Related
Dear StackOverflow Members
It's my first post, so please be nice :-)
I have a strange SQL behavior which i can't explain and don't find any resources which explains it.
I have built a web honeypot which record all access and attacks and display it on a statistic page.
However since the data increased, the generation of the statistic page is getting slower and slower.
I narrowed it down to a some select statements which takes a quite a long time.
The "issue" seems to be an index on a specific column.
*For sure the real issue is my lack of knowledge :-)
Database: mysql
DB schema
Event Table (removed unrelated columes):
Event table size: 30MB
Event table records: 335k
CREATE TABLE `event` (
`EventID` int(11) NOT NULL,
`EventTime` datetime NOT NULL DEFAULT current_timestamp(),
`WEBURL` varchar(50) COLLATE utf8_bin DEFAULT NULL,
`IP` varchar(15) COLLATE utf8_bin NOT NULL,
`AttackID` int(11) NOT NULL
) ENGINE=InnoDB DEFAULT CHARSET=utf8 COLLATE=utf8_bin;
ALTER TABLE `event`
ADD PRIMARY KEY (`EventID`),
ADD KEY `AttackID` (`AttackID`);
ALTER TABLE `event`
ADD CONSTRAINT `event_ibfk_1` FOREIGN KEY (`AttackID`) REFERENCES `attack` (`AttackID`);
Attack Table
attack table size: 32KB
attack Table records: 11
CREATE TABLE attack (
`AttackID` int(4) NOT NULL,
`AttackName` varchar(30) COLLATE utf8_bin NOT NULL,
`AttackDescription` varchar(70) COLLATE utf8_bin NOT NULL
) ENGINE=InnoDB DEFAULT CHARSET=utf8 COLLATE=utf8_bin;
ALTER TABLE `attack`
ADD PRIMARY KEY (`AttackID`),
SLOW Query:
SELECT Count(EventID), IP
-> FROM event
-> WHERE AttackID >0
-> GROUP BY IP
-> ORDER BY Count(EventID) DESC
-> LIMIT 5;
RESULT: 5 rows in set (1.220 sec)
(This seems quite long for me, for a simple query)
QuerySlow
Now the Strange thing:
If I remove the foreign key relationship the performance of the query is the same.
But if I remove the the index on event.AttackID same select statement is much faster:
(ALTER TABLE `event` DROP INDEX `AttackID`;)
The result of the SQL SELECT query:
5 rows in set (0.242 sec)
QueryFast
From my understanding indexes on columns which are used in "WHERE" should improve the performance.
Why does removing the index have such an impact on the query?
What can I do to keep the relations between the table and have a faster
SELECT execution?
Cheers
Why does removing the index improve performance?
The query optimizer has multiple ways to resolve a query. For instance, two methods for filtering data are:
Look up the rows that match the where clause in the index and then fetch related data from the data pages.
Scan the index.
This doesn't get into the use of indexes for joins or aggregations or alternative algorithms.
Which is better? Under some circumstances, the first method is horribly slower than the second. This occurs when the data for the table does not fit into memory. Under such circumstances, the index can read a record from page 124 and then from 1068 and then from 124 again and -- well, all sorts of random intertwined reading of pages. Reading data pages in order is usually faster. And when the data doesn't fit into memory, thrashing occurs, which means that a page in memory is aged (overwritten) -- and then needed again.
I'm not saying that is occurring in your case. I am simply saying that what optimizers do is not always obvious. The optimizer has to make judgements based on the nature of the data -- and those judgements are not right 100% of the time. They are usually correct. But there are borderline cases. Sometimes, the issue is out-of-date statistics. Sometimes the issue is that what looks best to the optimizer is not best in practice.
Let me emphasize that optimizers usually do a very good job, and a better job than a person would do. Even if they occasionally come up with suboptimal plans, they are still quite useful.
Get rid of your redundant UNIQUE KEYs. A primary key is a unique key.
Use COUNT(*) rather than COUNT(IP) in your query. They mean the same thing because you declared IP to be NOT NULL.
Your query can be much faster if you stop saying WHERE AttackId>0. Because that column is a FK to the PK of your other table, those values should be nonzero anyway. But to get that speedup you'll need an index on event(IP) something like this.
CREATE INDEX IpDex ON event (IP)
But you're still summarizing a large table, and that will always take time.
It looks like you want to display some kind of leaderboard. You could add a top_ips table, and use an EVENT to populate it, using your query, every few minutes. Then you could display it to your users without incurring the cost of the query every time. This of course would display slightly stale data; only you know whether that's acceptable in your app.
Pro Tip. Read https://use-the-index-luke.com by Marcus Winand.
Essentially every part of your query, except for the FKey, conspires to make the query slow.
Your query is equivalent to
SELECT Count(*), IP
FROM event
WHERE AttackID >0
GROUP BY IP
ORDER BY Count(*) DESC
LIMIT 5;
Please use COUNT(*) unless you need to avoid NULL.
