Mysql – Large table delete in thesql

Here's a method that does not interrupt service on the master, though it does lock the slave temporarily.

1. Stop the slave.

   mysql> STOP SLAVE;
   

2. Dump the offending table on the master with mysqldump in a repeatable-read transaction. Include the binary log coordinates in the output (--master-data). Output one INSERT per row (--skip-extended-insert).

   $ mysqldump --single-transaction --master-data=2 --skip-extended-insert 
     mydatabase myoffendingtable > master-dump.sql
   

3. Note the binlog coordinates in the dump file, and start slave until the coordinate shown in the master's dump file.

   mysql> START SLAVE UNTIL MASTER_LOG_FILE='xxxx' MASTER_LOG_POS=yyyy;
   

4. Wait until the slave thread catches up to that position and stops itself. This shouldn't take long.

   mysql> SELECT MASTER_POS_WAIT('xxxx', yyyy);
   

5. Dump the table on the slave, also with one row per INSERT.

   $ mysqldump --skip-extended-insert mydatabase myoffendingtable > slave-dump.sql
   

6. Resume replication.

   mysql> START SLAVE;
   

Now you can diff the dump files, and be assured they represent exactly the same logical point in the stream of changes. This allows you to find data discrepancies.

In MySQL there is a multi-table DELETE syntax. Your first DELETE will delete rows only from the players TABLE. If you want to delete from multiple tables you have to use something like:

DELETE FROM "players","stats","photos"
USING "players"
  LEFT JOIN "stats"
    ON "players".id = "stats".player_id
  LEFT JOIN "photos"
    ON "players".id = "photos".player_id
WHERE "players".born_on < "1950-01-01"

This doesn't address the problem with the long-running DELETE statement though. In fact above query should take even more time, because now it actually would delete rows from stats and photos tables. The workaround you could use is to split the large DELETE into smaller ones. Since you have a nice WHERE condition, you could manually split the deletes on that (for example one DELETE for each ten years of players.born_on) and run them in ascending order, that is:

DELETE FROM "players","stats","photos"
USING "players"
  LEFT JOIN "stats"
    ON "players".id = "stats".player_id
  LEFT JOIN "photos"
    ON "players".id = "photos".player_id
WHERE "players".born_on < "1930-01-01";

DELETE FROM "players","stats","photos"
USING "players"
  LEFT JOIN "stats"
    ON "players".id = "stats".player_id
  LEFT JOIN "photos"
    ON "players".id = "photos".player_id
WHERE "players".born_on < "1940-01-01";

DELETE FROM "players","stats","photos"
USING "players"
  LEFT JOIN "stats"
    ON "players".id = "stats".player_id
  LEFT JOIN "photos"
    ON "players".id = "photos".player_id
WHERE "players".born_on < "1950-01-01";

It this is too coarse (i.e. it takes too long do execute each query) you should make the WHERE conditions even more fine-grained (perhaps delete one year each chunk).

Also there is a --purge option for pt-archiver from Percona Toolkit which would split the data to be deleted in chunks automatically, but it doesn't seem to support the multi table case. See example usage of pt-archiver in this presentation

From your table definitions I see that you don't have an index on players.birth_date (I suppose that this is the column you relate to as born_at in your example queries). This makes the decade chunks approach useless, since every query would have to scan all players table.

If you can't afford having a long table lock for the DELETE to finish, you most likely can't afford to create an index on the birth_date column as well.

You could split the data on another column, PRIMARY KEY is a good bet. You can write a script which would process all the players in chunks of 10000 (or less or more, depending on the length of a single DELETE statement):

DELETE FROM "players","stats","photos"
USING "players"
  LEFT JOIN "stats"
    ON "players".id = "stats".player_id
  LEFT JOIN "photos"
    ON "players".id = "photos".player_id
WHERE "players".born_on < "1950-01-01"
  AND "players".id BETWEEN n*10000+1 AND (n+1)*10000

Where n would be a parameter ranging from 0 to MAX(players.id)/10000 . This way you will avoid a full table scan (which certainly is painful for a 100M table)

You could also try to estimate the DELETE complexity with an EXPLAIN SELECT instead of DELETE:

EXPLAIN SELECT *
FROM "players"
  LEFT JOIN "stats"
    ON "players".id = "stats".player_id
  LEFT JOIN "photos"
    ON "players".id = "photos".player_id
WHERE "players".born_on < "1950-01-01"
  AND "players".id BETWEEN 1 AND 10000