Balasore Best It Company|| Top 10 IT Company || Balasore Software company Odisha
MariaDB's join optimizer: how it works and current fixes
1. Sergei Petrunia
MariaDB devroom
FOSDEM 2021
Join Optimizer
1. How it works
2. What we’re working on to improve it
Optimizer Call
July 2022
Sergei Petrunia
MariaDB
2. 2
Join order search
●
Total number of possible join orders
for N-table join is:
N * (N-1) * (N-2) *… = N!
●
Join orders are built left-to-right
●
Cannot enumerate all possible join
orders.
t1
t1
t2
t2
t3
t4
t3
t4
t2
t4
t2
t3
t4
t3
t4
t2
t3
t2
t1
t3
t4
t3
t4
t1
t4
t1
t3
t4
t3
t4
t1
t3
t1
t3
t4
3. 3
Pruning
●
Enumerate promising join orders
first.
●
Do not explore join orders that are
apparently worse.
t1
t1
t2
t2
t3
t4
t3
t4
t2
t4
t2
t3
t4
t3
t4
t2
t3
t2
t1
t3
t4
t3
t4
t1
t4
t1
t3
t4
t3
t4
t1
t3
t1
t3
t4
4. 4
Pruning # 1: by cost
●
Cost of the current_prefix is
already higher than total cost of
best plan.
– Adding tables will make it even
higher
– No point to try.
●
This pruning is always done (no
switch)
●
Optimizer trace: pruned_by_cost
t1
t1
t2
t2
t3
t4
t3
t4
t2
t4
t2
t3
t4
t3
t4
t2
t3
t2
t1
t3
t4
t3
t4
t1
t4
t1
t3
t4
t3
t4
t1
t3
t1
t3
t4
5. 5
pruned_by_cost weaknesses
●
A really expensive table at the
end of the join order.
●
Any prefix that doesn’t include it is
relatively cheap
– Even if its comparably worse:
–
–
●
=> No pruning.
t1
t1
t2
t2
t3
t4
t3
t4
t2
t4
t2
t3
t4
t3
t4
t2
t3
t2
t1
t4
t3
t3
t4
t1
t3
t1
t4
t4
t3
t3
t1
t4
t1
t3
t4
t1
t1 t4
t2 t4
6. 6
Pruning # 2: by heuristic
●
Adding a table tX to a join prefix
– Adds read_time (time to read tX)
– Produces record_count row
combinations to be joined with further
tables (aka “join suffix”).
– Both have an effect on the total cost:
●
read_time is time spent right now.
●
record_count will affect cost of join
suffix.
– We don’t know the “exchange ratio”
because we don’t know the costs of
“join suffix”.
t0
t1
t2
t3
incoming_record_count
record_count_t1
record_count_t2
record_count_t3
read_time_t1
7. 7
The idea behind the heuristic
– … we don’t know the “exchange ratio”
because we don’t know the costs of “join
suffix”
●
Also the suffixes are different!
– Let’s assume the suffixes have similar costs.
– Then, if
●
read_time_t1 < read_time_t2, AND
●
record_count_t1 < record_count_t2
– Then t1 “is better” than t2.
– Can prune away t2.
t0
t1
t2
t3
incoming_record_count
record_count_t1
record_count_t2
record_count_t3
read_time_t1
8. 8
Applying heuristic pruning
●
Do it locally in each join prefix
●
First, consider more promising
tables first.
●
Less-promising tables second
– And try to prune them away.
t1
t1
t2
t2
t3
t4
t3
t4
t2
t4
t2
t3
t4
t3
t4
t2
t3
t2
t1
t3
t4
t3
t4
t1
t4
t1
t3
t4
t3
t4
t1
t3
t1
t3
t4
9. 9
Pruning # 2: by heuristic
●
A Model Table (yes, I’ve just invented this term):
– Lowest read_time AND record_count seen so
far
– Either
●
record_count < 2.0, or
●
there are no possible "key dependencies" on
tables not in the prefix
– A “possible key dependency” is an eqality in form:
tbl.keyXpartY=expr(tables_no_in_prefix)
●
^^ this is a “heuristic” to apply the heuristic.