If AttackID is rarely >0, the optimal index is probably
ADD INDEX(AttackID, -- for filtering
IP) -- for covering
Else, the optimal index is probably
ADD INDEX(IP, -- to avoid sorting
AttackID) -- for covering
You could simply add both indexes and let the Optimizer decide. Meanwhile, get rid of these, if they exist:
DROP INDEX(AttackID)
DROP INDEX(IP)
because any uses of them are handled by the new indexes.
Furthermore, leaving the 1-column indexes around can confuse the Optimizer into using them instead of the covering index. (This seems to be a design flaw in at least some versions of MySQL/MariaDB.)
"Covering" means that the query can be performed entirely in the index's BTree. EXPLAIN will indicate it with "Using index". A "covering" index speeds up a query by 2x -- but there is a very wide variation on this prediction. ("Using index condition" is something different.)
More on index creation: http://mysql.rjweb.org/doc.php/index_cookbook_mysql
Hi I currently have a query which is taking 11(sec) to run. I have a report which is displayed on a website which runs 4 different queries which are similar and all take 11(sec) each to run. I don't really want the customer having to wait a minute for all of these queries to run and display the data.
I am using 4 different AJAX requests to call an APIs to get the data I need and these all start at once but the queries are running one after another. If there was a way to get these queries to all run at once (parallel) so the total load time is only 11(sec) that would also fix my issue, I don't believe that is possible though.
Here is the query I am running:
SELECT device_uuid,
day_epoch,
is_repeat
FROM tracking_daily_stats_zone_unique_device_uuids_per_hour
WHERE day_epoch >= 1552435200
AND day_epoch < 1553040000
AND venue_id = 46
AND zone_id IN (102,105,108,110,111,113,116,117,118,121,287)
I can't think of anyway to speed this query up at all, below are pictures of the table indexes and the explain statement on this query.
I think the above query is using relevant indexes in the where conditions.
If there is anything you can think of to speed this query up please let me know, I have been working on it for 3 days and can't seem to figure out the problem. It would be great to get the query times down to 5(sec) maximum. If I am wrong about the AJAX issue please let me know as this would also fix my issue.
" EDIT "
I have came across something quite strange which might be causing the issue. When I change the day_epoch range to something smaller (5th - 9th) which returns 130,000 rows the query time is 0.7(sec) but then I add one more day onto that range (5th - 10th) and it returns over 150,000 rows the query time is 13(sec). I have ran loads of different ranges and have came to the conclusion if the amount of rows returned is over 150,000 that has a huge effect on the query times.
Table Definition -
CREATE TABLE `tracking_daily_stats_zone_unique_device_uuids_per_hour` (
`id` int(11) NOT NULL AUTO_INCREMENT,
`day_epoch` int(10) NOT NULL,
`day_of_week` tinyint(1) NOT NULL COMMENT 'day of week, monday = 1',
`hour` int(2) NOT NULL,
`venue_id` int(5) NOT NULL,
`zone_id` int(5) NOT NULL,
`device_uuid` binary(16) NOT NULL COMMENT 'binary representation of the device_uuid, unique for a single day',
`device_vendor_id` int(5) unsigned NOT NULL DEFAULT '0' COMMENT 'id of the device vendor',
`first_seen` int(10) unsigned NOT NULL DEFAULT '0',
`last_seen` int(10) unsigned NOT NULL DEFAULT '0',
`is_repeat` tinyint(1) NOT NULL COMMENT 'is the device a repeat for this day?',
`prev_last_seen` int(10) NOT NULL DEFAULT '0' COMMENT 'previous last seen ts',
PRIMARY KEY (`id`,`venue_id`) USING BTREE,
KEY `venue_id` (`venue_id`),
KEY `zone_id` (`zone_id`),
KEY `day_of_week` (`day_of_week`),
KEY `day_epoch` (`day_epoch`),
KEY `hour` (`hour`),
KEY `device_uuid` (`device_uuid`),
KEY `is_repeat` (`is_repeat`),
KEY `device_vendor_id` (`device_vendor_id`)
) ENGINE=InnoDB AUTO_INCREMENT=450967720 DEFAULT CHARSET=utf8
/*!50100 PARTITION BY HASH (venue_id)
PARTITIONS 100 */
The straight forward solution is to add this query specific index to the table:
ALTER TABLE tracking_daily_stats_zone_unique_device_uuids_per_hour
ADD INDEX complex_idx (`venue_id`, `day_epoch`, `zone_id`)
WARNING This query change can take a while on DB.
And then force it when you call:
SELECT device_uuid,
day_epoch,
is_repeat
FROM tracking_daily_stats_zone_unique_device_uuids_per_hour
USE INDEX (complex_idx)
WHERE day_epoch >= 1552435200
AND day_epoch < 1553040000
AND venue_id = 46
AND zone_id IN (102,105,108,110,111,113,116,117,118,121,287)
It is definitely not universal but should work for this particular query.
