New subquery
optimizations
in MySQL 6.0
Presented by,
MySQL AB® & O’Reilly Media, Inc.
Sergey Petrunia
sergefp@mysql.com
Background: subquery processing before 6.0
FROM subqueries are pre-materialized (early)
Scalar-context subqueries use straightforward
evaluation
Predicate subqueries
May perform two kinds of rewrites
Then use straightforward evaluation
Originally implemented in MySQL 4.1
by Sinisa (FROM subqueries) and Sanja (all other kinds)
Processing subqueries in the FROM clause
SELECT ... FROM (SELECT ...) AS tbl WHERE ...
Execution steps
1. Optimize the subquery SELECT
2. Execute it and capture result into a temporary table
3. Optimize and execute the parent SELECT
Properties
No optimization (can get some if you define/use a VIEW equivalent
to subquery)
EXPLAIN command runs the subquery and thus can be very slow
Straightforward subquery evaluation
Used for all kinds of
subqueries other than
FROM:
expr IN (SELECT ...)
EXISTS (SELECT ...)
expr (SELECT.... )
scalar context
subqueries
Subquery is optimized once,
all re-evaluations are done
using the same plan
Uncorrelated subqueries
are evaluated only once
Straightforward subquery evaluation (contd)
SELECT ... FROM outer_tbl1,outer_tbl2
WHERE expr IN (SELECT inner_expr
FROM inner_tbl1, inner_tbl2 WHERE ... )
select_type=“SUBQUERY” means subquery is run only once.
“DEPENDENT SUBUQERY” means it is re-run on every re-evaluation
Subqueries in WHERE/ON are re-evaluated when their WHERE
AND-part is evaluated, which is as soon as possible
Subqueries in select list, HAVING, etc are re-evaluated for every
record combination
Subquery rewrites: IN->EXISTS (1)
“Inform the subquery about which part of its resultset we're
interested in”
IN→EXISTS transformation:
OuterExpr IN (SELECT InnerExpr FROM ...
WHERE subq_where)
→
EXISTS (SELECT 1 FROM ...
WHERE subq_where AND
InnerExpr = OuterExpr)
Things to note
Uncorrelated subquery becomes correlated
This is a simplifed description, not counting cases with NULLs
Subquery rewrites: MIN/MAX (2)
“Inform the subquery about which part of its resultset we're
interested in”
MIN/MAX Transformation
OuterExpr > ALL(SELECT InnerExpr FROM ...)
→
OuterExpr > (SELECT MAX(InnerExpr)FROM ... )
handles all similar cases with
OuterExpr (SELECT...)
* simplifed description, not counting cases with NULLs or subqueries
returning zero rows
Current state of subqueries: summary
FROM subqueries
are always pre-materialized, exactly like in the query
Scalar and predicate subqueries
Optimized using two rule-based rewrites:
IN→EXISTS (pushdown the IN-equality)
ALL/ANY→MIN/MAX
Evaluated using straightforward outer-to-inner strategy
As early is possible if located in the WHERE
“Very late” if located in other parts of the query
Are evaluated only once if uncorellated
Subquery optimization map: 6.0
Subquery optimization map: 6.0
Semi-join subquery optimizations
A practically important kind of subqueries:
SELECT * FROM . . .
WHERE query_where AND
outer_expr IN (SELECT inner_expr
FROM . . . WHERE ...)
