Percona Toolkit’s pt-table-checksum is a great tool to find data inconsistencies between a MySQL master and its replicas. However it is sometimes not enough to know that there are inconsistencies and let pt-table-sync fix the issue: you may want to know which exact rows are different to identify the statements that created the inconsistency. This post shows one way to achieve that goal.
The issue
Let’s assume you have 2 servers running MySQL 5.5: db1 the master and db2 the replica. You want to upgrade to MySQL 5.6 using an in-place upgrade and to play safe, you will upgrade db2 (the slave) first. If all goes well you will promote it and upgrade db1.
A good thing to do after upgrading db2 is to check for potential data inconsistencies with pt-table-checksum. Once checksumming is done, you can run the following query on db2 to see if there is any data drift:
mysql> SELECT db, tbl, SUM(this_cnt) AS total_rows, COUNT(*) AS chunks FROM percona.checksums WHERE (master_cnt <> this_cnt OR master_crc <> this_crc OR ISNULL(master_crc) <> ISNULL(this_crc)) GROUP BY db, tbl; +------+-------+------------+--------+ | db | tbl | total_rows | chunks | +------+-------+------------+--------+ | mydb | items | 3745563 | 17 | +------+-------+------------+--------+
This indicates that inconsistencies can be found in mydb.items in 17 chunks. Now the question is: which rows are different on db1 and db2?
The solution
The previous query shows that we will find inconsistencies in 17 of the chunks pt-table-checksum used. But what is a chunk?
mysql> SELECT * FROM percona.checksums WHERE this_crc != master_crc AND tbl='items'G ***************** 1. row ***************** db: mydb tbl: items chunk: 28 chunk_time: 0.123122 chunk_index: PRIMARY lower_boundary: 7487511 upper_boundary: 7563474 this_crc: 2c11da8d this_cnt: 75964 master_crc: 66a1c22c master_cnt: 75964 ts: 2014-10-22 01:21:26 [...]
So the first chunk with inconsistencies is chunk #28, which is the set of rows where the primary key is >= 7487511 and <= 7563474.
Let’s export all these rows on db1 and db2 instance ::
# db1 mysql> SELECT * INTO outfile '/tmp/items_db1.txt' FROM mydb.items WHERE id BETWEEN 7487511 AND 7563474; # db2 mysql> SELECT * INTO outfile '/tmp/items_db2.txt' FROM mydb.items WHERE id BETWEEN 7487511 AND 7563474;
Then let’s use diff to isolate non-matching rows
# Using diff to compare rows # diff items_db1.txt items_db2.txt 75872,75874c75872,75874 < 7563382 2127002 3 0 2014-10-22 02:51:33 < 7563383 2127002 4 0 2014-10-22 02:51:33 < 7563384 2127002 5 0 2014-10-22 02:51:33 --- > 7563382 2127002 3 0 2014-10-22 02:51:34 > 7563383 2127002 4 0 2014-10-22 02:51:34 > 7563384 2127002 5 0 2014-10-22 02:51:34 [...]
We can see that some datetime fields are off by 1 second on the 5.6 instance.
In this case, the binlogs showed queries like:
INSERT INTO items ([...],posted_at) VALUES ([...],'2014-10-22 02:51:33.835249');
MySQL 5.5 rounds '2014-10-22 02:51:33.835249'
to '2014-10-22 02:51:33'
(ignoring the fractional part), while MySQL 5.6 rounds it to '2014-10-22 02:51:34'
.
Now it’s easy to fix the application so that it works both with MySQL 5.5 and 5.6 and then continue testing MySQL 5.6.
Conclusion
The method shown above is an easy way to find the exact records that are inconsistent between the MySQL master and a replica. It is not useful if you only want to resync the slave (in this case, just run pt-table-sync) but it can be a first step in understanding how inconsistencies are created.
The post Data inconsistencies on MySQL replicas: Beyond pt-table-checksum appeared first on MySQL Performance Blog.