UPDATE When you have partitioned table you can get profit by forcing particular PARTITION. In our case since that is venue_id just force it:
SELECT device_uuid,
day_epoch,
is_repeat
FROM tracking_daily_stats_zone_unique_device_uuids_per_hour
PARTITION (`p46`)
WHERE day_epoch >= 1552435200
AND day_epoch < 1553040000
AND zone_id IN (102,105,108,110,111,113,116,117,118,121,287)
Where p46 is concatenated string of p and venue_id = 46
And another trick if you go this way. You can remove AND venue_id = 46 from WHERE clause. Because there is no other data in that partition.
What happens if you change the order of conditions? Put venue_id = ? first. The order matters.
Now it first checks all rows for:
- day_epoch >= 1552435200
- then, the remaining set for day_epoch < 1553040000
- then, the remaining set for venue_id = 46
- then, the remaining set for zone_id IN (102,105,108,110,111,113,116,117,118,121,287)
When working with heavy queries, you should always try to make the first "selector" the most effective. You can do that by using a proper index for 1 (or combination) index and to make sure that first selector narrows down the most (at least for integers, in case of strings you need another tactic).
Sometimes, a query simply is slow. When you have a lot of data (and/or not enough resources) you just cant really do anything about that. Thats where you need another solution: Make a summary table. I doubt you show 150.000 rows x4 to your visitor. You can sum it, e.g., hourly or every few minutes and select from that way smaller table.
Offtopic: Putting an index on everything only slows you down when inserting/updating/deleting. Index the least amount of columns, just the once you actually filter on (e.g. use in a WHERE or GROUP BY).
450M rows is rather large. So, I will discuss a variety of issues that can help.
Shrink data A big table leads to more I/O, which is the main performance killer. ('Small' tables tend to stay cached, and not have an I/O burden.)
Any kind of INT, even INT(2) takes 4 bytes. An "hour" can easily fit in a 1-byte TINYINT. That saves over a 1GB in the data, plus a similar amount in INDEX(hour).
If hour and day_of_week can be derived, don't bother having them as separate columns. This will save more space.
Some reason to use a 4-byte day_epoch instead of a 3-byte DATE? Or perhaps you do need a 5-byte DATETIME or TIMESTAMP.
Optimal INDEX (take #1)
If it is always a single venue_id, then either this is a good first cut at the optimal index:
INDEX(venue_id, zone_id, day_epoch)
First is the constant, then the IN, then a range. The Optimizer does well with this in many cases. (It is unclear whether the number of items in an IN clause can lead to inefficiencies.)
Better Primary Key (better index)
With AUTO_INCREMENT, there is probably no good reason to include columns after the auto_inc column in the PK. That is, PRIMARY KEY(id, venue_id) is no better than PRIMARY KEY(id).
InnoDB orders the data's BTree according to the PRIMARY KEY. So, if you are fetching several rows and can arrange for them to be adjacent to each other based on the PK, you get extra performance. (cf "Clustered".) So:
PRIMARY KEY(venue_id, zone_id, day_epoch, -- this order, as discussed above;
id) -- to make sure that the entire PK is unique.
INDEX(id) -- to keep AUTO_INCREMENT happy
And, I agree with DROPping any indexes that are not in use, including the one I recommended above. It is rarely useful to index flags (is_repeat).
UUID
Indexing a UUID can be deadly for performance once the table is really big. This is because of the randomness of UUIDs/GUIDs, leading to ever-increasing I/O burden to insert new entries in the index.
Multi-dimensional
Assuming day_epoch is sometimes multiple days, you seem to have 2 or 3 "dimensions":
A date range
A list of zones
A venue.
INDEXes are 1-dimensional. Therein lies the problem. However, PARTITIONing can sometimes help. I discuss this briefly as "case 2" in http://mysql.rjweb.org/doc.php/partitionmaint .
There is no good way to get 3 dimensions, so let's focus on 2.
You should partition on something that is a "range", such as day_epoch or zone_id.
After that, you should decide what to put in the PRIMARY KEY so that you can further take advantage of "clustering".
Plan A: This assumes you are searching for only one venue_id at a time:
PARTITION BY RANGE(day_epoch) -- see note below
PRIMARY KEY(venue_id, zone_id, id)
Plan B: This assumes you sometimes srefineearch for venue_id IN (.., .., ...), hence it does not make a good first column for the PK:
Well, I don't have good advice here; so let's go with Plan A.
The RANGE expression must be numeric. Your day_epoch works fine as is. Changing to a DATE, would necessitate BY RANGE(TO_DAYS(...)), which works fine.
You should limit the number of partitions to 50. (The 81 mentioned above is not bad.) The problem is that "lots" of partitions introduces different inefficiencies; "too few" partitions leads to "why bother".
