maria-developers team mailing list archive

[Sergei Petrunia <psergey@xxxxxxxxxxxx>]

Hi Zardosht,

On Thu, Mar 07, 2013 at 11:08:01PM -0500, Zardosht Kasheff wrote:
> I have found an interesting issue with index condition pushdown. I do
> not know if it is a bug, or if I am misusing the feature.
> 
> I have found that in one scenario, when performing a join, an index
> condition is pushed down, but because handler->end_range is not set,
> handler_index_cond_check never returns ICP_OUT_OF_RANGE, even though
> we are out of range.

Confirm. I have looked how it works in MyISAM and InnoDB. Indeed, they consider
scan over equality to be a special case. They remember the lookup key and
check it themselves.

That is, if one does

  h->index_read(key='foo')=0
  h->index_next_same()

then index_next_same() will check whether the next index tuple has key='foo'. 
If the key is different, it will return HA_ERR_END_OF_FILE without checking 
the index condition.

I'm hesitant to say that this is "by design", I'd say that the available 
storage engines happened to work this way, and then ICP used that.

Maybe, this means that implementation of handler::index_next_same() should
set handler::end_range. If it did, a storage engine will be able to support 
ICP without implementing index_next_same() call.

The question for MariaDB is, does it make sense for us to make this change in
MariaDB, when Oracle doesn't support it? 

(As far as I understand, TokuDB will still need to support MySQL, and so will
have to work around the problem by supporting index_next_same() and comparing
the keys like MyISAM/InnoDB do.  When you have to do that anyway, there is no 
benefit from having this problem fixed in MariaDB...)

> 
> Here is the relevant stack trace, I cannot yet send the entire trace
> as it may include customer information, and I need to double check it
> is ok to send:
<skip>

> 
> In this stack, we are performing an index_next_same, and a condition
> has been pushed down. When we call handler_index_cond_check. we always
> get ICP_NO_MATCH and never get ICP_OUT_OF_RANGE, even though we go
> past the key that we should be retrieving.
> 
> Is this by design? Is this a bug?
> 
-- 
BR
 Sergei
-- 
Sergei Petrunia, Software Developer
Monty Program AB, http://askmonty.org
Blog: http://s.petrunia.net/blog