Ray Mattingly created HBASE-29070:
-------------------------------------
Summary: Balancer cost function epsilon is imprecise
Key: HBASE-29070
URL: https://issues.apache.org/jira/browse/HBASE-29070
Project: HBase
Issue Type: Bug
Affects Versions: 2.6.1
Reporter: Ray Mattingly
Assignee: Ray Mattingly
The balancer cost function is imprecise. This means that the balancer can
behave in unexpected ways at certain scales. For example, with enough regions
and replicas, it's very possible for the
StochasticLoadBalancers#areReplicasColocated (paraphrased) method to return a
false negative, and fail to trigger a balancer run despite bad replica
locations existing.
One immediate consequence of this bug is that our test suites are designed in
slow and flappy ways. They all have huge balancer run times, but really our
tests should be designed with tight balancer run times and iteration if
necessary (in most cases). If we had done that, then it would've exposed this
failure of the balancer to act as necessary, and our test suite would be much
faster.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