Note that almost always the optimal PK is different for a partitioned table than the equivalent non-partitioned table.
Note that I disagree with partitioning on venue_id since it is so easy to put that column at the start of the PK instead.
Analysis
Assuming you search for a single venue_id and use my suggested partitioning & PK, here's how the SELECT performs:
Filter on the date range. This is likely to limit the activity to a single partition.
Drill into the data's BTree for that one partition to find the one venue_id.
Hopscotch through the data from there, landing on the desired zone_ids.
For each, further filter based the date.
I have a simple MyISAM table resembling the following (trimmed for readability -- in reality, there are more columns, all of which are constant width and some of which are nullable):
CREATE TABLE IF NOT EXISTS `history` (
`id` bigint(20) NOT NULL AUTO_INCREMENT,
`time` int(11) NOT NULL,
`event` int(11) NOT NULL,
`source` int(11) DEFAULT NULL,
PRIMARY KEY (`id`),
KEY `event` (`event`),
KEY `time` (`time`),
);
Presently the table contains only about 6,000,000 rows (of which currently about 160,000 match the query below), but this is expected to increase. Given a particular event ID and grouped by source, I want to know how many events with that ID were logged during a particular interval of time. The answer to the query might be something along the lines of "Today, event X happened 120 times for source A, 105 times for source B, and 900 times for source C."
The query I concocted does perform this task, but it performs monstrously badly, taking well over a minute to execute when the timespan is set to "all time" and in excess of 30 seconds for as little as a week back:
SELECT COUNT(*) AS count FROM history
WHERE event=2000 AND time >= 0 AND time < 1310563644
GROUP BY source
ORDER BY count DESC
This is not for real-time use, so even if the query takes a second or two that would be fine, but several minutes is not. Explaining the query gives the following, which troubles me for obvious reasons:
id select_type table type possible_keys key key_len ref rows Extra
1 SIMPLE history ref event,time event 4 const 160399 Using where; Using temporary; Using filesort
I've experimented with various multi-column indexes (such as (event, time)), but with no improvement. This seems like such a common use case that I can't imagine there not being a reasonable solution, but my Googling all boil down to versions of the query I already have, with no particular suggestions on how to avoid the temporary (and even then, why performance is so abysmal).
Any suggestions?
You say you have tried multi-column indexes. Have you also tried single-column indexes, one per column?
UPDATE: Also, the COUNT(*) operation over a GROUP BY clause is probably a lot faster, if the grouped column also has an index on it... Of course, this depends on the number of NULL values that are actually in that column, which are not indexed.
For event, MySQL can execute a UNIQUE SCAN, which is quite fast, whereas for time, a RANGE SCAN will be applied, which is not so fast... If you separate indexes, I'd expect better performance than with multi-column ones.
Also, maybe you could gain something by partitioning your table by some expected values / value ranges:
http://dev.mysql.com/doc/refman/5.5/en/partitioning-overview.html
I offer you to try this multi-column index:
ALTER TABLE `history` ADD INDEX `history_index` (`event` ASC, `time` ASC, `source` ASC);
Then if it doesn't help, try to force index on this query:
SELECT COUNT(*) AS count FROM history USE INDEX (history_index)
WHERE event=2000 AND time >= 0 AND time < 1310563644
GROUP BY source
ORDER BY count DESC
If the source are known or you want to find the count for specific source, then you can try like this.
select count(source= 'A' or NULL) as A,count(source= 'B' or NULL) as B from history;
and for ordering you can do it in your application code. Also try with indexing event and source together.
This will be definitely faster than the older one.
I'm really struggling to get a query time down, its currently having to query 2.5 million rows and it takes over 20 seconds
here is the query
SELECT play_date AS date, COUNT(DISTINCT(email)) AS count
FROM log
WHERE play_date BETWEEN '2009-02-23' AND '2020-01-01'
AND type = 'play'
GROUP BY play_date
ORDER BY play_date desc;
`id` int(11) NOT NULL auto_increment,
`instance` varchar(255) NOT NULL,
`email` varchar(255) NOT NULL,
`type` enum('play','claim','friend','email') NOT NULL,
`result` enum('win','win-small','lose','none') NOT NULL,
`timestamp` timestamp NOT NULL default CURRENT_TIMESTAMP,
`play_date` date NOT NULL,
`email_refer` varchar(255) NOT NULL,
`remote_addr` varchar(15) NOT NULL,
PRIMARY KEY (`id`),
KEY `email` (`email`),
KEY `result` (`result`),
KEY `timestamp` (`timestamp`),
KEY `email_refer` (`email_refer`),
KEY `type_2` (`type`,`timestamp`),
KEY `type_4` (`type`,`play_date`),
KEY `type_result` (`type`,`play_date`,`result`)
id select_type table type possible_keys key key_len ref rows Extra
1 SIMPLE log ref type_2,type_4,type_result type_4 1 const 270404 Using where
The query is using the type_4 index.