In relational algebra, semi-join is defined as:
outertbl SEMI JOIN innertbl ON sj_cond =
{outertbl.row│∃ innertbl.row, sj_cond(outertbl.row, innertbl.row)}
A subquery is processed as a semi-join if
it is an IN/=ANY subquery it is an AND-part of the WHERE clause
it is not a UNION, has no aggregates or ORDER BY ... LIMIT
SELECT DISTINCT or “dummy” GROUP BY are allowed
Semi-join vs. inner join semantics
The difference is in duplicate outer row combinations
SELECT Country.Name FROM Country
WHERE Code IN (SELECT CountryCode FROM City
WHERE Population > 1M)
DEU │ Berlin
DEU │ Germany Germany
DEU │ Hamburg
SELECT Country.Name FROM Country, City
WHERE Country.Code=City.CountryCode AND
City.Population > 1M
DEU │ Berlin Germany
DEU │ Germany
DEU │ Hamburg Germany
=> semi-join is like inner join but we need some way to remove
the duplicates
Semi-join strategy #1: Table pullout
If a subquery table is functionally dependent on the parent query
tables, it can be “pulled out” of the subquery
SELECT City.Name FROM City
WHERE City.Country IN (SELECT Country.Code FROM Country
WHERE Country.SurfaceArea < 2K)
Victoria│HKG HKG│Hong Kong Victoria
is converted into
SELECT City.Name FROM City, Country
WHERE City.Country = Country.Code AND
Country.SurfaceArea < 2K)
•
If the subquery has several tables, will pull out those tables that don't
generate duplicate matches
Semi-join strategy #1: Table pullout: example
EXPLAIN EXTENDED SELECT City.Name FROM City
WHERE City.Country IN (SELECT Country.Code FROM Country
WHERE Country.SurfaceArea < 10);
SHOW WARNINGS;
In MySQL 4.1/5.x :
id select_type table type possible_keys key key_len ref rows Extra
1 PRIMARY City ALL NULL NULL NULL NULL 4079 Using where
DEPENDENT PRIMARY,
2 Country unique_subquery PRIMARY 3 func 1 Using where
SUBQUERY SurfaceArea
select `world`.`City`.`Name` AS `Name` from `world`.`City` where
<in_optimizer>(`world`.`City`.`Country`,<exists>(<primary_index_lookup>(<cach
e>(`world`.`City`.`Country`) in Country on PRIMARY where ...
In MySQL 6.0 :
id select_type table type possible_keys key key_len ref rows Extra
PRIMARY, Using index condition;
1 PRIMARY Country range SurfaceArea 4 NULL 3
SurfaceArea Using MRR
1 PRIMARY City ref Country Country 3 Country.Code 18
select `world`.`City`.`Name` AS `Name` from `world`.`Country` join `world`.`City`
where ((`world`.`City`.`Country` = `world`.`Country`.`Code`) and (`world`.`...
Semi-join strategy #1: Table pullout: summary
In two words, this is subquery-to-join conversion
Properties
It is rule-based, pullout is done whenever possible
It enables the join optimizer to make a cost-based choice
from a greater variety of query plans (including a plan that is
eqivalent to pre-6.0 server strategy)
Applicability
Pullout is done before any other semi-join strategy
considerations
Can handle correlated subqueries (analogous functionality
in PostgreSQL, surprisingly, doesn't)
Can handle arbitarily deep subquery nesting
Semi-join strategy #2: FirstMatch
Short-cut enumeration of subquery tables as soon as we get
first matching row combination
SELECT * FROM ot1,ot2,nt1, ...
WHERE expr(ot1,ot2) IN (SELECT ... FROM it1,it2 ...)
Semi-join strategy #2: FirstMatch: example
EXPLAIN EXTENDED
SELECT Name FROM Country
WHERE
Country.Continent='Europe' AND
Country.Code IN (SELECT City.Country FROM City
WHERE City.ID != Country.Capital AND
Population > 1M)
id select_type table type possible_keys key key_len ref rows Extra
1 PRIMARY Country ref PRIMARY, Continent Continent 21 const 8 Using index condition;
Using where;
1 PRIMARY City ref Population, Country Country 3 Country.Code 18
FirstMatch(Country)
select `world`.`Country`.`Name` AS `Name` from `world`.`Country` semi join
(`world`.`City`) where ((`world`.`City`.`Country` = `world`.`Country`.`Code`) and
(`world`.`Country`.`Continent` = 'Europe') and (`world`.`City`.`ID` <> `world`.