●
Prune away tables that have both worse read_time
and record_count than the Model Table.
t1
t1
t2
t2
t3
t4
t3
t4
t2
t4
t2
t3
t4
t3
t4
t2
t3
t2
t1
t3
t4
t3
t4
t1
t4
t1
t3
t4
t3
t4
t1
t3
t1
t3
t4
10. 10
How one can see heuristic pruning
●
@@optimizer_prune_level
– 0 – not enabled.
– 1 – enabled (the default)
●
Optimizer trace: grep for
“pruned_by_heuristic”
t1
t1
t2
t2
t3
t4
t3
t4
t2
t4
t2
t3
t4
t3
t4
t2
t3
t2
t1
t3
t4
t3
t4
t1
t4
t1
t3
t4
t3
t4
t1
t3
t1
t3
t4
12. 12
Greedy search
●
Consider only prefixes of limited size
– Based on that, pick the first table
– Repeat
●
@@optimizer_search_depth
– Default: 62
(both MySQL and MariaDB)
– 0 – “pick depth automatically”
●
Why is this not default yet?
t1
t1
t2
t2
t3
t4
t3
t4
t2
t4
t2
t3
t4
t3
t4
t2
t3
t2
t1
t3
t4
t3
t4
t1
t4
t1
t3
t4
t3
t4
t1
t3
t1
t3
t4
14. 14
MDEV-28073: patch #1: “edge tables”
●
If the suffix t1-t4-t3 uses only eq_ref or similar:
– It is [nearly] the best
– Don’t enumerate other table combinations.
●
They can’t be much better.
●
Optimizer trace: pruned_by_hanging_leaf
t1
t1
t2
t2
t3
t4
t3
t4
t2
t4
t2
t3
t4
t3
t4
t2
t3
t2
t1
t3
t4
t3
t4
t1
t4
t1
t3
t4
t3
t4
t1
t3
t1
commit b729896d00e022f6205399376c0cc107e1ee0704
Author: Monty <monty@mariadb.org>
Date: Tue May 10 11:47:20 2022 +0300
MDEV-28073 Query performance degradation in newer MariaDB versions when
using many tables
The issue was that best_extension_by_limited_search() had to go through
too many plans with the same cost as there where many EQ_REF tables.
Fixed by shortcutting EQ_REF (AND REF) when the result only contains one
row. This got the optimization time down from hours to sub seconds.
t0
15. 15
MDEV-28073: patch #2: key_dependent
select ...
from
person, car_rides, bicycle_rides
where
person.name=car_rides.rider and
person.name=bicycle_rides.rider and
...
car_rides bicycle_rides
person
●
Remember the “heuristics to apply the heuristic” a few slides above:
– there are no possible "key dependencies" on tables not in the prefix
It can be false due to multi-equalities:
●
person.name=bicycle_rides.riders is a “possible key dependency”.
●
But we already have person.name from car_rides.rider (the equality is “bound”)
– Trying join orders with bicycle_rides before person won’t produce a better plan.
●
Solution: adjust the heuristic: there are no possible key_dependencies on tables not in
the prefix that are not already bound.
name
16. 16
MDEV-28073: patch #3: table order de-scrambling
●
The optimizer should try good tables first
●
Implemented by taking tables off the unused portion of
join->best_ref array.
– Initially it’s ordered (“promising” tables first)
– But due to bug eventually gets out of order
●
Plan searches that enumerate many options could
suffer from poor pruning towards the end.
t1
t1
t2
t2
t3
t4
t3
t4
t2
t4
t2
t3
t4
t3
t4
t2
t3
t2
t1
t3
t4
t3
t4
t1
t4
t1
t3
t4
t3
t4
t1
t3
t1
Author: Michael Widenius <monty@mariadb.org>
Date: Sun May 15 15:46:29 2022 +0300
greedy_search() and best_extension_by_limited_search() scrambled table order
best_extension_by_limited_search() assumes that tables should be sorted
according to size to be able to quickly disregard bad plans. However the
current usage of swap_variables() will change the table order to a not
sorted one for the next recursive call. This breaks the assumtion and
causes performance issues when using many tables (we have to examine
many more plans).
t0
19. 19
In which order do we try the tables?
