Mysql count performance on very big tables
I have a table with more than 100 millions rows in Innodb.
I have to know if there is more than 5000 rows where the foreign key = 1. I don't need the exact number.
I made some testing :
SELECT COUNT(*) FROM table WHERE fk = 1 => 16 seconds SELECT COUNT(*) FROM table WHERE fk = 1 LIMIT 5000 => 16 seconds SELECT primary FROM table WHERE fk = 1 => 0.6 seconds
I will have a bigger network and treatment time but it can be an overload of 15.4 seconds !
Do you have a better idea ?
Edit: [Added OP's relevant comments]
I tried SELECT SQL_NO_CACHE COUNT(fk) FROM table WHERE fk = 1 but it took 25 seconds
Mysql was tuned for Innodb with Mysql Tuner.
CREATE TABLE table ( pk bigint(20) NOT NULL AUTO_INCREMENT, fk tinyint(3) unsigned DEFAULT '0', PRIMARY KEY (pk), KEY idx_fk (fk) USING BTREE ) ENGINE=InnoDB AUTO_INCREMENT=100380914 DEFAULT CHARSET=latin1
'have_innodb', 'YES' 'ignore_builtin_innodb', 'OFF' 'innodb_adaptive_hash_index', 'ON' 'innodb_additional_mem_pool_size', '20971520' 'innodb_autoextend_increment', '8' 'innodb_autoinc_lock_mode', '1' 'innodb_buffer_pool_size', '25769803776' 'innodb_checksums', 'ON' 'innodb_commit_concurrency', '0', 'innodb_concurrency_tickets', '500' 'innodb_data_file_path', 'ibdata1:10M:autoextend' 'innodb_data_home_dir', '', 'innodb_doublewrite', 'ON' 'innodb_fast_shutdown', '1' 'innodb_file_io_threads', '4' 'innodb_file_per_table', 'OFF', 'innodb_flush_log_at_trx_commit', '1' 'innodb_flush_method', '' 'innodb_force_recovery', '0' 'innodb_lock_wait_timeout', '50' 'innodb_locks_unsafe_for_binlog', 'OFF' 'innodb_log_buffer_size', '8388608' 'innodb_log_file_size', '26214400' 'innodb_log_files_in_group', '2' 'innodb_log_group_home_dir', './' 'innodb_max_dirty_pages_pct', '90' 'innodb_max_purge_lag', '0' 'innodb_mirrored_log_groups', '1' 'innodb_open_files', '300' 'innodb_rollback_on_timeout', 'OFF' 'innodb_stats_on_metadata', 'ON' 'innodb_support_xa', 'ON' 'innodb_sync_spin_loops', '20' 'innodb_table_locks', 'ON' 'innodb_thread_concurrency', '8' 'innodb_thread_sleep_delay', '10000' 'innodb_use_legacy_cardinality_algorithm', 'ON'
Update '15: I used the same method up to now with 600 millions rows and 640 000 new rows per day. It's still working fine.
Counter tables or other caching mechanism is the solution:
InnoDB does not keep an internal count of rows in a table because concurrent transactions might “see” different numbers of rows at the same time. To process a SELECT COUNT(*) FROM t statement, InnoDB scans an index of the table, which takes some time if the index is not entirely in the buffer pool. If your table does not change often, using the MySQL query cache is a good solution. To get a fast count, you have to use a counter table you create yourself and let your application update it according to the inserts and deletes it does. If an approximate row count is sufficient, SHOW TABLE STATUS can be used. See Section 18.104.22.168, “InnoDB Performance Tuning Tips”.
You don't seem interested in the actual count so give this a try:
SELECT 1 FROM table WHERE fk = 1 LIMIT 5000, 1
If a row is returned, you have 5000 and more records. I presume the fk column is indexed.
I gotta add another Answer -- I have many corrections/additions to the comments and Answers so far.
For MyISAM, SELECT COUNT(*) without WHERE is dead-reckoned -- very fast. All other situations (include the InnoDB in the Question) must count through either the data's BTree or an index's BTree to get the answer. So we need to see how much to count through.
InnoDB caches data and index blocks (16KB each). But when the table's data or index BTree is bigger than innodb_buffer_pool_size, you are guaranteed to hit the disk. Hitting the disk is almost always the slowest part of any SQL.
The Query Cache, when involved, usually results in query times of about 1 millisecond; this does not seem to be an issue with any of the timings quoted. So I won't dwell on it.
But... Runing the same query twice in a row will often exhibit:
- First run: 10 seconds
- Second run: 1 second
This is symptomatic of the first run having to fetch most of the blocks from disk, while the second found it all in RAM (the buffer_pool). I suspect that some of the timings listed are bogus because of not realizing this caching issue. (16 sec vs 0.6 sec may be explained by this.)
