Vitess supports both managed, online schema migrations (aka Online DDL) as well as unmanaged migrations. How Vitess runs a schema migration depends on the DDL strategy. Vitess allows these strategies:

• vitess (formerly known as online): utilizes Vitess's built-in VReplication mechanism. This is the preferred strategy in Vitess.
• gh-ost: uses 3rd party GitHub's gh-ost tool.
• pt-osc: uses 3rd party Percona's pt-online-schema-change as part of Percona Toolkit. pt-osc strategy is experimental.
• mysql: managed by the Online DDL scheduler, but executed via normal MySQL statement. Whether it is blocking or not is up to the specific query.
• direct: unmanaged. The direct apply of DDL to your database. Whether it is blocking or not is up to the specific query.

CREATE and DROP are managed in the same way, by Vitess, whether strategy is vitess, gh-ost or pt-osc.

See also ddl_strategy flags.

Specifying a DDL strategy #

You can apply DDL strategies to your schema changes in these different ways:

• The command vtctldclient ApplySchema takes a --ddl-strategy flag. The strategy applies to the specific changes requested in the command. The following example applies the vitess strategy to three migrations submitted together:

$ vtctldclient ApplySchema --ddl-strategy "vitess" --sql "ALTER TABLE demo MODIFY id bigint UNSIGNED; CREATE TABLE sample (id int PRIMARY KEY); DROP TABLE another;" commerce
ab185fdf_6e46_11ee_8f23_0a43f95f28a3

• Set vtgate --ddl_strategy flag. Migrations executed from within vtgate will use said strategy.

$ vtgate --ddl_strategy="vitess"

$ mysql

mysql> alter table corder force;
+--------------------------------------+
| uuid                                 |
+--------------------------------------+
| 2015f08d_6e46_11ee_a918_0a43f95f28a3 |
+--------------------------------------+

• Set the @@ddl_strategy session variable to override the value of vtgate's --ddl_strategy flag, for the current session.

mysql> set @@ddl_strategy="vitess --postpone-completion --allow-concurrent";
Query OK, 0 rows affected (0.01 sec)

mysql> alter table corder force;
+--------------------------------------+
| uuid                                 |
+--------------------------------------+
| 861f7de9_6e46_11ee_a918_0a43f95f28a3 |
+--------------------------------------+

Choosing a DDL strategy #

Different strategies have different behavior for ALTER statements. Sections below first break down specific handling and notes for each strategy, followed by an evaluation of the differences.

vitess #

The vitess strategy invokes Vitess's built in VReplication mechanism. It is the mechanism behind resharding, materialized views, imports from external databases, and more. VReplication migrations use the same logic for copying data as do other VReplication operations, and as such the vitess strategy is known to be compatible with overall Vitess behavior. VReplication is authored by the maintainers of Vitess.

VReplication migrations enjoy the general features of VReplication:

• Seamless integration with Vitess.
• Seamless use of the throttler mechanism.
• Visibility into internal working and status of VReplication.
• Agnostic to planned reparenting and to unplanned failovers. A migration will resume from point of interruption shortly after a new primary is available.

VReplication migrations further:

• Are revertible: you may switch back to the pre-migration schema without losing any data accumulated during and post migration.
• Support a wider range of schema changes. For example, while gh-ost has a strict requirement for a shared unique key pre/post migration, vitess migrations may work with different keys, making it possible to modify a table's PRIMARY KEY without having to rely on an additional UNIQUE KEY.

gh-ost #

gh-ost was developed by GitHub as a lightweight and safe schema migration tool.

To be able to run online schema migrations via gh-ost:

• If you're on Linux/amd64 architecture, and on glibc 2.3 or similar, there are no further dependencies. Vitess comes with a built-in gh-ost binary, that is compatible with your system. Note that the Vitess Docker images use this architecture, and gh-ost comes pre-bundled and compatible.
• On other architectures:
  • Have gh-ost executable installed
  • Run vttablet with --gh-ost-path=/full/path/to/gh-ost flag

Vitess automatically creates a MySQL account for the migration, with a randomly generated password. The account is destroyed at the end of the migration.

Vitess takes care of setting up the necessary command line flags. It automatically creates a hooks directory and populates it with hooks that report gh-ost's progress back to Vitess. You may supply additional flags for your migration as part of @@ddl_strategy session variable (using VTGate) or --ddl-strategy command line flag (using vtctldclient). Examples:

• set @@ddl_strategy='gh-ost --max-load Threads_running=200';
• set @@ddl_strategy='gh-ost --max-load Threads_running=200 --critical-load Threads_running=500 --critical-load-hibernate-seconds=60 --default-retries=512';
• vtctldclient --ddl-strategy "gh-ost --allow-nullable-unique-key --chunk-size 200" ...

