InnoDB Online DDL Overview

In MariaDB 10.0 and later, the ALTER TABLE statement supports two clauses that are used to implement online DDL:

• ALGORITHM - This clause controls how the DDL operation is performed.
• LOCK - This clause controls how much concurrency is allowed while the DDL operation is being performed.

Alter Algorithms

In MariaDB 10.0 and later, InnoDB supports multiple algorithms for performing DDL operations. This offers a significant performance improvement over previous versions. The supported algorithms are:

• DEFAULT - This implies the default behavior for the specific statement.
• COPY
• INPLACE
• NOCOPY - This was added in MariaDB 10.3.7.
• INSTANT - This was added in MariaDB 10.3.7.

Specifying an Alter Algorithm

Specifying an Alter Algorithm Using the ALGORITHM Clause

The ALGORITHM clause for the ALTER TABLE statement can be used to pick the least efficient algorithm that the user is willing to accept.

Specifying an Alter Algorithm Using System Variables

Supported Alter Algorithms

The supported algorithms are described in more details below.

DEFAULT Algorithm

The default behavior, which occurs if ALGORITHM=DEFAULT is specified, or if ALGORITHM is not specified at all, usually only makes a copy if the operation doesn't support being done in-place at all. In this case, the most efficient available algorithm will usually be used.

This means that, if an operation supports the INSTANT algorithm, then it will use that algorithm by default. If an operation does not support the INSTANT algorithm, but it does support the NOCOPY algorithm, then it will use that algorithm by default. If an operation does not support the NOCOPY algorithm, but it does support the INPLACE algorithm, then it will use that algorithm by default.

COPY Algorithm

When the COPY algorithm is used, MariaDB essentially does the following operations:

-- Create a temporary table with the new definition
CREATE TEMPORARY TABLE tmp_tab (
...
);

-- Copy the data from the original table
INSERT INTO tmp_tab
   SELECT * FROM original_tab;

-- Drop the original table
DROP TABLE original_tab;

-- Rename the temporary table, so that it replaces the original one
RENAME TABLE tmp_tab TO original_tab;

This algorithm is very inefficient, but it is generic, so it works for all storage engines.

If the COPY algorithm is specified with the ALGORITHM clause or with the alter_algorithm system variable, then the COPY algorithm will be used even if it is not necessary. This can result in a lengthy table copy. If multiple ALTER TABLE operations are required that each require the table to be rebuilt, then it is best to specify all operations in a single ALTER TABLE statement, so that the table is only rebuilt once.

Using the COPY Algorithm with InnoDB

If the COPY algorithm is used with an InnoDB table, then the following statements apply:

• The table will be rebuilt using the current values of the innodb_file_per_table, innodb_file_format, and innodb_default_row_format system variables.

• The operation will have to create a temporary table to perform the the table copy. This temporary table will be in the same directory as the original table, and it's file name will be in the format #sql${PID}_${THREAD_ID}_${TMP_TABLE_COUNT}, where ${PID} is the process ID of mysqld, ${THREAD_ID} is the connection ID, and ${TMP_TABLE_COUNT} is the number of temporary tables that the connection has open. Therefore, the datadir may contain files with file names like #sql1234_12_1.ibd.

• The operation inserts one record at a time into each index, which is very inefficient.

• InnoDB does not use a sort buffer.

• In MariaDB 10.2.13, MariaDB 10.3.5 and later, the table copy operation creates a lot fewer InnoDB undo log writes. See MDEV-11415 for more information.

• The table copy operation creates a lot of InnoDB redo log writes.

INPLACE Algorithm

The COPY algorithm can be incredibly slow, because the whole table has to be copied and rebuilt. The INPLACE algorithm was introduced as a way to avoid this by performing operations in-place and avoiding the table copy and rebuild, when possible.