Does anyone know how I could speed this query up?
Thanks
Tom
That's relatively good, already. The performance sink is that the query has to compare 270404 varchars for equality for the COUNT(DISTINCT(email)), meaning that 270404 rows have to be read.
You could be able to make the count faster by creating a covering index. This means that the actual rows do not need to be read because all the required information is present in the index itself.
To do this, change the index as follows:
KEY `type_4` (`type`,`play_date`, `email`)
I would be surprised if that wouldn't speed things up quite a bit.
(Thanks to MarkR for the proper term.)
Your indexing is probably as good as you can get it. You have a compound index on the 2 columns in your where clause and the explain you posted indicates that it is being used. Unfortunately, there are 270,404 rows that match the criteria in your where clause and they all need to be considered. Also, you're not returning unnecessary rows in your select list.
My advice would be to aggregate the data daily (or hourly or whatever makes sense) and cache the results. That way you can access slightly stale data instantly. Hopefully this is acceptable for your purposes.
Try an index on play_date, type (same as type_4, just reversed fields) and see if that helps
There are 4 possible types, and I assume 100's of possible dates. If the query uses the type, play_date index, it basically (not 100% accurate, but general idea) says.
(A) Find all the Play records (about 25% of the file)
(B) Now within that subset, find all of the requested dates
By reversing the index, the approach is
> (A) Find all the dates within range
> (Maybe 1-2% of file) (B) Now find all
> PLAY types within that smaller portion
> of the file
Hope this helps
Extracting email to separate table should be a good performance boost since counting distinct varchar fields should take awhile. Other than that - the correct index is used and the query itself is as optimized as it could be (except for the email, of course).
The COUNT(DISTINCT(email)) part is the bit that's killing you. If you only truly need the first 2000 results of 270,404, perhaps it would help to do the email count only for the results instead of for the whole set.
SELECT date, COUNT(DISTINCT(email)) AS count
FROM log,
(
SELECT play_date AS date
FROM log
WHERE play_date BETWEEN '2009-02-23' AND '2020-01-01'
AND type = 'play'
ORDER BY play_date desc
LIMIT 2000
) AS shortlist
WHERE shortlist.id = log.id
GROUP BY date
Try creating an index only on play_date.
Long term, I would recommend building a summary table with a primary key of play_date and count of distinct emails.
Depending on how up to date you need it to be - either allow it to be updated daily (by play_date) or live via a trigger on the log table.
There is a good chance a table scan will be quicker than random access to over 200,000 rows:
SELECT ... FROM log IGNORE INDEX (type_2,type_4,type_result) ...
Also, for large grouped queries you may see better performance by forcing a file sort rather than a hashtable-based group (since if this turns out to need more than tmp_table_size or max_heap_table_size performance collapses):
SELECT SQL_BIG_RESULT ...
I'm a relative novice when it comes to databases. We are using MySQL and I'm currently trying to speed up a SQL statement that seems to take a while to run. I looked around on SO for a similar question but didn't find one.
The goal is to remove all the rows in table A that have a matching id in table B.
I'm currently doing the following:
DELETE FROM a WHERE EXISTS (SELECT b.id FROM b WHERE b.id = a.id);
There are approximately 100K rows in table a and about 22K rows in table b. The column 'id' is the PK for both tables.
This statement takes about 3 minutes to run on my test box - Pentium D, XP SP3, 2GB ram, MySQL 5.0.67. This seems slow to me. Maybe it isn't, but I was hoping to speed things up. Is there a better/faster way to accomplish this?
EDIT:
Some additional information that might be helpful. Tables A and B have the same structure as I've done the following to create table B:
CREATE TABLE b LIKE a;
Table a (and thus table b) has a few indexes to help speed up queries that are made against it. Again, I'm a relative novice at DB work and still learning. I don't know how much of an effect, if any, this has on things. I assume that it does have an effect as the indexes have to be cleaned up too, right? I was also wondering if there were any other DB settings that might affect the speed.
Also, I'm using INNO DB.
Here is some additional info that might be helpful to you.
Table A has a structure similar to this (I've sanitized this a bit):
DROP TABLE IF EXISTS `frobozz`.`a`;
CREATE TABLE `frobozz`.`a` (
`id` bigint(20) unsigned NOT NULL auto_increment,
`fk_g` varchar(30) NOT NULL,
`h` int(10) unsigned default NULL,
`i` longtext,
`j` bigint(20) NOT NULL,
`k` bigint(20) default NULL,
`l` varchar(45) NOT NULL,
`m` int(10) unsigned default NULL,
`n` varchar(20) default NULL,
`o` bigint(20) NOT NULL,
`p` tinyint(1) NOT NULL,
PRIMARY KEY USING BTREE (`id`),
KEY `idx_l` (`l`),
KEY `idx_h` USING BTREE (`h`),
KEY `idx_m` USING BTREE (`m`),
KEY `idx_fk_g` USING BTREE (`fk_g`),
KEY `fk_g_frobozz` (`id`,`fk_g`),
CONSTRAINT `fk_g_frobozz` FOREIGN KEY (`fk_g`) REFERENCES `frotz` (`g`)
) ENGINE=InnoDB AUTO_INCREMENT=179369 DEFAULT CHARSET=utf8 ROW_FORMAT=DYNAMIC;
I suspect that part of the issue is there are a number of indexes for this table.