Note: Do not override the following flags: alter, database, table, execute, max-lag, force-table-names, serve-socket-file, hooks-path, hooks-hint-token, panic-flag-file. Overriding any of these may cause Vitess to lose control and track of the migration, or even to migrate the wrong table.

gh-ost throttling is done via Vitess's own tablet throttler, based on replication lag.

Using pt-online-schema-change #

pt-online-schema-change is part of Percona Toolkit, a set of Perl scripts. To be able to use pt-online-schema-change, you must have the following setup on all your tablet servers (normally tablets are co-located with MySQL on same host and so this implies setting up on all MySQL servers):

• pt-online-schema-change tool installed and is executable
• Perl libdbi and libdbd-mysql modules installed. e.g. on Debian/Ubuntu, sudo apt-get install libdbi-perl libdbd-mysql-perl
• Run vttablet with -pt-osc-path=/full/path/to/pt-online-schema-change flag.

Note that on Vitess Docker images, pt-online-schema-change and dependencies are pre-installed.

Vitess automatically creates a MySQL account for the migration, with a randomly generated password. The account is destroyed at the end of the migration.

Vitess takes care of supplying the command line flags, the DSN, the username & password. It also sets up PLUGINS used to communicate migration progress back to the tablet. You may supply additional flags for your migration as part of @@ddl_strategy session variable (using VTGate) or -ddl-strategy command line flag (using vtctldclient). Examples:

• set @@ddl_strategy='pt-osc --null-to-not-null';
• set @@ddl_strategy='pt-osc --max-load Threads_running=200';
• vtctldclient ApplySchema --ddl-strategy "pt-osc --alter-foreign-keys-method auto --chunk-size 200" ...

Vitess tracks the state of the pt-osc migration. If it fails, Vitess makes sure to drop the migration triggers. Vitess keeps track of the migration even if the tablet itself restarts for any reason. Normally that would terminate the migration; Vitess will cleanup the triggers if so, or will happily let the migration run to completion if not.

Do not override the following flags: alter, pid, plugin, dry-run, execute, new-table-name, [no-]drop-new-table, [no-]drop-old-table.

pt-osc throttling is done via Vitess's own tablet throttler, based on replication lag, and via a pt-online-schema-change plugin.

Comparing the options #

There are pros and cons to using any of the strategies. Some notable differences:

General #

• All three options mimic an ALTER TABLE statement by creating and populating a shadow/ghost table behind the scenes, slowly bringing it up to date, and finally switching between the original and shadow tables.
• All three options utilize the Vitess throttler.

Support #

• VReplication (vitess strategy) is internal to Vitess and supported by the Vitess maintainers.
• gh-ost and pt-online-schema-change are not supported by the Vitess maintainers.

Setup #

• VReplication is part of Vitess
• A gh-ost binary is embedded within the Vitess binary, compatible with glibc 2.3 and Linux/amd64. The user may choose to use their own gh-ost binary, configured with --gh-ost-path.
  The embedded gh-ost binary will be removed in future versions. The user will need to install their own gh-ost binary.
• pt-online-schema-change is not included in Vitess, and the user needs to set it up on tablet hosts.
  • Note that on Vitess Docker images, pt-online-schema-change and dependencies are pre-installed.

Load #

• pt-online-schema-change uses triggers to propagate changes. This method is traditionally known to generate high load on the server. Both VReplication and gh-ost tail the binary logs to capture changes, and this approach is known to be more lightweight.
• When throttled, pt-online-schema-change still runs trigger actions, whereas both VReplication and gh-ost cease transfer of data (they may keep minimal bookkeeping operations).

Cut-over #

• All strategies use an atomic cut-over based on MySQL locking. At the end of the migration, the tables are switched, and incoming queries are momentarily blocked, but not lost.
• In addition, vitess offers a buffering layer, that reduces the contention on the database server at cut-over time.

MySQL compatibility #

• pt-online-schema-change partially supports foreign keys. Neither gh-ost nor VReplication support foreign keys.

Vitess functionality comparison #

• Managed: whether Vitess schedules and operates the migration
• Online:
  • MySQL supports limited online (INPLACE) DDL as well as INSTANT DDL. See support chart. INSTANT DDL is instant on both primary and replicas. INPLACE is non-blocking on parent but serialized on replicas, causing replication lag. Otherwise migrations are blocking on both primary and replicas.
  • gh-ost, vitess do not support foreign keys
  • pt-osc has support for foreign keys (may apply collateral blocking operations)
• Trackable: able to determine migration state (ready, running, complete etc)
  • vitess and gh-ost strategies also makes available progress % and ETA seconds
• Declarative: support --declarative flag
• Revertible: vitess strategy supports revertible ALTER statements (or ALTERs implied by --declarative migrations). All managed strategies supports revertible CREATE and DROP.
• Recoverable: a vitess migration interrupted by planned/unplanned failover, automatically resumes work from point of interruption. gh-ost and pt-osc will not resume after failover, but Vitess will automatically retry the migration (by marking the migration as failed and by initiating a RETRY), exactly once for any migration.