When the INPLACE algorithm is used, the underlying storage engine uses optimizations to perform the operation while avoiding the table copy and rebuild. However, INPLACE is a bit of a misnomer, since some operations may still require the table to be rebuilt for some storage engines. Regardless, several operations can be performed without a full copy of the table for some storage engines.

A more accurate name for the algorithm would have been the ENGINE algorithm, since the storage engine decides how to implement the algorithm.

If an ALTER TABLE operation supports the INPLACE algorithm, then it can be performed using optimizations by the underlying storage engine, but it may rebuilt.

If the INPLACE algorithm is specified with the ALGORITHM clause or with the alter_algorithm system variable and if the ALTER TABLE operation does not support the INPLACE algorithm, then an error will be raised. For example:

SET SESSION alter_algorithm='INPLACE';

ALTER TABLE tab MODIFY COLUMN c int;
ERROR 1846 (0A000): ALGORITHM=INPLACE is not supported. Reason: Cannot change column type INPLACE. Try ALGORITHM=COPY

In this case, raising an error is preferable, if the alternative is for the operation to make a copy of the table, and perform unexpectedly slowly.

Using the INPLACE Algorithm with InnoDB

If the INPLACE algorithm is used with an InnoDB table, then the following statements apply:

• The operation might have to write sort files in the directory defined by the innodb_tmpdir system variable.

• The operation might also have to write a temporary log file to track data changes by DML queries executed during the operation. The maximum size for this log file is configured by the innodb_online_alter_log_max_size system variable.

• Some operations require the table to be rebuilt, even though the algorithm is inaccurately called "in-place". This includes operations such as adding or dropping columns, adding a primary key, changing a column to NULL, etc.

• If the operation requires the table to be rebuilt, then the operation might have to create temporary tables.
  • It may have to create a temporary intermediate table for the actual table rebuild operation.
    • In MariaDB 10.2.19 and later, this temporary table will be in the same directory as the original table, and it's file name will be in the format #sql${PID}_${THREAD_ID}_${TMP_TABLE_COUNT}, where ${PID} is the process ID of mysqld, ${THREAD_ID} is the connection ID, and ${TMP_TABLE_COUNT} is the number of temporary tables that the connection has open. Therefore, the datadir may contain files with file names like #sql1234_12_1.ibd.
    • In MariaDB 10.2.18 and before, this temporary table will be in the same directory as the original table, and it's file name will be in the format #sql-ib${TABLESPACE_ID}-${RAND}, where ${TABLESPACE_ID} is the table's tablespace ID within InnoDB and ${RAND} is a randomly initialized number. Therefore, the datadir may contain files with file names like #sql-ib230291-1363966925.ibd.
  • When it replaces the original table with the rebuilt table, it may also have to rename the original table using a temporary table name.
    • If the server is MariaDB 10.3 or later or if it is running MariaDB 10.2 and the innodb_safe_truncate system variable is set to OFF, then the format will actually be #sql-ib${TABLESPACE_ID}-${RAND}, where ${TABLESPACE_ID} is the table's tablespace ID within InnoDB and ${RAND} is a randomly initialized number. Therefore, the datadir may contain files with file names like #sql-ib230291-1363966925.ibd.
    • If the server is running MariaDB 10.1 or before or if it is running MariaDB 10.2 and the innodb_safe_truncate system variable is set to ON, then the renamed table will have a temporary table name in the format #sql-ib${TABLESPACE_ID}, where ${TABLESPACE_ID} is the table's tablespace ID within InnoDB. Therefore, the datadir may contain files with file names like #sql-ib230291.ibd.

• The storage needed for the above items can add up to the size of the original table, or more in some cases.

• Some operations are instantaneous, if they only require the table's metadata to be changed. This includes operations such as renaming a column, changing a column's DEFAULT value, etc.

Operations Supported by InnoDB with the INPLACE Algorithm

With respect to the allowed operations, the INPLACE algorithm supports a subset of the operations supported by the COPY algorithm, and it supports a superset of the operations supported by the NOCOPY algorithm.

