Using histograms to get better performance

Using histograms
to get better performance
Sergei Petrunia
Varun Gupta

Database performance
● Performance is a product of many
factors
● One of them is Query optimizer
● It produces query plans
– A “good” query plan only
reads rows that contribute to
the query result
– A “bad” query plan means
unnecessary work is done

Do my queries use bad query plans?
● Queries take a long time
● Some are just inherently hard to
compute
● Some look good but turn out bad
due to factors that were not
accounted for

Query plan cost depends on data statistics
select *
from
lineitem, orders
where
o_orderkey=l_orderkey and
o_orderdate between '1990-01-01' and '1998-12-06' and
l_extendedprice > 1000000
● orders->lineitem
vs
lineitem->orders
● Depends on
condition selectivity

Another choice optimizer has to make
select *
from
orders
where
o_orderstatus='F'
order by
order_date
limit 10
● Use index(order_date)
– Stop as soon as we find 10 matches
● Find rows with o_orderstatus='F'
– Sort by o_orderdate picking first 10
● Again, it depends on condition
selectivity.

Data statistics in MariaDB
● Table: #rows in the table
● Index
– cardinality: AVG(#lineitems per order)
– “range estimates” - #rows(t.key BETWEEN const1 and
const2)
● Non-index column? Histogram

Histogram
● Partition the value space into buckets
– Store bucket bounds and #values in the bucket
– Imprecise
– Very compact

Summary so far
● Good database performance requires good query plans
● To pick those, optimizer needs statistics about the data
– Condition selectivity is important
● Certain kinds of statistics are always available
– Indexes
– For non-indexed columns, histograms may be needed.

Do my query plans suffer
from bad statistics?

Will my queries benefit?
● Very complex question
● No definite answer
● Suggestions
– ANALYZE for statements, r_filtered.
– Slow query log

ANALYZE for statements and r_filtered
● filtered – % of rows left after applying condition (expectation)
– r_filtered - ... - the reality
● r_filtered << filtered – the optimizer didn’t know the condition is selective
– Happens on a non-first table? We are filtering out late!
●
Add histogram on the column (Check the cond in FORMAT=JSON)
analyze select *
from lineitem, orders
where o_orderkey=l_orderkey and
o_orderdate between '1990-01-01' and '1998-12-06' and
l_extendedprice > 1000000
+--+-----------+--------+----+-------------+-------+-------+-----------------+-------+-------+--------+----------+-----------+
|id|select_type|table |type|possible_keys|key |key_len|ref |rows |r_rows |filtered|r_filtered|Extra |
+--+-----------+--------+----+-------------+-------+-------+-----------------+-------+-------+--------+----------+-----------+
|1 |SIMPLE |orders |ALL |PRIMARY,i_...|NULL |NULL |NULL |1504278|1500000| 50.00 | 100.00 |Using where|
|1 |SIMPLE |lineitem|ref |PRIMARY,i_...|PRIMARY|4 |orders.o_orderkey|2 |4.00 | 100.00 | 0.00 |Using where|
+--+-----------+--------+----+-------------+-------+-------+-----------------+-------+-------+--------+----------+-----------+

# Query_time: 1.961549 Lock_time: 0.011164 Rows_sent: 1 Rows_examined: 11745000
# Rows_affected: 0 Bytes_sent: 73
# Full_scan: Yes Full_join: No Tmp_table: No Tmp_table_on_disk: No
# Filesort: No Filesort_on_disk: No Merge_passes: 0 Priority_queue: No
#
# explain: id select_type table type possible_keys key key_len ref rows r_rows
filtered r_filtered Extra
# explain: 1 SIMPLE inventory ALL NULL NULL NULL NULL 11837024
11745000.00 100.00 0.00 Using where
#
SET timestamp=1551155484;
select count(inv_date_sk) from inventory where inv_quantity_on_hand>10000;
Slow Query Log
slow-query-log
long-query-time=...
log-slow-verbosity=query_plan,explain
my.cnf
hostname-slow.log
● Rows_examined >> Rows_sent? Grouping,or a poor query plan
● log_slow_query=explain will shows ANALYZE output

Histograms in MariaDB
● Available since MariaDB 10.0 (Yes)
● Used by advanced users
● Not enabled by default
● Have limitations, not user-friendly
● MariaDB 10.4
– Fixes some of the limitations
– Makes histograms easier to use

Configuration for collecting histograms
histogram_size=0
histogram_type=SINGLE_PREC_HB
histogram_size=254
histogram_type=DOUBLE_PREC_HB
● MariaDB before 10.4: change the default histogram size
● MariaDB 10.4 : enable automatic sampling
histogram_size=254
histogram_type=DOUBLE_PREC_HB
analyze_sample_percentage=100

ANALYZE ... PERSISTENT collects histograms
– Collect statistics for everything:
analyze table t1 persistent
for columns (col1,...) indexes (idx1,...);
+---------+---------+----------+-----------------------------------------+
+---------+---------+----------+-----------------------------------------+
| test.t1 | analyze | status | Engine-independent statistics collected |
+---------+---------+----------+-----------------------------------------+
analyze table t1 persistent for all;

Can make histogram collection automatic
set use_stat_tables='preferably';
analyze table t1;
+---------+---------+----------+-----------------------------------------+
+---------+---------+----------+-----------------------------------------+
| test.t1 | analyze | status | Engine-independent statistics collected |
+---------+---------+----------+-----------------------------------------+
● Beware: this may be *much* slower than ANALYZE TABLE
you’re used to
● Great for migrations

