[HBASE-24554] Improve/stable read replica

• Type: Task
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 2.3.0
• Component/s: read replicas
• Assignee: Huaxiang Sun
• Reporter: Huaxiang Sun
• Created: 12/Jun/20 23:27
• Updated: 05/Aug/20 21:10

Description

Tracing some read replica issues recently, this is the umbrella Jira to track this effort. A few observations so far:

1. balancer balances replica regions too often, need to spend time on it. Replica region does not serve write and rarely serve reads (unless the client specifically selects the replica region). So data locality should be a very minimum factor for replica regions. 
2. Need to study split/merge for regions with replica, need to make them more robust. With proc-v2, probably it is already robust. 

 

 

Issue Links

• contains: HBASE-24582 The current implementation of assignMetaReplicas() may assign replica meta regions to the same server hosting primary meta region.
• contains: HBASE-24552 Replica region needs to check if primary region directory exists at file system in TransitRegionStateProcedure
• contains: HBASE-24705 MetaFixer#fixHoles() does not include the case for read replicas (i.e, replica regions are not created)
• contains: HBASE-24720 Meta replicas not cleaned when disabled
• contains: HBASE-24581 Skip compaction request/check for replica regions at the early stage.
• contains: HBASE-24633 Remove data locality and StoreFileCostFunction for replica regions out of balancer's cost calculation
• contains: HBASE-24643 Replace Cluster#primariesOfRegionsPerServer from int array to treemap
• contains: HBASE-24707 Fix Empty region in meta with read replica
• contains: HBASE-24562 Stabilize master startup with meta replicas enabled
• contains: HBASE-24824 Add more stats in PE for read replica