●
Current:
– join_tab_cmp() orders all tables by their
JOIN_TAB::found_records
(records after table’s condition is checked)
– The same ordering is used everywhere
– This *ignores* the join prefix and efficien
table read plans we can use
– e.g. here, ignores the prefix of t1:
t1
t1
t2
t2
t3
t4
t3
t4
t2
t4
t2
t3
t4
t3
t4
t2
t3
t2
t1
t3
t4
t3
t4
t1
t4
t1
t3
t4
t3
t4
t1
t3
t1
t3
t4
t1
t1
t2
t3
t4
20. 20
In which order do we try the tables?
●
First, evaluate possible table accesses for {t2,t3,t4}.
●
Sort them by #found_rows
●
Then try extending join orders
– Do all kinds of pruning while doing this
t1
t1
t2
t3
t4
t3
t4
t2
t4
t2
t3
t4
t3
t4
t2
t3
t2
commit 0762dd9283185c72c6955f44fc4d862a0a928569
Author: Monty <monty@mariadb.org>
Date: Tue May 31 17:36:32 2022 +0300
Improve pruning in greedy_search by sorting tables during search
MDEV-28073 Slow query performance in MariaDB when using many tables
The faster we can find a good query plan, the more options we have for
finding and pruning (ignoring) bad plans.
This patch adds sorting of plans to best_extension_by_limited_search().
22. 22
Remember: the idea behind the heuristic
– … we don’t know the “exchange ratio”
because we don’t know the costs of “join
suffix”
●
Also the suffixes are different!
– Let’s assume the suffixes have similar costs.
– Then, if
●
read_time_t1 < read_time_t2, AND
●
record_count_t1 < record_count_t2
– Then t1 “is better” than t2.
– Can prune away t2.
–
t0
t1
t2
t3
incoming_record_count
record_count_t1
record_count_t2
record_count_t3
read_time_t1
24. 24
Let’s plot the tables
records_read
read_time
t1
Better than t1
Worse than t1
25. 25
Let’s plot the tables
●
“Typical” situation: plan with
higher cost produce more #rows.
– A lot of opportunities to do
pruning
●
The optimizer orders plans by
records_read
– (that is, goes left-to-right)
– Pick the first plan as “Model”,
prune those that are worse.
records_read
read_time
t1
t0
t2
t3
t4
26. 26
When pruning doesn’t work
records_read
read_time
t1
t0
t2
t3
t4
●
“Bad” situation:
– plans with high cost produce
few rows
– And vice versa
●
Can’t do pruning.
27. 27
When pruning could work but doesn’t
records_read
read_time
t1
t2
t3
t4
t0
●
Walking left-to-right, the optimizer
picks t0 as Model table.
●
And then can’t prune away any
other table.
Tables that are worse than t0 are here
28. 28
How to do as much pruning as possible?
records_read
read_time
●
Pick a minimal set of Model
tables that allow to prune away
the rest?
●
Complexity seems to be at least
N^2.
●
Some approximate algorithm?
– Use the table with min_cost
– Use the table with
min_records_read
●
Have a patch with some
approximate implementation
30. 30
Motivation
●
Tables with “attributes” that are joined using Primary Key
select * from base_table, attr1, attr2, ... attrN
where
attr1.pk = base_table.pk and
attr2.pk = base_table.pk and
...
attrN.pk = base_table.pk
●
Lots of nearly-identical query plans: There are factorial(n_attributes) permutations
– Have the same or very close cost
●
=> Can’t do pruning
●
The fix with “Edge tables” aka pruned_by_hanging_leaf helps but only if the
attributes are at the end of the join order.
31. 31
eq_ref chaining
●
The idea: if we see a eq_ref access, try
considering only eq_refs as long as we can.
●
MySQL 5.7 has a similar optimization
– TODO: describe the differences.
t1
t1
t2
t2
t3
t4
t3
t4
t2
t4
t2
t3
t4
t3
t4
t2
t3
t2
t1
t3
t4
t3
t4
t1
t4
t1
t3
t4
t3
t4
t1
t3
t1
t3
t4
commit 5abb6bff6cfb5cb5d87520f1e32e9b41db46bd7b
Author: Monty <monty@mariadb.org>
Date: Thu Jun 2 19:47:23 2022 +0300
Added EQ_REF chaining to the greedy_optimizer
MDEV-28073 Slow query performance in MariaDB when using many table
The idea is to prefer and chain EQ_REF tables (tables that uses an
unique key to find a row) when searching for the best table combination.
This significantly reduces row combinations that has to be examined.
This is optimization is enabled when setting optimizer_prune_level=2 (default)