Histogram collection performance
● MariaDB 10.0: uses all data in the table to build histogram
– Precise, but expensive
– Particularly so for VARCHARs
● A test on a real table:
– Real table, 740M rows, 90GB
– CHECKSUM TABLE: 5 min
– ANALYZE TABLE ... PERSISTENT FOR ALL – 30 min

MariaDB 10.4: Bernoulli sampling
● Default: analyze_sample_percentage=100
– Uses the entire table, slow
● Suggested: analyze_sample_percentage=0
– “Roll the dice” sampling, size picked automatically
analyze table t1 persistent for columns (...) indexes();
analyze table t1 persistent for all;
– full table and secondary index scans
– does a full table scan

Further plans: genuine sampling
● Work on avoiding full table scans is in progress
● Will allow to make ANALYZE TABLE collect all histograms

Making the optimizer
use histograms

Make the optimizer use histograms
@@use_stat_tables=NEVER
@@optimizer_use_condition_selectivity=1
@@use_stat_tables=PREFERABLY // also affects ANALYZE!
● MariaDB before 10.4: does not use histograms
● MariaDB 10.4 : uses histograms if they are collected
@@use_stat_tables=PREFERABLY_FOR_QUERIES
– remember to re-collect!

Conclusions: how to start using histograms
● MariaDB before 10.4
use_stat_tables=PREFERABLY # Changes optimizer
optimizer_use_condition_selectivity=4 # behavior
● MariaDB 10.4
● Both: ANALYZE TABLE ... PERSISTENT FOR ...
histogram_size=254 # No risk
histogram_type=DOUBLE_PREC_HB #

Can I just have histograms
for all columns?

A stored procedure to analyze every table
CREATE PROCEDURE analyze_persistent_for_all(db_name VARCHAR(64))
BEGIN
DECLARE done INT DEFAULT FALSE;
DECLARE x VARCHAR(64);
DECLARE cur1 CURSOR FOR
SELECT TABLE_NAME
FROM INFORMATION_SCHEMA.TABLES
WHERE TABLE_TYPE = 'BASE TABLE' AND TABLE_SCHEMA=db_name;
DECLARE CONTINUE HANDLER FOR NOT FOUND SET done = TRUE;
OPEN cur1;
read_loop: LOOP
FETCH cur1 INTO x;
IF done THEN
LEAVE read_loop;
END IF;
SET @sql = CONCAT('analyze table ', x, ' persistent for all');
PREPARE stmt FROM @sql;
EXECUTE stmt;
DEALLOCATE PREPARE stmt;
END LOOP;
CLOSE cur1;
END|

Should I ANALYZE ... PERSISTENT every table?
● New application
– Worth giving it a try
– Provision for periodic ANALYZE
– Column correlations?
● Existing application
– Performance fixes on a case-by-case basis.

TPC-DS benchmark
● scale=1
● The same dataset
– without histograms: ~20 min
– after ‘call analyze_persistent_for_all(‘tpcds’) from two slides
prior: 5 min.

A customer case with ORDER BY ... LIMIT
● table/column names replaced
CREATE TABLE cars (
type varchar(10),
company varchar(20),
model varchar(20),
quantity int,
KEY quantity (quantity),
KEY type (type)
);
select * from cars
where
type='electric' and
company='audi'
order by
quantity
limit 3;
● table/column names replaced
● quantity matches the ORDER BY, but need to match condition
● type is a Restrictive index

● Uses ORDER-BY compatible index by default
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: cars
type: index
possible_keys: type
key: quantity
key_len: 5
ref: const
rows: 994266
r_rows: 700706.00
filtered: 0.20
r_filtered: 0.00
Extra: Using where
1 row in set (2.098 sec)
select * from cars
where
type='electric' and
company='audi'
order by
quantity
limit 3;

● Providing the optimizer with histogram
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: cars
type: ref
possible_keys: type
key: type
key_len: 13
ref: const
rows: 2022
r_rows: 3.00
filtered: 100.00
r_filtered: 100.00
Extra: Using index condition; Using where; Using filesort
1 row in set (0.010 sec)
analyze table cars persistent for all;
select * from cars
where
type='electric' and
company='audi'
order by
quantity
limit 3;

Histograms are stored in a table
CREATE TABLE mysql.column_stats (
db_name varchar(64) NOT NULL,
table_name varchar(64) NOT NULL,
column_name varchar(64) NOT NULL,
min_value varbinary(255) DEFAULT NULL,
max_value varbinary(255) DEFAULT NULL,
nulls_ratio decimal(12,4) DEFAULT NULL,
avg_length decimal(12,4) DEFAULT NULL,
avg_frequency decimal(12,4) DEFAULT NULL,
hist_size tinyint unsigned,
hist_type enum('SINGLE_PREC_HB','DOUBLE_PREC_HB'),
histogram varbinary(255),
PRIMARY KEY (db_name,table_name,column_name)
);

TPC-DS benchmark
● Can save/restore histograms
● Can set @@optimizer_use_condition_selectivity to disable
histogram use per-thread

Problem with correlated conditions
● Possible selectivities
– MIN(1/n, 1/m)
– (1/n) * (1/m)
– 0
select ...
from order_items
where shipdate='2015-12-15' AND item_name='christmas light'
'swimsuit'

Problem with correlated conditions
● PostgreSQL: Multi-variate statistics
– Detects functional dependencies, col1=F(col2)
– Only used for equality predicates
– Also #DISTINCT(a,b)
● MariaDB: MDEV-11107: Use table check constraints in optimizer
– In development
select ...
from order_items
where shipdate='2015-12-15' AND item_name='christmas light'
'swimsuit'

Using histograms to get better performance

More Related Content

What's hot

Similar to Using histograms to get better performance

More from Sergey Petrunya

Recently uploaded

Using histograms to get better performance