← Back to team overview

maria-developers team mailing list archive

Updated (by Psergey): Subquery optimization: Avoid recalculating subquery if external fields values found in subquery cache (66)

 

-----------------------------------------------------------------------
                              WORKLOG TASK
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
TASK...........: Subquery optimization: Avoid recalculating subquery if external fields
		values found in subquery cache
CREATION DATE..: Wed, 25 Nov 2009, 22:25
SUPERVISOR.....: Monty
IMPLEMENTOR....: Sanja
COPIES TO......: 
CATEGORY.......: Server-BackLog
TASK ID........: 66 (http://askmonty.org/worklog/?tid=66)
VERSION........: Server-5.2
STATUS.........: Assigned
PRIORITY.......: 60
WORKED HOURS...: 0
ESTIMATE.......: 0 (hours remain)
ORIG. ESTIMATE.: 0

PROGRESS NOTES:

-=-=(Psergey - Tue, 12 Jan 2010, 18:39)=-=-
High-Level Specification modified.
--- /tmp/wklog.66.old.31666     2010-01-12 18:39:43.000000000 +0200
+++ /tmp/wklog.66.new.31666     2010-01-12 18:39:43.000000000 +0200
@@ -4,3 +4,99 @@
 
 To check/discuss:
  To put subquery cache on all levels of subqueries or on highest level only.
+
+
+<contents>
+1. Scope of the task
+2. Data structure used for the cache
+3. Cache size
+4. Interplay with other subquery optimizations
+5. User interface
+</contents>
+
+1. Scope of the task
+--------------------
+This WL should handle all subquery predicates, i.e. it should handle these
+cases:
+
+  outer_expr IN (SELECT correlated_select)
+  outer_expr $CMP$ ALL/ANY (SELECT correlated_select)
+  EXISTS (SELECT correlated_select)
+  scalar-context subquery: (SELECT correlated_select)
+
+The cache will maintain 
+  
+  (outer_expr, correlation_references)-> subquery_item_result 
+
+mapping, where
+- correlation_references is a list of tablename.column_name that are referred 
+  from the correlated_select but tablename is a table that is ouside the
+  subquery. 
+- subquery_item_result is 'bool' for subquery predicates, and is of
+some scalar or  ROW(scalar1,...scalarN) type for scalar-context subquery.
+
+We dont support cases when outer_expr or correlation_references are blobs.
+
+2. Data structure used for the cache
+------------------------------------
+There are two data structures available in the codebase that will allow fast 
+equality lookups:
+
+1. HASH (mysys/hash.c) tables
+2. Temporary tables (the ones that are used for e.g. GROUP BY)
+
+None of them has any support for element eviction on overflow (using LRU or
+some other policy). 
+
+Query cache and MyISAM/Maria's key/page cache ought to support some eviction 
+mechanism, but code-wise it is not readily reusable, one will need to factor 
+it out (or copy it).
+
+We choose to use #2, and not to have any eviction policy. See subsequent
+sections for details and reasoning behind the decision.
+
+3. Cache size
+-------------
+Typically, a cache has some maximum size and a policy which is used to
+select a cache entry for removal when the cache becomes full (e.g. find 
+and remove the least [recently] used entry)
+
+For this WL entry we will use a cache of infinite size. The reasoning behind
+this is that:
+- is is easy to do: we have temporary tables that can grow to arbitrarily
+  large size while still providing the same insert/lookup interface.
+- it suits us: unless the subquery is resolved with one index lookup,
+  hitting the cache would be many times cheaper than re-running the
+  subquery, so cache is worth having.
+
+4. Interplay with other subquery optimizations
+----------------------------------------------
+* This WL entry should not care about IN->EXISTS transformation: caching for 
+  IN subquery and result of its conversion to EXISTS would work in the same
+  way.
+
+* This optimization is orthogonal to <=>ANY -> MIN/MAX rewrite (it will
+  work/be useful irrespectively of whether the rewrite has been performed or
+  not)
+
+* TODO: compare this with materialization for uncorrelated IN-subqueries. Is 
+  this basically the same?
+  A: no, it is not:
+  - IN-Materialization has to perform full materialization before it can 
+    do the first subquery evaluation. This WL's code has almost no startup 
+    costs.
+  - This optimization has temp.table of (corr_reference, predicate_value),
+    while IN-materialization will have (corr_reference) only.
+
+5. User interface
+-----------------
+* There will be an @@optimizer_switch flag to turn this optimization on and
+  off (TODO: name of the flag?)
+
+* TODO: how do we show this in EXPLAIN [EXTENDED]?  The most easiest is to 
+  print something in the warning text of EXPLAIN EXTEDED that would indicate 
+  use of cache.
+
+* temporary table sizing (max size for heap table, whether to use MyISAM or
+  Maria) will be controlled with common temp.table control variables.
+

-=-=(Psergey - Mon, 11 Jan 2010, 13:25)=-=-
As of today, there is code that
- collects outside references 
- creates a temporary table with index that would allow for fast lookups.

there is no code to
- fill the temporary table 
- make lookups into it

Reported zero hours worked. Estimate unchanged.

-=-=(Sanja - Fri, 11 Dec 2009, 15:09)=-=-
High-Level Specification modified.
--- /tmp/wklog.66.old.28164     2009-12-11 15:09:15.000000000 +0200
+++ /tmp/wklog.66.new.28164     2009-12-11 15:09:15.000000000 +0200
@@ -3,4 +3,4 @@
 
 
 To check/discuss:
-Are there sens to put subquery cache on all levels of subqueries of on highest.
+ To put subquery cache on all levels of subqueries or on highest level only.

