Horton

Re: Write Ahead Logging for Hash Indexes

Amit Kapila <amit.kapila16@gmail.com>

View thread View on postgr.es
From: Amit Kapila <amit.kapila16@gmail.com>
To: Mark Kirkwood <mark.kirkwood@catalyst.net.nz>
Cc: pgsql-hackers <pgsql-hackers@postgresql.org>
Date: 2016-08-24T06:29:46Z
Lists: pgsql-hackers

Attachments

On Wed, Aug 24, 2016 at 11:44 AM, Mark Kirkwood
<mark.kirkwood@catalyst.net.nz> wrote:
> On 24/08/16 17:01, Mark Kirkwood wrote:
>>
>>
>> ...actually I was wrong there, only 2 of them were the same. So I've
>> attached a new log of bt's of them all.
>>
>>
>>
>
> And I can reproduce with only 1 session, figured that might be a helpful
> piece of the puzzle (trace attached).
>

Thanks.

I think I know the problem here.  Basically _hash_freeovflpage() is
trying to take a lock on a buffer previous to overflow page to update
the links and it is quite possible that the same buffer is already
locked for moving the tuples while squeezing the bucket.  I am working
on a fix for the same.

Coincidently, Ashutosh Sharma a colleague of mine who was also testing
this patch found the same issue by an attached sql script.  So we
might be able to inculcate a test case  in the regression suite as
well after fix.


-- 
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com

Commits

  1. Add a regression test for snapshot too old with hash indexes.

  2. hash: Add write-ahead logging support.

  3. Improve coding in _hash_addovflpage.

  4. Remove _hash_wrtbuf() in favor of calling MarkBufferDirty().

  5. Improve hash index bucket split behavior.