Umesh Kumar Kumawat created HBASE-30331:
-------------------------------------------
Summary: RegionMover.filterRSGroupServers short-circuits on
default RSGroup, causing unnecessary region churn on clusters with custom
RSGroups
Key: HBASE-30331
URL: https://issues.apache.org/jira/browse/HBASE-30331
Project: HBase
Issue Type: Bug
Components: rsgroup
Affects Versions: 3.0.0
Reporter: Umesh Kumar Kumawat
Problem
RegionMover.filterRSGroupServers() in hbase-server contains a short-circuit
for the default RSGroup:
// RegionMover.java line 530
if (rsgroup.getName().equals(RSGroupInfo.DEFAULT_GROUP)) {
return onlineServers; // returns ALL online servers — including
custom-group servers
}
When unloadRegions() is called on a server that belongs to the default
RSGroup, filterRSGroupServers skips the membership check and returns every live
RegionServer as a valid move target — including
servers in custom RSGroups. HMaster's RSGroup-aware balancer will eventually
move those regions back to the correct group, but RegionMover is causing work
that the master then has to undo.
Root Cause
The short-circuit was added as an optimization: on a cluster where all
servers are in the default group, iterating the membership set is redundant.
However, once any server is moved to a custom RSGroup,
the default-group filter no longer reflects actual membership — it
over-includes servers that have left the default group.
Impact
- A RegionMover unload on a default-group server may place regions onto
custom-group servers.
- HMaster detects the misplacement and moves those regions back via the
RSGroup-aware balancer, causing an unnecessary round of region transitions and
load imbalance during the window between RegionMover
finishing and the next balancer cycle.
- Affects any cluster that uses RSGroups and runs RegionMover (rolling
restarts, decommission workflows).
The corresponding fix for branch-2 was shipped as part of HBASE-30323. This
ticket tracks the same correction on master/branch-3.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