Table B looks similar to table B, though it only contains the columns id and h.
Also, the profiling results are as follows:
starting 0.000018
checking query cache for query 0.000044
checking permissions 0.000005
Opening tables 0.000009
init 0.000019
optimizing 0.000004
executing 0.000043
end 0.000005
end 0.000002
query end 0.000003
freeing items 0.000007
logging slow query 0.000002
cleaning up 0.000002
SOLVED
Thanks to all the responses and comments. They certainly got me to think about the problem. Kudos to dotjoe for getting me to step away from the problem by asking the simple question "Do any other tables reference a.id?"
The problem was that there was a DELETE TRIGGER on table A which called a stored procedure to update two other tables, C and D. Table C had a FK back to a.id and after doing some stuff related to that id in the stored procedure, it had the statement,
DELETE FROM c WHERE c.id = theId;
I looked into the EXPLAIN statement and rewrote this as,
EXPLAIN SELECT * FROM c WHERE c.other_id = 12345;
So, I could see what this was doing and it gave me the following info:
id 1
select_type SIMPLE
table c
type ALL
possible_keys NULL
key NULL
key_len NULL
ref NULL
rows 2633
Extra using where
This told me that it was a painful operation to make and since it was going to get called 22500 times (for the given set of data being deleted), that was the problem. Once I created an INDEX on that other_id column and reran the EXPLAIN, I got:
id 1
select_type SIMPLE
table c
type ref
possible_keys Index_1
key Index_1
key_len 8
ref const
rows 1
Extra
Much better, in fact really great.
I added that Index_1 and my delete times are in line with the times reported by mattkemp. This was a really subtle error on my part due to shoe-horning some additional functionality at the last minute. It turned out that most of the suggested alternative DELETE/SELECT statements, as Daniel stated, ended up taking essentially the same amount of time and as soulmerge mentioned, the statement was pretty much the best I was going to be able to construct based on what I needed to do. Once I provided an index for this other table C, my DELETEs were fast.
Postmortem:
Two lessons learned came out of this exercise. First, it is clear that I didn't leverage the power of the EXPLAIN statement to get a better idea of the impact of my SQL queries. That's a rookie mistake, so I'm not going to beat myself up about that one. I'll learn from that mistake. Second, the offending code was the result of a 'get it done quick' mentality and inadequate design/testing led to this problem not showing up sooner. Had I generated several sizable test data sets to use as test input for this new functionality, I'd have not wasted my time nor yours. My testing on the DB side was lacking the depth that my application side has in place. Now I've got the opportunity to improve that.
Reference: EXPLAIN Statement
Deleting data from InnoDB is the most expensive operation you can request of it. As you already discovered the query itself is not the problem - most of them will be optimized to the same execution plan anyway.
While it may be hard to understand why DELETEs of all cases are the slowest, there is a rather simple explanation. InnoDB is a transactional storage engine. That means that if your query was aborted halfway-through, all records would still be in place as if nothing happened. Once it is complete, all will be gone in the same instant. During the DELETE other clients connecting to the server will see the records until your DELETE completes.
To achieve this, InnoDB uses a technique called MVCC (Multi Version Concurrency Control). What it basically does is to give each connection a snapshot view of the whole database as it was when the first statement of the transaction started. To achieve this, every record in InnoDB internally can have multiple values - one for each snapshot. This is also why COUNTing on InnoDB takes some time - it depends on the snapshot state you see at that time.
For your DELETE transaction, each and every record that is identified according to your query conditions, gets marked for deletion. As other clients might be accessing the data at the same time, it cannot immediately remove them from the table, because they have to see their respective snapshot to guarantee the atomicity of the deletion.
Once all records have been marked for deletion, the transaction is successfully committed. And even then they cannot be immediately removed from the actual data pages, before all other transactions that worked with a snapshot value before your DELETE transaction, have ended as well.
So in fact your 3 minutes are not really that slow, considering the fact that all records have to be modified in order to prepare them for removal in a transaction safe way. Probably you will "hear" your hard disk working while the statement runs. This is caused by accessing all the rows.
To improve performance you can try to increase InnoDB buffer pool size for your server and try to limit other access to the database while you DELETE, thereby also reducing the number of historic versions InnoDB has to maintain per record.
With the additional memory InnoDB might be able to read your table (mostly) into memory and avoid some disk seeking time.
