Sid Khillon created HBASE-30529:
-----------------------------------
Summary: Throttle the region reopen when a replication peer
becomes serial
Key: HBASE-30529
URL: https://issues.apache.org/jira/browse/HBASE-30529
Project: HBase
Issue Type: Bug
Components: proc-v2, Replication
Affects Versions: 2.6.5, 3.0.0-beta-1, 2.6.0
Reporter: Sid Khillon
Assignee: Sid Khillon
When a replication peer is changed from non-serial to serial,
ModifyPeerProcedure.reopenRegions reopens every region of the peer's tables
with {{{}new ReopenTableRegionsProcedure(tn){}}}. That constructor turns off
batching, so {{hbase.reopen.table.regions.progressive.batch.size.max}} and
{{hbase.reopen.table.regions.progressive.batch.backoff.ms}} are ignored and
every region of the table closes and reopens at the same time.
The same happens when a serial peer is added for tables that already have
global replication scope.
Table alters (ModifyTableProcedure) and HMaster#reopenRegionsThrottled already
use {{{}ReopenTableRegionsProcedure.throttled(conf, td){}}}. HBASE-29809 moved
HMaster#reopenRegions and RegionsRecoveryChore to the throttled path and
mentioned replication procedures, but did not change ModifyPeerProcedure.
*Example*
A table with 932 regions, batch size max 50 and backoff 1000ms:
* An alter reopens it in 24 batches (1, 2, 4, 8, 16, 32, then 50 at a time).
* Making its replication peer serial reopens all 932 regions in one batch.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
