hi sergei, there's a file that constants are declared? i could start changing the constants to variables without problem

Part of this task is, exactly, finding and identifying these constants. Some of them are implicit and not easy to find.

Another big part if building the framework for collecting and storing time measurements and solving equations against these constants.

ok, what about files that i should start finding? sql_select.cc, sql_select.h, opt_range.cc (i found a constant inside NOT IN that rewrite to IN(), it's a kind of constant that i should look or not?), update and delete should be find too?
they are CAPS_LOCK_ON or they could be CAPS_LOCK_ON_and_off?

I'm afraid, it's something that not even I will be able to do, one of
our optimizer gurus should.

Sometimes there is no constant at all, it is implicitly assumed to be 1.

For example, optimizer compares table->file->read_time() values for
different tables. Here it assumes that results of ::read_time() for
different storage engines are directly comparable. They're not, they
should be multiplied by a storage engine specific factor.

Those places are very difficult to find.

We did have a GSoC-2013 project for this task, but it was assumed that
we will provide a list of constants to tune, and the student will only
work on time measurements and aggregation.

ok, should i start a list of constants that i find in the optimizer? or i will waste my time?

I'll be pushing the code on my github repo: https://github.com/igniting/server/tree/selfTuningOptimizer . @serg: I guess you can assign this issue to me.

So, to summarize: there are two approaches to obtaining timing data:

• counting operations, timing the query as a whole
• actually timing individual operations, P_S-style

In the first approach we only count number of index lookups (N _i), rows read in a table scan (N _s), etc. And time the query as a whole (T _q). This gives an equation:

T _i * N _i + T _s * N _s + ... = T _q

Collecting the data from many queries, we have a system of linear equations (may be even over-defined). It can be solved to give durations on individual operations.

In the second approach we directly time individual operations, much like PERFORMANCE_SCHEMA does. Preferably, reusing P_S measurements, not duplicating them.
These approaches both have their benefits:

• The first one is potentially much cheaper, it does not add timing overhead to every operation (and with constants like TIME_FOR_COMPARE, the individual operation is a comparison, timing it might be prohibitively expensive).
• Also, it uses the complete query execution time — the only time users care about. If we see that values of the coefficients don't converge over time, it means that there's an unaccounted-for part in the query time, and we add another coefficient.
• Making the second approach to use P_S when P_S is disabled is not exactly trivial.

But

• The second needs less memory: for N coefficients, it'll need — per THD — 4N doubles at most (value, sum, sum squared, counter). The first one needs to store(in addition to the above) N equations, that's N² values. Although for large N this might be a sparse matrix, I don't know.
• The second one measures only what we use in the optimizer, directly. No hidden parameters, and most coefficients should converge nicely. Still, it's meaningless from the user point of view.

I can see this going horribly wrong both when system load changes and due to hardware anomalies. For example when a hard drive hits a bad sector it may do an i/o which takes up to 8 seconds by default which may throw off all of your measurement for that query. If that query ends up updating the global statistics then all future queries may be dramatically punished all because of a single isolated problem.

cpu overload can also impact timing issues. For example if you have a burst of threads running then certain parts of a query may take longer as they fight for locks or cpu time.

It would be nice to run a sample workload to gather these statistics then be able to freeze them on an instance. This would take care of the cases where the optimizer makes assumptions about disk seek time that were true on hard drives but aren't true on flash hardware but without having to worry about hosts all the sudden wildly changing their query plans due to hardware blips.

A specific value (say, "read_time") will be measured many times and the on-disk table will store the sum of all measurements and the number of them. The optimizer will use the average (sum/count). This means:

• hopefully, short spikes won't disrupt the average too much
• one can effectively "freeze" the average by updating the table and putting a huge "count" value.

Eric, yes that's a problem, but... that's a nice solution to many problems
The statistic is the main part here, if the standard mean (or any other usable variable) become too distance from "current" measures we could alert DBA about disk problems or too much queries/second, think about a usefull feature to alert DBA and developers/engeneer about a possible problem =)
about increasing robustness of the self-tunning cost coefficients algorithm we can develop with more time and use cases, for the first version, a useless (no changes to cost coefficients/optimizer) but measurable feature is ok, with time we get the experience to write a good (robust/inteligent) control =]

Hi guys any news here?

I would also give +1 for a global statistic approach more that gathering fixe constants like in a benchmark

But i would add that a per table set of constant would be better because in any workload some specifics tables may cause more often non determinist cache overflow

• Data Cache
• Index Cache
• FS Cache
• Controller Cache
• CPU Caches

Non determinism's is also driven by the workload itself, like :

• writing reading from last time range ,
• number of secondary indexes in a write,
• storage engine
• randomness data access pattern
• lock time
• column and row size fetching

looks like such metrics are mostly per table dependent , and can help Query Plan to better estimate join cost at each depth

Hi All,

Went to MariaDB road Show yesterday at London, I suggested:

1). look at top slow queries in performance_schema table
2). record the query execution plan
3). try different index (especially compound index)
4). if new index works better, use it for the same query in future.

Hope this makes sense

1). look at top slow queries in performance_schema table
not only slow queries, but stats about these queries, they should be grouped by hash, like percona do "select * from table where column=? and column1>=? and column2>=?" constants are changed to "?" and operations (graph and/or/>/</=/etc...) are ordered

3). try different index (especially compound index)
gridsearching index is ok to do, consume resources but find many solutions, it's import to record not only time to execute, but resources and metadata information to create a good metric function, for example: one index with (int,int) is "better" than an index with (double,double,string,double) with the same time or not? a performace metric should be created to evaluate the best search path

4). if new index works better, use it for the same query in future.
at some point the problem is how to "same query" (the same of 1) at parser/optimizer and don't penalize global performace