Comments on: Innodb performance gotcha w Larger queries.

By: 5.0.86-build19 Percona binaries | MySQL Performance Blog

5.0.86-build19 Percona binaries | MySQL Performance Blog — Tue, 13 Oct 2009 06:05:05 +0000

[...] fix InnoDB slowness with large queries [...]

By: Chetan Ahuja

Chetan Ahuja — Mon, 28 Sep 2009 16:48:09 +0000

I just trawled through the source and it’s been fixed somewhere between 5.1.30 and 5.1.37. That is to say, I didn’t find the ” thd_is_select(trx->mysql_thd)” clause in the 5.1.30 codebase but did find it in the 5.1.37 codebase. And using vtune profiles, we even saw the actual dict_scan function that was dominating in the 5.1.30 codebase, disappear from the top 10 functions in 5.1.37 codebase. Though unlike your success with REPEATABLE-READ, we didn’t see any jump in our actual query throughput (and yes, we even tried the REPEATABLE-READ isolation level thing with 5.1.37 just in case… no cigar).

Chetan

By: peter

peter — Wed, 23 Sep 2009 17:51:20 +0000

Bhushan,

I do not know exact version – it is surely fixed in recent 5.1 versions. If you’re looking for us to dig into revision history and find in which version this change was done we can do it on consulting basics.

By: Bhushan Uparkar

Bhushan Uparkar — Wed, 23 Sep 2009 17:01:26 +0000

Could you please provide details about the mysql 5.1 version in which this got fixed ?

By: Log Buffer #160: a Carnival of the Vanities for DBAs | Pythian Group Blog

Log Buffer #160: a Carnival of the Vanities for DBAs | Pythian Group Blog — Fri, 04 Sep 2009 16:59:42 +0000

[...] Not that those other engines are without flaw. Peter Zaitsev reports on an InnoDB performance gotcha with larger queries. [...]

By: peter

peter — Fri, 28 Aug 2009 22:31:07 +0000

Arjen,

I do not know. Sometimes it is just assumed it would help. Plus it actually does help a little bit – as you can see from the code in the example you can have less locks if you’re running in low isolation level.

I also checked for single row replacement statements there was performance gain of about 5% for READ-UNCOMMITTED.
I think this is not enough to bother but I also remember some use cases with larger gains.

By: Arjen Lentz

Arjen Lentz — Fri, 28 Aug 2009 22:26:45 +0000

So why was the isolation level lowered in this case, can you tell?

And, when do you see measurable gain by lowering isolation level. As I understand from previous info, lowering the isolation level in InnoDB tends to not actaully save any time in real terms.