MySQL Index creation failing on table is full

UPDATE: tl;dr: The problem was MySQL uses the TMPDIR when creating indexes. And my TMPDIR was the one running out of disk space.

Original Q:

I'm trying to add an index to an InnoDB table, and getting a table is full error.
I have enough disk space and the MySQL configuration has a file-per-table=1. The table data is 85GB and I assume the index will be around 20GB - 30GB and I have much more disk space than that. I'm also using ext3 so I don't feel there's any problem with the file size limit from the OS point of view.

The logged error looks like this:

140616 13:04:33  InnoDB: Error: Write to file (merge) failed at offset 3 1940914176.
InnoDB: 1048576 bytes should have been written, only 970752 were written.
InnoDB: Operating system error number 0.
InnoDB: Check that your OS and file system support files of this size.
InnoDB: Check also that the disk is not full or a disk quota exceeded.
InnoDB: Error number 0 means 'Success'.
InnoDB: Some operating system error numbers are described at
InnoDB: http://dev.mysql.com/doc/refman/5.5/en/operating-system-error-codes.html
140616 13:04:33 [ERROR] /usr/libexec/mysqld: The table 'my_table' is full

What is causing this and how can I resolve?

The create table:

`CREATE TABLE `my_table` (
 `uid_from` bigint(11) NOT NULL,
 `uid_to` bigint(11) NOT NULL,
 `counter` int(11) NOT NULL,
 `updated` date NOT NULL,
 PRIMARY KEY (`uid_to`,`uid_from`)
) ENGINE=InnoDB DEFAULT CHARSET=utf8`

As is the Data is 87.4GB and I estimate there are around 1.5B rows.

SHOW GLOBAL VARIABLES LIKE 'tmpdir';
Variable_name   Value
tmpdir  /tmp

[root@web ~]# df -h /tmp
Filesystem      Size  Used Avail Use% Mounted on
/dev/xvda1       40G   24G   14G  64% /

Please don't fooled by the error message The table 'my_table' is full. This rare scenario has absolutely nothing to do with diskspace. This table full condition has to do with the internal plumbing of InnoDB.

First, take a look at this diagram of the InnoDB Architecture

Please note that the system tablespace (the file ibdata) only has 128 Rollback Segments and 1023 Rollback Slots Per Rollback Segment. This places limits on the size of a transaction's rollback capacity. In other words, if a single rollback segment needs more than 1023 slots to support a transaction, the transaction will hit that table is full condition.

Think of the restaurant Red Lobster in New Jersey. It may have a capacity of 200 people. If the restaurant is full, a line of people may go outside to wait. If the people on the line get impatient, they may leave because the restaurant is full. Obviously, the solution would not be to make New Jersey bigger (or getting more diskspace). The solution would be to make the Red Lobster restaurant bigger. That way you can increase the seating capacity to, let's say, 240. Even with that, a line may form outside if more than 240 people decide to come to Red Lobster.

Just to give you an example, I had a client with 2TB of system tablespace and innodb_file_per_table was disabled. (346G for ibdata1, the reset for ibdata2). I ran this query

SELECT SUM(data_length+index+length) InnoDBDataIndexSpace
FROM information_schema.tables WHERE engine='InnoDB';

Next, I substracted InnoDBDataIndexSpace from the sum of the filesizes for ibdata1 and ibdata2. I got two things that shocked me

• There was 106GB remaining inside the system tablespace.

• I got this same Table is Full condition

This means that the 106GB was using for InnoDB's internal plumbing. The client was using ext3 at the time.

The solution for me was to add ibdata3. I discussed this in my old post How to solve "The table ... is full" with "innodb_file_per_table"?

I have discussed this in other posts as well

• Mar 31, 2014 : mysql directory grow to 246G after one query, which failed due to table is full

• Nov 25, 2011 : ERROR 1114 (HY000) at line 6308 in file & The table user_analysis is full

Keep in mind that this condition can happen even if innodb_file_per_table was enabled. How? Rollbacks and Undo Logs are the source of uncontrolled spikes is growth for ibdata1.

YOUR ACTUAL QUESTION

Since your are adding an index to a table and getting Table is Full, the table must be huge and cannot fit inside a single rollback segment. You must do the following:

STEP 01

Get the data into a dump file

mysqldump --no-create-info mydb mytable > table_data.sql

STEP 02

Login to MySQL and run this

USE mydb
CREATE TABLE mytable_new LIKE mytable;
ALTER TABLE mytable_new ADD INDEX ... ;
ALTER TABLE mytable RENAME mytable_old;
ALTER TABLE mytable_new RENAME mytable;

STEP 03

Load the table with the additional index with the data

mysql -Dmydb < table_data.sql

That's all.

See, the problem is that the ALTER TABLE will try to inject all the rows in your huge table as a single transaction. Using the mysqldump will insert the data into table (now with a new index) thousands of rows at a time, not all the rows in a single transaction.

Don't worry about what if this doesn't work? The original table will be named mytable_old in case of anything. it can serve as a backup. You can drop the backup when you know the new table works for you.

Give it a Try !!!

UPDATE 2014-06-16 11:13 EDT

If you are worried about the dump of the data bigger huge, just gzip it.

You can do the same steps, but as follows

STEP 01

Get the data into a dump file

mysqldump --no-create-info mydb mytable | gzip > table_data.sql.gz

STEP 02

Login to MySQL and run this

USE mydb
CREATE TABLE mytable_new LIKE mytable;
ALTER TABLE mytable_new ADD INDEX ... ;
ALTER TABLE mytable RENAME mytable_old;
ALTER TABLE mytable_new RENAME mytable;

STEP 03

Load the table with the additional index with the data

gzip -d < table_data.sql.gz | mysql -Dmydb

or

gunzip < table_data.sql.gz | mysql -Dmydb

UPDATE 2014-06-16 12:55 EDT

I just thought of another aspect with regard to this issue.

Since you are doing DDL and not DML, it is possible that this is not InnoDB internal plumbing. Since DDL cannot rollback for InnoDB, the issue has to be external plumbing. Where is this external plumbing getting clogged? I suspect the temp folder for the OS. Why?

```
140616 13:04:33 InnoDB: Error: Write to file (merge) failed at offset 3 1940914176.
InnoDB: 1048576 bytes should have been written, only 970752 were written.
InnoDB: Operating system error number 0.
InnoDB: Check that your OS and file system support files of this size.
InnoDB: Check also that the disk is not full or a disk quota exceeded.
InnoDB: Error number 0 means 'Success'.
InnoDB: Some operating system error numbers are described at
InnoDB: http://dev.mysql.com/do

StackExchange Database Administrators Q#68319, answer score: 9