Most DDL operations (e.g. adding a column, removing a column or adding a default value to a column) take an ACCESS EXCLUSIVE lock on the table they are altering. While these operations wait to acquire their lock, other statements are blocked. Before running a migration, Safe PG Migrations sets a short lock timeout (default to 5 seconds) so that statements are not blocked for too long.

See PostgreSQL Alter Table and Long Transactions and Migrations and Long Transactions for detailed explanations of the matter.

Adding a foreign key or a not-null constraint can take a lot of time on a large table. The problem is that those operations take ACCESS EXCLUSIVE locks. We clearly don't want them to hold these locks for too long. Thus, Safe PG Migrations runs them with a short statement timeout (default to 5 seconds).

See Zero-downtime Postgres migrations - the hard parts for a detailed explanation on the subject.

When Safe PG Migrations is used, migrations are not wrapped in a transaction. This is for several reasons:

• We want to release locks as soon as possible.
• In order to be able to retry statements that have failed because of a lock timeout, we have to be outside a transaction.
• In order to add an index concurrently, we have to be outside a transaction.

Note that if a migration fails, it won't be rolled back. This can result in migrations being partially applied. In that case, they need to be manually reverted.

Safe PG Migrations gracefully handle the upgrade to PG11 by not backfilling default value for existing rows, as the database engine is now natively handling it.

Beware though, when adding a volatile default value:

add_column :users, :created_at, default: 'clock_timestamp()'

PG will still needs to update every row of the table, and will most likely statement timeout for big table. In this case, Safe PG Migrations can automatically backfill data when the option default_value_backfill: is set to :update_in_batches.

Safe PG Migrations provides the extra option parameter default_value_backfill:. When your migration is adding a volatile default value, the option :update_in_batches can be set. It will automatically backfill the value in a safe manner.

safety_assured do
  add_column :users, :created_at, default: 'clock_timestamp()', default_value_backfill: :update_in_batches
end

More specifically, it will:

1. create the column without default value and without null constraint. This ensure the ACCESS EXCLUSIVE lock is acquired for the least amount of time;
2. add the default value, without data backfill. An ACCESS EXCLUSIVE lock is acquired and released immediately;
3. backfill data, in batch of SafePgMigrations.config.backfill_batch_size and with a pause of SafePgMigrations.config.backfill_pause between each batch;
4. change the column to null: false, if defined in the parameters, following the algorithm we have defined below.

NOTE

Data backfill take time. If your table is big, your migrations will (safely) hangs for a while. You might want to backfill data manually instead, to do so you will need two migrations

1. First migration :

   a. adds the column without default and without null constraint;

   b. add the default value.

2. manual data backfill (rake task, manual operation, ...)

3. Second migration which change the column to null false (with Safe PG Migrations, change_column_null is safe and can be used; see section below)

default_value_backfill: also accept the value :auto which is set by default. In this case, Safe PG Migrations will not backfill data and will let PostgreSQL handle it itself.

add_column with default_value_backfill: :update_in_batches can be dangerous on big tables. To avoid unwanted long migrations, Safe PG Migrations does not automatically mark this usage as safe when used with strong-migrations, usage of safety_assured is required.

It is also possible to set a threshold for the table size, above which the migration will fail. This can be done by setting the default_value_backfill_threshold: option in the configuration.

Creating an index requires a SHARE lock on the target table which blocks all write on the table while the index is created (which can take some time on a large table). This is usually not practical in a live environment. Thus, Safe PG Migrations ensures indexes are created concurrently.

As CREATE INDEX CONCURRENTLY and DROP INDEX CONCURRENTLY are non-blocking operations (ie: read/write operations on the table are still possible), Safe PG Migrations sets a lock timeout to 30 seconds for those 2 specific statements.

If you still get lock timeout while adding / removing indexes, it might be for one of those reasons:

• Long-running queries are active on the table. To create / remove an index, PG needs to wait for the queries that are actually running to finish before starting the index creation / removal. The blocking activity logger might help you to pinpoint the culprit queries.
• A vacuum / autovacuum is running on the table, holding a ShareUpdateExclusiveLock, you are most likely out of luck for the current migration, but you may try to optimize your autovacuums settings.

Adding a foreign key requires a SHARE ROW EXCLUSIVE lock, which prevent writing in the tables while the migration is running.

Adding the constraint itself is rather fast, the major part of the time is spent on validating this constraint. Thus Safe PG Migrations ensures that adding a foreign key holds blocking locks for the least amount of time by splitting the foreign key creation in two steps:

1. adding the constraint without validation, will not validate existing rows;
2. validating the constraint, will validate existing rows in the table, without blocking read or write on the table

Adding a check constraint requires an ACCESS EXCLUSIVE lock, which prevent writing and reading in the tables as soon as the lock is requested.