Try this:
DELETE a
FROM a
INNER JOIN b
on a.id = b.id
Using subqueries tend to be slower then joins as they are run for each record in the outer query.
This is what I always do, when I have to operate with super large data (here: a sample test table with 150000 rows):
drop table if exists employees_bak;
create table employees_bak like employees;
insert into employees_bak
select * from employees
where emp_no > 100000;
rename table employees to employees_todelete;
rename table employees_bak to employees;
drop table employees_todelete;
In this case the sql filters 50000 rows into the backup table.
The query cascade performs on my slow machine in 5 seconds.
You can replace the insert into select by your own filter query.
That is the trick to perform mass deletion on big databases!;=)
Your time of three minutes seems really slow. My guess is that the id column is not being indexed properly. If you could provide the exact table definition you're using that would be helpful.
I created a simple python script to produce test data and ran multiple different versions of the delete query against the same data set. Here's my table definitions:
drop table if exists a;
create table a
(id bigint unsigned not null primary key,
data varchar(255) not null) engine=InnoDB;
drop table if exists b;
create table b like a;
I then inserted 100k rows into a and 25k rows into b (22.5k of which were also in a). Here's the results of the various delete commands. I dropped and repopulated the table between runs by the way.
mysql> DELETE FROM a WHERE EXISTS (SELECT b.id FROM b WHERE a.id=b.id);
Query OK, 22500 rows affected (1.14 sec)
mysql> DELETE FROM a USING a LEFT JOIN b ON a.id=b.id WHERE b.id IS NOT NULL;
Query OK, 22500 rows affected (0.81 sec)
mysql> DELETE a FROM a INNER JOIN b on a.id=b.id;
Query OK, 22500 rows affected (0.97 sec)
mysql> DELETE QUICK a.* FROM a,b WHERE a.id=b.id;
Query OK, 22500 rows affected (0.81 sec)
All the tests were run on an Intel Core2 quad-core 2.5GHz, 2GB RAM with Ubuntu 8.10 and MySQL 5.0. Note, that the execution of one sql statement is still single threaded.
Update:
I updated my tests to use itsmatt's schema. I slightly modified it by remove auto increment (I'm generating synthetic data) and character set encoding (wasn't working - didn't dig into it).
Here's my new table definitions:
drop table if exists a;
drop table if exists b;
drop table if exists c;
create table c (id varchar(30) not null primary key) engine=InnoDB;
create table a (
id bigint(20) unsigned not null primary key,
c_id varchar(30) not null,
h int(10) unsigned default null,
i longtext,
j bigint(20) not null,
k bigint(20) default null,
l varchar(45) not null,
m int(10) unsigned default null,
n varchar(20) default null,
o bigint(20) not null,
p tinyint(1) not null,
key l_idx (l),
key h_idx (h),
key m_idx (m),
key c_id_idx (id, c_id),
key c_id_fk (c_id),
constraint c_id_fk foreign key (c_id) references c(id)
) engine=InnoDB row_format=dynamic;
create table b like a;
I then reran the same tests with 100k rows in a and 25k rows in b (and repopulating between runs).
mysql> DELETE FROM a WHERE EXISTS (SELECT b.id FROM b WHERE a.id=b.id);
Query OK, 22500 rows affected (11.90 sec)
mysql> DELETE FROM a USING a LEFT JOIN b ON a.id=b.id WHERE b.id IS NOT NULL;
Query OK, 22500 rows affected (11.48 sec)
mysql> DELETE a FROM a INNER JOIN b on a.id=b.id;
Query OK, 22500 rows affected (12.21 sec)
mysql> DELETE QUICK a.* FROM a,b WHERE a.id=b.id;
Query OK, 22500 rows affected (12.33 sec)
As you can see this is quite a bit slower than before, probably due to the multiple indexes. However, it is nowhere near the three minute mark.
Something else that you might want to look at is moving the longtext field to the end of the schema. I seem to remember that mySQL performs better if all the size restricted fields are first and text, blob, etc are at the end.
You're doing your subquery on 'b' for every row in 'a'.
Try:
DELETE FROM a USING a LEFT JOIN b ON a.id = b.id WHERE b.id IS NOT NULL;
Try this out:
DELETE QUICK A.* FROM A,B WHERE A.ID=B.ID
It is much faster than normal queries.
Refer for Syntax : http://dev.mysql.com/doc/refman/5.0/en/delete.html
I know this question has been pretty much solved due to OP's indexing omissions but I would like to offer this additional advice, which is valid for a more generic case of this problem.
I have personally dealt with having to delete many rows from one table that exist in another and in my experience it's best to do the following, especially if you expect lots of rows to be deleted. This technique most importantly will improve replication slave lag, as the longer each single mutator query runs, the worse the lag would be (replication is single threaded).
So, here it is: do a SELECT first, as a separate query, remembering the IDs returned in your script/application, then continue on deleting in batches (say, 50,000 rows at a time).