`Country`.`Capital`) and (`world`.`City`.`Population` > 1000000))
Semi-join strategy #2: FirstMatch (contd)
Similar to IN->EXISTS
Strictly outer-to-inner join order
Better than IN->EXISTS
Don't have to evaluate subquery
immediately after outer tables it
refers to:
SELECT employee.*
FROM employee NATURAL JOIN office
WHERE employee.hire_date > '2008-01-01' AND
office.country='EU' AND
employee_id IN (SELECT employee_id
FROM conference_speaker)
4.1/5.x: employee--conference_speaker, office
6.0 employee, office, conference_speaker
Doesn't force IN->EXISTS rewrite so allows for other
optimizations
Semi-join strategy #3: duplicate elimination
•Use temporary table with unique key (or constraint) to
eliminate duplicate row combinations of outer tables
SELECT Country.Name FROM Country
WHERE Code IN (SELECT Country FROM City
WHERE Population > 1M)
City (inner) Country (outer)
Berlin │ DEU DEU│Germany│ROWID
temp DEU│Germany
Paris │ FRA FRA│France │ROWID table
filter FRA│France
Hamburg│ DEU DEU│Germany│ROWID
UNIQUE(Countries.ROWID)
Duplicate elimination: example
SELECT Country.Name FROM Country
WHERE Code IN (SELECT Country FROM City
WHERE Population > 1M)
select_type table type possible_keys key key_len ref rows Extra
Using index condition;
Population,
PRIMARY City range Population 4 NULL 246 Using MRR;
Country
Start temporary
PRIMARY Country eq_ref PRIMARY PRIMARY 3 City.Country 1 End temporary
select `world`.`Country`.`Name` AS `Name` from `world`.`Country` semi join
(`world`.`City`) where ((`world`.`Country`.`Code` = `world`.`City`.`Country`) and
(`world`.`City`.`Population` > 1000000))
Can have outer/inner tables in any order
Including Inner-to-outer which wasn't possible till 6.0
Can handle correlated subqueries
Similar to materialization (uses memory, hash lookups)
But different – temp table gets outer table's rows (not inner)
that have matches (materialization stores all inner table rows)
Semi-join strategy #4: InsideOut
Scan inner table(s) in a way that doesn't produce duplicates
SELECT ... FROM outer_tbl
WHERE outer_expr IN (SELECT it1.poor_key
FROM it1,it2 WHERE cond(it1, it2))
Semi-join strategy #4: InsideOut: example
SELECT * FROM outer_tbl
WHERE key1 IN (SELECT poor_key FROM inner_tbl);
id select_type table type possible_keys key key_len ref rows Extra
Using index;
1 PRIMARY inner_tbl index poor_key poor_key 5 NULL 10000
LooseScan
1 PRIMARY outer_tbl ref key1 key1 5 inner_tbl.poor_key 1 Using index
Only inner-to-outer join orders
Subquery must be uncorrelated
There must be an index that covers subquery' select list
At the moment usable with 'ref' or 'index'
Should be also usable with 'range' but not yet
CAUTION there are known bugs.
Materialization strategy
Use temporary table with unique constraint to
• - Materialize the subquery result
• - Create unique hash index on it
• - Lookup outer tuples into temp table index
• SELECT Country.Name FROM Country
WHERE Population IN (SELECT SUM(Population)
FROM City
GROUP BY CountryCode)
SELECT SUM(Population) temp
For each Country.Population
FROM City
table
+ if (hash lookup into
GROUP BY CountryCode hash temp table)
index
output result
Remove duplicates Hash
lookup
Controlling new subquery optimizations
Currently, a server variable:
@@optimizer_switch =
'no_semijoin,no_materialization'
(like set-type column: any order, no space after comma)
this will likely to change into being a part of a bigger
optimization on/off scheme (WL#4046)
Already seeing a need for hints but no WL entry for this yet
thinking of syntax like
outer_expr IN (SELECT no_materialize ...)