-=-=(Sanja - Fri, 11 Dec 2009, 15:08)=-=-
Low Level Design modified.
--- /tmp/wklog.66.old.28072     2009-12-11 15:08:16.000000000 +0200
+++ /tmp/wklog.66.new.28072     2009-12-11 15:08:16.000000000 +0200
@@ -1 +1,6 @@
+All items on which subquery depend could be collected in
+st_select_lex::mark_as_dependent (direct of indirect reference?)
+
+Temporary table index should be created by all fields except result field
+(TMP_TABLE_PARAM::keyinfo).
 

-=-=(Sanja - Fri, 11 Dec 2009, 15:05)=-=-
High-Level Specification modified.
--- /tmp/wklog.66.old.27795     2009-12-11 15:05:04.000000000 +0200
+++ /tmp/wklog.66.new.27795     2009-12-11 15:05:04.000000000 +0200
@@ -1 +1,6 @@
+Attach subquery cache to each Item_subquery. Interface should allow to use hash
+or temporary table inside.
 
+
+To check/discuss:
+Are there sens to put subquery cache on all levels of subqueries of on highest.



DESCRIPTION:

Collect all outer items/references (left part of the subquiery and outer
references inside the subquery) in key string. Compare the string (which
represents certain value set of the references) against values in hash table and
return cached result of subquery if the reference values combination has already
been used.

For example in the following subquery:
(L1, L2) IN (SELECT A, B FROM T WHERE T.F1>OTER_FIELD)
set of references to look into the subquery cache is (L1, L2, OTER_FIELD).

The subquery cache should be implemented as simple LRU connected to the subquery.

Size of the subquery cache (in number of results (but maybe in used memory
amount)) is limited by session variable (query parameter?).


HIGH-LEVEL SPECIFICATION:



Attach subquery cache to each Item_subquery. Interface should allow to use hash
or temporary table inside.


To check/discuss:
 To put subquery cache on all levels of subqueries or on highest level only.


<contents>
1. Scope of the task
2. Data structure used for the cache
3. Cache size
4. Interplay with other subquery optimizations
5. User interface
</contents>

1. Scope of the task
--------------------
This WL should handle all subquery predicates, i.e. it should handle these
cases:

  outer_expr IN (SELECT correlated_select)
  outer_expr $CMP$ ALL/ANY (SELECT correlated_select)
  EXISTS (SELECT correlated_select)
  scalar-context subquery: (SELECT correlated_select)

The cache will maintain 
  
  (outer_expr, correlation_references)-> subquery_item_result 

mapping, where
- correlation_references is a list of tablename.column_name that are referred 
  from the correlated_select but tablename is a table that is ouside the
  subquery. 
- subquery_item_result is 'bool' for subquery predicates, and is of
some scalar or  ROW(scalar1,...scalarN) type for scalar-context subquery.

We dont support cases when outer_expr or correlation_references are blobs.

2. Data structure used for the cache
------------------------------------
There are two data structures available in the codebase that will allow fast 
equality lookups:

1. HASH (mysys/hash.c) tables
2. Temporary tables (the ones that are used for e.g. GROUP BY)

None of them has any support for element eviction on overflow (using LRU or
some other policy). 

Query cache and MyISAM/Maria's key/page cache ought to support some eviction 
mechanism, but code-wise it is not readily reusable, one will need to factor 
it out (or copy it).

We choose to use #2, and not to have any eviction policy. See subsequent
sections for details and reasoning behind the decision.

3. Cache size
-------------
Typically, a cache has some maximum size and a policy which is used to
select a cache entry for removal when the cache becomes full (e.g. find 
and remove the least [recently] used entry)

For this WL entry we will use a cache of infinite size. The reasoning behind
this is that:
- is is easy to do: we have temporary tables that can grow to arbitrarily
  large size while still providing the same insert/lookup interface.
- it suits us: unless the subquery is resolved with one index lookup,
  hitting the cache would be many times cheaper than re-running the
  subquery, so cache is worth having.

4. Interplay with other subquery optimizations
----------------------------------------------
* This WL entry should not care about IN->EXISTS transformation: caching for 
  IN subquery and result of its conversion to EXISTS would work in the same
  way.

* This optimization is orthogonal to <=>ANY -> MIN/MAX rewrite (it will
  work/be useful irrespectively of whether the rewrite has been performed or
  not)

* TODO: compare this with materialization for uncorrelated IN-subqueries. Is 
  this basically the same?
  A: no, it is not:
  - IN-Materialization has to perform full materialization before it can 
    do the first subquery evaluation. This WL's code has almost no startup 
    costs.
  - This optimization has temp.table of (corr_reference, predicate_value),
    while IN-materialization will have (corr_reference) only.

5. User interface
-----------------
* There will be an @@optimizer_switch flag to turn this optimization on and
  off (TODO: name of the flag?)

* TODO: how do we show this in EXPLAIN [EXTENDED]?  The most easiest is to 
  print something in the warning text of EXPLAIN EXTEDED that would indicate 
  use of cache.

* temporary table sizing (max size for heap table, whether to use MyISAM or
  Maria) will be controlled with common temp.table control variables.


LOW-LEVEL DESIGN:



All items on which subquery depend could be collected in
st_select_lex::mark_as_dependent (direct of indirect reference?)

Temporary table index should be created by all fields except result field
(TMP_TABLE_PARAM::keyinfo).


ESTIMATED WORK TIME

ESTIMATED COMPLETION DATE
-----------------------------------------------------------------------
WorkLog (v3.5.9)