This will achieve the following:
each one of the delete statements will not lock the table for too long, thus not letting replication lag to get out of control. It is especially important if you rely on your replication to provide you relatively up-to-date data. The benefit of using batches is that if you find that each DELETE query still takes too long, you can adjust it to be smaller without touching any DB structures.
another benefit of using a separate SELECT is that the SELECT itself might take a long time to run, especially if it can't for whatever reason use the best DB indexes. If the SELECT is inner to a DELETE, when the whole statement migrates to the slaves, it will have to do the SELECT all over again, potentially lagging the slaves because it has to do the long select all over again. Slave lag, again, suffers badly. If you use a separate SELECT query, this problem goes away, as all you're passing is a list of IDs.
Let me know if there's a fault in my logic somewhere.
For more discussion on replication lag and ways to fight it, similar to this one, see MySQL Slave Lag (Delay) Explained And 7 Ways To Battle It
P.S. One thing to be careful about is, of course, potential edits to the table between the times the SELECT finishes and DELETEs start. I will let you handle such details by using transactions and/or logic pertinent to your application.
DELETE FROM a WHERE id IN (SELECT id FROM b)
Maybe you should rebuild the indicies before running such a hugh query. Well, you should rebuild them periodically.
REPAIR TABLE a QUICK;
REPAIR TABLE b QUICK;
and then run any of the above queries (i.e.)
DELETE FROM a WHERE id IN (SELECT id FROM b)
The query itself is already in an optimal form, updating the indexes causes the whole operation to take that long. You could disable the keys on that table before the operation, that should speed things up. You can turn them back on at a later time, if you don't need them immediately.
Another approach would be adding a deleted flag-column to your table and adjusting other queries so they take that value into account. The fastest boolean type in mysql is CHAR(0) NULL (true = '', false = NULL). That would be a fast operation, you can delete the values afterwards.
The same thoughts expressed in sql statements:
ALTER TABLE a ADD COLUMN deleted CHAR(0) NULL DEFAULT NULL;
-- The following query should be faster than the delete statement:
UPDATE a INNER JOIN b SET a.deleted = '';
-- This is the catch, you need to alter the rest
-- of your queries to take the new column into account:
SELECT * FROM a WHERE deleted IS NULL;
-- You can then issue the following queries in a cronjob
-- to clean up the tables:
DELETE FROM a WHERE deleted IS NOT NULL;
If that, too, is not what you want, you can have a look at what the mysql docs have to say about the speed of delete statements.
BTW, after posting the above on my blog, Baron Schwartz from Percona brought to my attention that his maatkit already has a tool just for this purpose - mk-archiver. http://www.maatkit.org/doc/mk-archiver.html.
It is most likely your best tool for the job.
Obviously the SELECT query that builds the foundation of your DELETE operation is quite fast so I'd think that either the foreign key constraint or the indexes are the reasons for your extremely slow query.
Try
SET foreign_key_checks = 0;
/* ... your query ... */
SET foreign_key_checks = 1;
This would disable the checks on the foreign key. Unfortunately you cannot disable (at least I don't know how) the key-updates with an InnoDB table. With a MyISAM table you could do something like
ALTER TABLE a DISABLE KEYS
/* ... your query ... */
ALTER TABLE a ENABLE KEYS
I actually did not test if these settings would affect the query duration. But it's worth a try.
Connect datebase using terminal and execute command below, look at the result time each of them, you'll find that times of delete 10, 100, 1000, 10000, 100000 records are not Multiplied.
DELETE FROM #{$table_name} WHERE id < 10;
DELETE FROM #{$table_name} WHERE id < 100;
DELETE FROM #{$table_name} WHERE id < 1000;
DELETE FROM #{$table_name} WHERE id < 10000;
DELETE FROM #{$table_name} WHERE id < 100000;
The time of deleting 10 thousand records is not 10 times as much as deleting 100 thousand records.
Then, except for finding a way delete records more faster, there are some indirect methods.
1, We can rename the table_name to table_name_bak, and then select records from table_name_bak to table_name.
2, To delete 10000 records, we can delete 1000 records 10 times. There is an example ruby script to do it.
#!/usr/bin/env ruby
require 'mysql2'
$client = Mysql2::Client.new(
:as => :array,
:host => '10.0.0.250',
:username => 'mysql',
:password => '123456',
:database => 'test'
)
$ids = (1..1000000).to_a
$table_name = "test"
until $ids.empty?
ids = $ids.shift(1000).join(", ")
puts "delete =================="
$client.query("
DELETE FROM #{$table_name}
WHERE id IN ( #{ids} )
")
end
The basic technique for deleting multiple Row form MySQL in single table through the id field
DELETE FROM tbl_name WHERE id <= 100 AND id >=200;
This query is responsible for deleting the matched condition between 100 AND 200 from the certain table