Benchmarking new optimizations
A look at standard benchmarks: DBT-{1,2,3}
DBT-3 has 10 subquery cases
Of which 8 are not covered by new optimizations (2 are
covered)
Query #18: covered (materialization), execution times:
Engine Query time
MySQL 6.0, no new optimizations > 3 hours
MySQL 6.0, materialization 3.76 sec ~1800 times
PostgreSQL 6.52 sec faster now
Query #16: will be covered by NULL-aware materialization
Engine Query time
MySQL 6.0, no new optimizations 0.55 sec
PostgreSQL 1.16 sec
* used DBT3 scale=1, InnoDB, all default settings
Benchmarking new optimizations (contd)
MySQL bugs/customer cases and DBT-3 have different
subquery populations
MySQL bugs Support cases DBT-3
No idea about the reason of the difference
We intend to develop some subquery benchmark to cover
subqueries like in MySQL bugs/support db
Benchmarking new optimizations (contd)
MySQL bugs/customer issues that are easily repeatable
Found 10 subquery cases
Taking PostgreSQL's speed as 1.0:
No 6.0 optimizations Materialization Semi-join Run parameters
67285.714 34.286 1.429 MySQL 6.0.3
59490.000 780.000 n/a
9.477 2.109 0.004 PostreSQL 8.3.0
151.429 206.667 0.476 No tuning, all
1360.000 490.000 10.000 default settings
670.453 0.264 1.052
Small query
16.364 0.455 0.182
10.000 0.625 n/a
population
5648.649 3.243 0.270 => numbers only
962.500 1.500 n/a show order of
Medians: 816.48 2.68 0.48 magnitude
Semj-join and materialization together: 0.84
BTW, about PostgreSQL
We compare against PostgreSQL often
That's not because we have a goal to compete or
outperform PostgreSQL
It's just an
OSS DBMS
That is easy to use
Has a feature-rich optimizer
Does some things differently than MySQL
And some of us have expirience with
=> Natural first choice but we'd like to compare
with other databases too
Coverage of new optimizations
Subquery classification:
Correlated/uncorrelated:
MySQL 4.1/5.x: correlate if
possible
MySQL 6.0:
Flattening, FirstMatch,
temptable – don't care
InsideOut, Materialization:
uncorrelated only.
PostgreSQL:
Flattening (equvalent),
Hash/Merge Join – seem
to handle uncorrelated only
Coverage of new optimizations
Uncorrelated semi-join subquery classification
MySQL 4/5.x MySQL 6.0 PostgreSQL *
Conversion to inner join Nothing Flattening Flattening
Outer-to-inner
IN->EXISTS (with
Lookup index available all its limitations)
First Match IN NL-Join
Duplicate
Nested Loop join IN NL-Join, Hash
No usable index without buffering
Elimination,
join
Materialization
Inner-to-outer
Merge join,
Can use index to remove duplicates Nothing InsideOut
Hash join
Duplicate Hash join
No index for duplicate elimination Nothing Elimination, variants,
Materialization Sort+Unique
* based on our observations, may be incomplete
Future subquery work
Doing now
Bug fixing
WL#3485 FROM subquery flattening
WL#3985 Smart choice between semi-join and
materialization
WL#3830 Partial matching of tuples with NULL
components: let materialization handle
nullable_col IN (SELECT ...)
smth IN (SELECT nullable_expr ...)
Intend to do
SQL-level subquery hints
subquery_predicate_value(correlation_values) cache
Ongoing and future subquery work
Smart choice
between
semi-join and
NULL handling materialization
FROM flattening
Semi-join
W
PR OR
O K
G IN
FROM subquery flattening RE
SS
Merge the FROM subquery into the upper join
SELECT ...
FROM (SELECT * FROM inner_tbl WHERE ...) tbl,
...
WHERE tbl.col='foo' AND ...
Work in progress (done by Evgen Potemkin)
Applicability conditions are same as for materialized
VIEWs: subquery must
Not be a UNION
No GROUP BY or aggregates
No ORDER BY ... LIMIT or DISTINCT
W
PR OR
O K
G IN
Materialization and NULLs RE
SS
.. outer_expr [NOT] IN (SELECT innercol
FROM inner_tbl WHERE ...)
NULL problems:
On the left
NULL IN (SELECT something) = NULL
NULL IN (SELECT nothing) = FALSE
On the right:
'bar'
'foo' IN (SELECT col FROM 'baz' ) = TRUE/FALSE
...
'bar'
'foo' IN (SELECT col FROM 'baz' ) = NULL
NULL
References
6.0 Subquery optimizations cheatsheet
http://forge.mysql.com/wiki/6.0_Subquery_Optimization_Cheatsheet
Technlical specs: Subquery optimizations: semijoin:
WL#3985 and its subtasks
http://forge.mysql.com/worklog/task.php?id=3985
Technical specs: Subquery optimizations:
materialization: WL#1110
http://forge.mysql.com/worklog/task.php?id=1110
MySQL 6.0 Subquery optimization benchmarks
http://forge.mysql.com/wiki/6.0_Subquery_Optimization_Benchmarks
Observations and news about subquery development
http://s.petrunia.net/blog/
The end
Thank you
Q&A