See InnoDB Online DDL Operations with ALGORITHM=INPLACE for more information.

NOCOPY Algorithm

The INPLACE algorithm can sometimes be surprisingly slow in instances where it has to rebuild the clustered index, because when the clustered index has to be rebuilt, the whole table has to be rebuilt. The NOCOPY algorithm was introduced as a way to avoid this.

If an ALTER TABLE operation supports the NOCOPY algorithm, then it can be performed without rebuilding the clustered index.

If the NOCOPY algorithm is specified with the ALGORITHM clause or with the alter_algorithm system variable and if the ALTER TABLE operation does not support the NOCOPY algorithm, then an error will be raised. For example:

SET SESSION alter_algorithm='NOCOPY';

ALTER TABLE tab MODIFY COLUMN c int;
ERROR 1846 (0A000): ALGORITHM=NOCOPY is not supported. Reason: Cannot change column type INPLACE. Try ALGORITHM=COPY

In this case, raising an error is preferable, if the alternative is for the operation to rebuild the clustered index, and perform unexpectedly slowly.

Operations Supported by InnoDB with the NOCOPY Algorithm

With respect to the allowed operations, the NOCOPY algorithm supports a subset of the operations supported by the INPLACE algorithm, and it supports a superset of the operations supported by the INSTANT algorithm.

See InnoDB Online DDL Operations with ALGORITHM=NOCOPY for more information.

INSTANT Algorithm

The INPLACE algorithm can sometimes be surprisingly slow in instances where it has to modify data files. The INSTANT algorithm was introduced as a way to avoid this.

If an ALTER TABLE operation supports the INSTANT algorithm, then it can be performed without modifying any data files.

If the INSTANT algorithm is specified with the ALGORITHM clause or with the alter_algorithm system variable and if the ALTER TABLE operation does not support the INSTANT algorithm, then an error will be raised. For example:

SET SESSION alter_algorithm='INSTANT';

ALTER TABLE tab MODIFY COLUMN c int;
ERROR 1846 (0A000): ALGORITHM=INSTANT is not supported. Reason: Cannot change column type INPLACE. Try ALGORITHM=COPY

In this case, raising an error is preferable, if the alternative is for the operation to modify data files, and perform unexpectedly slowly.

Operations Supported by InnoDB with the INSTANT Algorithm

With respect to the allowed operations, the INSTANT algorithm supports a subset of the operations supported by the NOCOPY algorithm.

See InnoDB Online DDL Operations with ALGORITHM=INSTANT for more information.

LOCK

In MariaDB 10.0 and later, the ALTER TABLE statement supports the LOCK clause. This clause is one of the clauses that is used to implement online DDL. ALTER TABLE supports several different locking strategies. A locking strategy can be explicitly chosen for an ALTER TABLE operation by setting the LOCK clause. The supported values are:

• DEFAULT: Acquire the least restrictive lock on the table that is supported for the specific operation. Permit the maximum amount of concurrency that is supported for the specific operation.
• NONE: Acquire no lock on the table. Permit all concurrent DML. If this locking strategy is not permitted for an operation, then an error is raised.
• SHARED: Acquire a read lock on the table. Permit read-only concurrent DML. If this locking strategy is not permitted for an operation, then an error is raised.
• EXCLUSIVE: Acquire a write lock on the table. Do not permit concurrent DML.

If the LOCK clause is not explicitly set, then the operation uses LOCK=DEFAULT.

ALTER ONLINE TABLE is equivalent to LOCK=NONE. Therefore, the ALTER ONLINE TABLE statement can be used to ensure that your ALTER TABLE operation allows all concurrent DML.

To see which locking strategies InnoDB supports for each operation, see the pages that describe which operations are supported for each algorithm:

• InnoDB Online DDL Operations with ALGORITHM=INPLACE
• InnoDB Online DDL Operations with ALGORITHM=NOCOPY
• InnoDB Online DDL Operations with ALGORITHM=INSTANT