Adding the constraint itself is rather fast, the major part of the time is spent on validating this constraint. Thus Safe PG Migrations ensures that adding a constraints holds blocking locks for the least amount of time by splitting the constraint addition in two steps:

1. adding the constraint without validation, will not validate existing rows;
2. validating the constraint, will validate existing rows in the table, without blocking read or write on the table

Changing the nullability of a column requires an ACCESS EXCLUSIVE lock, which prevent writing and reading in the tables as soon as the lock is requested.

Adding the constraint itself is rather fast, the major part of the time is spent on validating this constraint.

Safe PG Migrations acts differently depending on the version you are on.

Starting on PostgreSQL versions 12, adding the column NOT NULL constraint is safe if a check constraint validates the nullability of the same column. Safe PG Migrations also relies on add_check_constraint, which was introduced in ActiveRecord 6.1.

If these requirements are met, Safe PG Migrations ensures that adding a constraints holds blocking locks for the least amount of time by splitting the constraint addition in several steps:

1. adding a IS NOT NULL constraint without validation, will not validate existing rows but block read or write;
2. validating the constraint, will validate existing rows in the table, without blocking read or write on the table;
3. changing the not null status of the column, thanks to the NOT NULL constraint without having to scan the table sequentially;
4. dropping the IS NOT NULL constraint.

If the version of PostgreSQL is below 12, or if the version of ActiveRecord is below 6.1, Safe PG Migrations will only wrap ActiveRecord method into a statement timeout and lock timeout.

Calling change_column_null with a default parameter is dangerous and is likely not to finish in the statement timeout defined by Safe PG Migrations. For this reason, when the default parameter is given, Safe PG Migrations will simply forward it to activerecord methods without trying to improve it

Dropping a null constraint still requires an ACCESS EXCLUSIVE lock, but does not require extra operation to reduce the amount of time during which the lock is held. Safe PG Migrations only wrap methods of activerecord in lock and statement timeouts

When a statement fails with a lock timeout, Safe PG Migrations retries it (5 times max) list of retriable statements

If a statement fails with a lock timeout, Safe PG Migrations will try to tell you what was the blocking statement.

NOTE

Data logged by the Blocking activity logger can be sensitive (it will contain raw SQL queries, which can be hashes of password, user information, ...)

If you cannot afford to log this type of data, you can either

• Set SafePgMigrations.config.blocking_activity_logger_verbose = false. In this case, the logger will only log the pid of the blocking statement, which should be enough to investigate;
• Provide a different logger through SafePgMigrations.config.sensitive_logger = YourLogger.new. Instead of using the default IO stream, SafePgMigrations will send sensitive data to the given logger which can be managed as you wish.

Dropping a table can be difficult to achieve in a small amount of time if it holds several foreign keys to busy tables. To remove the table, PostgreSQL will have to acquire an access exclusive lock on all the tables referenced by the foreign keys.

To solve this issue, Safe Pg Migrations will drop the foreign keys before dropping the table.

NOTE

Dropping a table is a dangerous operation by nature. Safe Pg Migrations will not prevent the deletion of a table which would still be in use.

For any operation, Safe PG Migrations can output the performed SQL queries. This feature is enabled by default in a production Rails environment. If you want to explicit enable it, for example in your development environment you can use:

export SAFE_PG_MIGRATIONS_VERBOSE=1

Instead of the traditional output:

add_index :users, :age

== 20191215132355 SampleIndex: migrating ======================================
-- add_index(:users, :age)
   -> add_index("users", :age, {:algorithm=>:concurrently})
   -> 0.0175s
== 20191215132355 SampleIndex: migrated (0.0200s) =============================

Safe PG Migrations will output the following logs:

add_index :users, :age

== 20191215132355 SampleIndex: migrating ======================================
   (0.3ms)  SHOW lock_timeout
   (0.3ms)  SET lock_timeout TO '5s'
-- add_index(:users, :age)
   -> add_index("users", :age, {:algorithm=>:concurrently})
   (0.3ms)  SHOW statement_timeout
   (0.3ms)  SET statement_timeout TO 0
   (0.3ms)  SHOW lock_timeout
   (0.3ms)  SET lock_timeout TO '30s'
   (3.5ms)  CREATE INDEX CONCURRENTLY "index_users_on_age" ON "users"  ("age")
   (0.3ms)  SET lock_timeout TO '5s'
   (0.2ms)  SET statement_timeout TO '1min'
   -> 0.0093s
   (0.2ms)  SET lock_timeout TO '0'
== 20191215132355 SampleIndex: migrated (0.0114s) =============================

So you can actually check that the CREATE INDEX statement will be performed concurrently, without any statement timeout and with a lock timeout of 30 seconds.

Nb: The SHOW statements are used by Safe PG Migrations to query settings for their original values in order to restore them after the work is done