[HBASE-16581] Optimize Replication queue transfers after server fail over

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Duplicate
• Assignee: Unassigned
• Reporter: Jerry He
• Created: 08/Sep/16 04:11
• Updated: 17/Jun/22 18:41
• Resolved: 08/Sep/16 17:04

Description

Currently if a region server fails, the replication queue of this server will be picked up by another region server. The problem is this queue can possibly be huge and contains queues from other multiple or cascading server failures.

We had such a case in production. From zk_dump:

...
/hbase/replication/rs/r01data10-va-pub.xxx.ext,60020,1472680007498/1-r01data10-va-pub.xxx.ext,60020,1455993417125-r01data07-va-pub.xxx.ext,60020,1472680008225-r01data08-va-pub.xxx.ext,60020,1472680007318/r01data10-va-pub.xxx.ext%2C60020%2C1455993417125.1467603735059: 18748267

/hbase/replication/rs/r01data10-va-pub.xxx.ext,60020,1472680007498/1-r01data10-va-pub.xxx.ext,60020,1455993417125-r01data07-va-pub.xxx.ext,60020,1472680008225-r01data08-va-pub.xxx.ext,60020,1472680007318/r01data10-va-pub.xxx.ext%2C60020%2C1455993417125.1471723778060: 

/hbase/replication/rs/r01data10-va-pub.xxx.ext,60020,1472680007498/1-r01data10-va-pub.xxx.ext,60020,1455993417125-r01data07-va-pub.xxx.ext,60020,1472680008225-r01data08-va-pub.xxx.ext,60020,1472680007318/r01data10-va-pub.xxx.ext%2C60020%2C1455993417125.1468258960080: 

/hbase/replication/rs/r01data10-va-pub.xxx.ext,60020,1472680007498/1-r01data10-va-pub.xxx.ext,60020,1455993417125-r01data07-va-pub.xxx.ext,60020,1472680008225-r01data08-va-pub.xxx.ext,60020,1472680007318/r01data10-va-pub.xxx.ext%2C60020%2C1455993417125.1468204958990: 

/hbase/replication/rs/r01data10-va-pub.xxx.ext,60020,1472680007498/1-r01data10-va-pub.xxx.ext,60020,1455993417125-r01data07-va-pub.xxx.ext,60020,1472680008225-r01data08-va-pub.xxx.ext,60020,1472680007318/r01data10-va-pub.xxx.ext%2C60020%2C1455993417125.1469701010649: 

/hbase/replication/rs/r01data10-va-pub.xxx.ext,60020,1472680007498/1-r01data10-va-pub.xxx.ext,60020,1455993417125-r01data07-va-pub.xxx.ext,60020,1472680008225-r01data08-va-pub.xxx.ext,60020,1472680007318/r01data10-va-pub.xxx.ext%2C60020%2C1455993417125.1470409989238: 

/hbase/replication/rs/r01data10-va-pub.xxx.ext,60020,1472680007498/1-r01data10-va-pub.xxx.ext,60020,1455993417125-r01data07-va-pub.xxx.ext,60020,1472680008225-r01data08-va-pub.xxx.ext,60020,1472680007318/r01data10-va-pub.xxx.ext%2C60020%2C1455993417125.1471838985073: 

/hbase/replication/rs/r01data10-va-pub.xxx.ext,60020,1472680007498/1-r01data10-va-pub.xxx.ext,60020,1455993417125-r01data07-va-pub.xxx.ext,60020,1472680008225-r01data08-va-pub.xxx.ext,60020,1472680007318/r01data10-va-pub.xxx.ext%2C60020%2C1455993417125.1467142915090: 57804890

/hbase/replication/rs/r01data10-va-pub.xxx.ext,60020,1472680007498/1-r01data10-va-pub.xxx.ext,60020,1455993417125-r01data07-va-pub.xxx.ext,60020,1472680008225-r01data08-va-pub.xxx.ext,60020,1472680007318/r01data10-va-pub.xxx.ext%2C60020%2C1455993417125.1472181000614: 

/hbase/replication/rs/r01data10-va-pub.xxx.ext,60020,1472680007498/1-r01data10-va-pub.xxx.ext,60020,1455993417125-r01data07-va-pub.xxx.ext,60020,1472680008225-r01data08-va-pub.xxx.ext,60020,1472680007318/r01data10-va-pub.xxx.ext%2C60020%2C1455993417125.1471464567365: 

/hbase/replication/rs/r01data10-va-pub.xxx.ext,60020,1472680007498/1-r01data10-va-pub.xxx.ext,60020,1455993417125-r01data07-va-pub.xxx.ext,60020,1472680008225-r01data08-va-pub.xxx.ext,60020,1472680007318/r01data10-va-pub.xxx.ext%2C60020%2C1455993417125.1469486466965: 

/hbase/replication/rs/r01data10-va-pub.xxx.ext,60020,1472680007498/1-r01data10-va-pub.xxx.ext,60020,1455993417125-r01data07-va-pub.xxx.ext,60020,1472680008225-r01data08-va-pub.xxx.ext,60020,1472680007318/r01data10-va-pub.xxx.ext%2C60020%2C1455993417125.1467787339841: 47812951
...

There were hundreds of wals hanging under this queue, coming from diferent region servers, which took a long time to replicate.

We should have a better strategy which lets live region servers each grep part of this nested queue, and replicate in parallel.

Issue Links

• is duplicated by: HBASE-12770 Don't transfer all the queued hlogs of a dead server to the same alive server