[
https://issues.apache.org/jira/browse/HIVE-1293?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12899320#action_12899320
]
Joydeep Sen Sarma commented on HIVE-1293:
-----------------------------------------
a little bummed that locks need to be held for entire query execution. that
could mean a writer blocking readers for hours.
hive's query plans seem to be of two distinct stages:
1. read a bunch of stuff, compute intermediate/final data
2. move final data into output locations
ie. - a single query never reads what it writes (into a final output location).
even if #1 and #2 are mingled today - they can easily be put in order.
in that sense - we only need to get shared locks for all read entities involved
in #1 to begin with. once phase #1 is done, we can drop all the read locks and
get the exclusive locks for all the write entities in #2, perform #2 and quit.
that way exclusive locks are held for a very short duration. i think this
scheme is similarly deadlock free (now there are two independent lock
acquire/release phases - and each of them can lock stuff in lex. order).
> Concurreny Model for Hive
> -------------------------
>
> Key: HIVE-1293
> URL: https://issues.apache.org/jira/browse/HIVE-1293
> Project: Hadoop Hive
> Issue Type: New Feature
> Components: Query Processor
> Reporter: Namit Jain
> Assignee: Namit Jain
> Fix For: 0.7.0
>
> Attachments: hive.1293.1.patch, hive.1293.2.patch, hive.1293.3.patch,
> hive.1293.4.patch, hive.1293.5.patch, hive_leases.txt
>
>
> Concurrency model for Hive:
> Currently, hive does not provide a good concurrency model. The only
> guanrantee provided in case of concurrent readers and writers is that
> reader will not see partial data from the old version (before the write) and
> partial data from the new version (after the write).
> This has come across as a big problem, specially for background processes
> performing maintenance operations.
> The following possible solutions come to mind.
> 1. Locks: Acquire read/write locks - they can be acquired at the beginning of
> the query or the write locks can be delayed till move
> task (when the directory is actually moved). Care needs to be taken for
> deadlocks.
> 2. Versioning: The writer can create a new version if the current version is
> being read. Note that, it is not equivalent to snapshots,
> the old version can only be accessed by the current readers, and will be
> deleted when all of them have finished.
> Comments.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
