (every ~5 minutes by
default) for every table that doesn't have as many regions as there are region
servers.
Given the other caveats against using "balancer per table" maybe we could
expressly only do the short circuit when we are not doing per-table?
> Balancer should avoid leaving idle region servers
> -------------------------------------------------
>
> Key: HBASE-24139
> URL: https://issues.apache.org/jira/browse/HBASE-24139
> Project: HBase
> Issue Type: Improvement
> Components: Balancer, Operability
> Reporter: Sean Busbey
> Assignee: Sean Busbey
> Priority: Critical
>
> After HBASE-15529 the StochasticLoadBalancer makes the decision to run based
> on its internal cost functions rather than the simple region count skew of
> BaseLoadBalancer.
> Given the default weights for those cost functions, the default minimum cost
> to indicate a need to rebalance, and a regions per region server density of
> ~90 we are not very responsive to adding additional region servers for
> non-trivial cluster sizes:
> * For clusters ~10 nodes, the defaults think a single RS at 0 regions means
> we need to balance
> * For clusters >20 nodes, the defaults will not consider a single RS at 0
> regions to mean we need to balance. 2 RS at 0 will cause it to balance.
> * For clusters ~100 nodes, having 6 RS with no regions will still not meet
> the threshold to cause a balance.
> Note that this is the decision to look at balancer plans at all. The
> calculation is severely dominated by the region count skew (it has weight 500
> and all other weights are ~105), so barring a very significant change in all
> other cost functions this condition will persist indefinitely.
> Two possible approaches:
> * add a new cost function that's essentially "don't have RS with 0 regions"
> that an operator can tune
> * add a short circuit condition for the {{needsBalance}} method that checks
> for empty RS similar to the check we do for colocated region replicas
> For those currently hitting this an easy work around is to set
> {{hbase.master.balancer.stochastic.minCostNeedBalance}} to {{0.01}}. This
> will mean that a single RS having 0 regions will cause the balancer to run
> for clusters of up to ~90 region servers. It's essentially the same as the
> default slop of 0.01 used by the BaseLoadBalancer.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
