[
https://issues.apache.org/jira/browse/HIVE-11388?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15061006#comment-15061006
]
Eugene Koifman commented on HIVE-11388:
---------------------------------------
here is one general purpose mechanism:
create table MUTEX_TABLE(keyname varchar(512) PRIMARY KEY)
Then any process that requires a mutex needs to insert a row into this table
(as long as everyone agrees on the key) and then do a "Select for update" on
this row. If the process dies, "select for update" lock is automatically
released.
For example, if 2 Initiator instances want to schedule a compaction, each could
1. select * from MUTEX_TABLE where keyname="initiator" for update.
If the "initiator" row is already there, only 1 will succeed. The other one,
once it unblocks, will already see "this" compaction scheduled.
2. if select in 1 misses, then Initiator can insert "initiator" row and then
goto 1. Because of PK only 1 will succeed.
Since the keyname is arbitrary, it can be "db/table/partition" to coordinate
Workers if necessary.
A little primitive, but workable and avoids ZooKeeper and allows all parts of
Compaction/HouseKeeping to run on multiple MS nodes.
> there should only be 1 Initiator for compactions per Hive installation
> ----------------------------------------------------------------------
>
> Key: HIVE-11388
> URL: https://issues.apache.org/jira/browse/HIVE-11388
> Project: Hive
> Issue Type: Bug
> Components: Transactions
> Affects Versions: 1.0.0
> Reporter: Eugene Koifman
> Assignee: Eugene Koifman
> Priority: Critical
>
> org.apache.hadoop.hive.ql.txn.compactor.Initiator is a thread that runs
> inside the metastore service to manage compactions of ACID tables. There
> should be exactly 1 instance of this thread (even with multiple Thrift
> services).
> This is documented in
> https://cwiki.apache.org/confluence/display/Hive/Hive+Transactions#HiveTransactions-Configuration
> but not enforced.
> Should add enforcement, since more than 1 Initiator could cause concurrent
> attempts to compact the same table/partition - which will not work.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