I will harp on "disk hits" or "blocks needed to be touched" as the real metric of which SQL is faster.
COUNT(x) checks x for IS NOT NULL before tallying. This adds a tiny amount of processing, but does not change the number of disk hits.
The proffered table has a PK and a second column. I wonder if that is the real table?? It makes a difference --
- If the Optimizer decides to read the data -- that is, scan in PRIMARY KEY order -- it will be reading the data BTree, which is usually (but not in this lame example) much wider than secondary index BTrees.
- If the Optimizer decides to read a secondary index (but not need to do a sort), there will be fewer blocks to touch. Hence, faster.
Comments on the original queries:
SELECT COUNT(*) FROM table WHERE fk = 1 => 16 seconds -- INDEX(fk) is optimal, but see below SELECT COUNT(*) FROM table WHERE fk = 1 LIMIT 5000 => 16 seconds -- the LIMIT does nothing, since there is only one row in the result SELECT primary FROM table WHERE fk = 1 => 0.6 seconds -- Again INDEX(fk), but see below
WHERE fk = 1 begs for INDEX(fk, ...), preferably just INDEX(fk). Note that in InnoDB, each secondary index contains a copy of the pk. That is, INDEX(fk) is effectively INDEX(fk, primary). Hence, the 3rd query can use that as "covering" and not need to touch the data.
If the table is truly just the two columns then probably the secondary index BTree will be fatter than the data BTree. But in realistic tables, the secondary index will be smaller. Hence an index scan will be faster (fewer blocks to touch) than a table scan.
The third query is also delivering a large resultset; this could cause the query to take a long time -- but it won't be included in the quoted "time"; it is network time, not query time.
innodb_buffer_pool_size = 25,769,803,776 I would guess that the table and its secondary index (from the FK) are each about 3-4GB. So, any timing might first have to load a lot of stuff. Then a second run would be entirely cached. (Of course, I don't know how many rows have fk=1; presumably less than all the rows?)
But... At 600M rows, the table and its index are each approaching the 25GB buffer_pool. So, the day may come soon that it becomes I/O bound -- this will make you wish to get back to 16 (or 25) seconds; yet you won't be able to. We can then talk about alternatives to doing the COUNT.
SELECT 1 FROM tbl WHERE fk = 1 LIMIT 5000,1 -- Let's analyze this. It will scan the index, but it will stop after 5000 rows. Of all you need is "more than 5K", that is the best way to get it. It will be consistently fast (touching only a dozen blocks), regardless of total number of rows in the table. (It is still subject to buffer_pool_size and cache characteristics of the system. But a dozen blocks takes much less than a second, even with a cold cache.)
MariaDB's LIMIT ROWS_EXAMINED may be worth looking into. Without that, you could do
SELECT COUNT(*) AS count_if_less_than_5K FROM ( SELECT 1 FROM tbl WHERE fk = 1 LIMIT 5000 );
It may be faster than delivering the rows to the client; it will have to collect the rows internally in a tmp table, but deliver only the COUNT.
A side note: 640K rows inserted per day -- this approaches the limit for single-row INSERTs in MySQL with your current settings on a HDD (not SDD). If you need to discuss the potential disaster, open another Question.
- Be sure to avoid the Query cache. (by using SQL_NO_CACHE or turning the QC off)
- Run any timing query twice; use the second time.
- Understand the structure and size of the BTree(s) involved.
- Don't use COUNT(x) unless you need the null check.
- Do not use PHP's mysql_* interface; switch to mysqli_* or PDO.
If you are using PHP you could do mysql_num_rows on the result you got from SELECT primary FROM table WHERE fk = 1 => 0.6 seconds, I think that will be efficient.
But depends on what server-side language you are using
Finally the fastest was to query the first X rows using C# and counting the rows number.
My application is treating the data in batches. The amount of time between two batches are depending the number of rows who need to be treated
SELECT pk FROM table WHERE fk = 1 LIMIT X
I got the result in 0.9 seconds.
Thanks all for your ideas!
If you're not interested to know the number of rows and you just want to test the COUNT against some value, you can use the standard script bellow:
SELECT 'X' FROM mytable WHERE myfield='A' HAVING COUNT(*) >5
This will return one single row or no row at all, depending if condition is met.
This script is ANSI compliant and can be fully run without evaluating the complete value of COUNT(*). If MySQL implemented optimization to stop evaluating rows after some condition is met (I really hope it does), then you'll get a performance improvement. Unfortunately I can't test this behavior myself because I don't have a big MySQL database available. If you do this test, please share the result here :)