From: Jorgen Loland
Date: January 14 2011 3:44pm
Subject: Re: bzr commit into mysql-next-mr-bugfixing branch (jorgen.loland:3242)
 WL#4800
List-Archive: http://lists.mysql.com/commits/128816
Message-Id: <4D306F7A.10904@oracle.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 8bit

On 01/14/2011 03:51 PM, Guilhem Bichot wrote:
> Hello,
>
> I deleted all "ok, no comment" points.
>
> Jorgen Loland a écrit, Le 14.01.2011 12:49:
>> > But wait, this is interesting. I have the following issue (in the
>> > current wl4800 tree): make_join_select() looks like this:
>> >
>> > open array in trace;
>> > for (take tables from first to last)
>> > {
>> > create object in trace for this table;
>> > write table's name to trace;
>> > determine attached condition;
>> > write attached condition to trace;
>> > }
>> >
>> > But "determine attached condition" contains pushdown_on_conditions()
>> > which may attach condition to tables located *before* the one which the
>> > iteration is positioned on. And as I have already printed the condition
>> > for this earlier table, the earlier printed condition is now
>> incomplete.
>> > The example is "t1 LEFT JOIN (t2 LEFT JOIN t3 ON ...) ON ...": when
>> > positioned on t3, bits of the ON conditions get attached to t1 and t2.
>> > As a solution, I considered switching to this "summarizing" pseudocode:
>> >
>> > for (take tables from first to last)
>> > {
>> > determine attached condition;
>> > }
>> > open array in trace;
>> > for (take tables from first to last)
>> > {
>> > create object in trace for this table;
>> > write table's name to trace;
>> > write attached condition to trace;
>> > }
>> > But "determine attached condition" calls the range opt (right below in
>> > your patch), so this blows up (range opt tracing without the proper
>> > context => crash).
>> > So maybe we need:
>> >
>> > open array in trace, for key "attached_conditions_computation"
>> > for (take tables from first to last)
>> > {
>> > determine attached condition; as part of it, if using the range opt
>> > then create proper enclosing context with table's name as you do in the
>> > patch
>> > }
>> > close array
>> > open array in trace, for key "attached_conditions_summary"
>> > for (take tables from first to last)
>> > {
>> > create object in trace for this table;
>> > write table's name to trace;
>> > write attached condition to trace;
>> > }
>> >
>> > If you find this is ok, I can handle implementation.
>>
>> If I understand correctly, you want to execute range optimization
>> all over again for all tables?
>
> No. Current code does:
> open array in trace;
> for (take tables from first to last)
> {
> create object in trace for this table;
> write table's name to trace;
> determine attached condition;
> write attached condition to trace;
> }
> and I would change to
> open array in trace, for key "attached_conditions_computation"
> for (take tables from first to last)
> {
> determine attached condition; as part of it, if using the range opt
> then create proper enclosing context with table's name as you do in the
> patch
> }
> close array
> open array in trace, for key "attached_conditions_summary"
> for (take tables from first to last)
> {
> create object in trace for this table;
> write table's name to trace;
> write attached condition to trace;
> }
> so the second for() loop only does tracing, nothing else.
> The first loop does no tracing except if range opt is used, in which
> case tracing is done by your:
> Opt_trace_object trace_table(join->thd->opt_trace,
> "rechecking_index_usage");
> trace_table.add_utf8_table(tab->table);

Ok, sounds like a good idea

-- 
Jørgen Løland | Senior Software Engineer | +47 73842138
Oracle MySQL
Trondheim, Norway