A non-exhaustive guide on common migration recipes and how to avoid trouble.

• Adding an index
• Adding a reference or foreign key
• Adding a column with a default value
• Changing the type of a column
• Removing a column
• Renaming a column
• Renaming a table
• Adding a check constraint
• Setting NOT NULL on an existing column
• Adding a JSON column
• Squashing migrations

Read more about safe migrations at Fly.io Phoenix Files where we dive into how to safely backfill data and go through Ecto Migration options.

Creating an index will block both reads and writes.

def change do
  create index("posts", [:slug])

  # This obtains a ShareLock on "posts" which will block writes to the table
end

Instead, have Postgres create the index concurrently which does not block reads. You will need to disable the migration transactions to use CONCURRENTLY.

@disable_ddl_transaction true
@disable_migration_lock true

def change do
  create index("posts", [:slug], concurrently: true)
end

The migration may still take a while to run, but reads and updates to rows will continue to work. For example, for 100,000,000 rows it took 165 seconds to add run the migration, but SELECTS and UPDATES could occur while it was running.

Do not have other changes in the same migration; only create the index concurrently and separate other changes to later migrations.

Adding a foreign key blocks writes on both tables.

def change do
  alter table("posts") do
    add :group_id, references("groups")
  end
end

In the first migration

def change do
  alter table("posts") do
    add :group_id, references("groups", validate: false)
  end
end

In the second migration

def change do
  execute "ALTER TABLE posts VALIDATE CONSTRAINT group_id_fkey", ""
end

These migrations can be in the same deployment, but make sure they are separate migrations.

Adding a column with a default value to an existing table may cause the table to be rewritten. During this time, reads and writes are blocked in Postgres, and writes are blocked in MySQL and MariaDB.

Note: This becomes safe in:

• Postgres 11+
• MySQL 8.0.12+
• MariaDB 10.3.2+

def change do
  alter table("comments") do
    add :approved, :boolean, default: false
    # This took 10 minutes for 100 million rows with no fkeys,

    # Obtained an AccessExclusiveLock on the table, which blocks reads and
    # writes.
  end
end

Add the column first, then alter it to include the default.

First migration:

def change do
  alter table("comments") do
    add :approved, :boolean
    # This took 0.27 milliseconds for 100 million rows with no fkeys,
  end
end

Second migration:

def change do
  alter table("comments") do
    modify :approved, :boolean, default: false
    # This took 0.28 milliseconds for 100 million rows with no fkeys,
  end
end

Schema change to read the new column:

schema "comments" do
+ field :approved, :boolean, default: false
end

Changing the type of a column may cause the table to be rewritten. During this time, reads and writes are blocked in Postgres, and writes are blocked in MySQL and MariaDB.

Safe in Postgres:

• increasing length on varchar or removing the limit
• changing varchar to text
• changing text to varchar with no length limit
• Postgres 9.2+ - increasing precision (NOTE: not scale) of decimal or numeric columns. eg, increasing 8,2 to 10,2 is safe. Increasing 8,2 to 8,4 is not safe.
• Postgres 9.2+ - changing decimal or numeric to be unconstrained
• Postgres 12+ - changing timestamp to timestamptz when session TZ is UTC

Safe in MySQL/MariaDB:

• increasing length of varchar from < 255 up to 255.
• increasing length of varchar from > 255 up to max.

def change do
  alter table("posts") do
    modify :my_column, :boolean, from: :text
  end
end

Take a phased approach:

1. Create a new column
2. In application code, write to both columns
3. Backfill data from old column to new column
4. In application code, move reads from old column to the new column
5. In application code, remove old column from Ecto schemas.
6. Drop the old column.

If Ecto is still configured to read a column in any running instances of the application, then queries will fail when loading data into your structs. This can happen in multi-node deployments or if you start the application before running migrations.

# Without a code change to the Ecto Schema

def change do
  alter table("posts") do
    remove :no_longer_needed_column
  end
end

Safety can be assured if the application code is first updated to remove references to the column so it's no longer loaded or queried. Then, the column can safely be removed from the table.

1. Deploy code change to remove references to the field.
2. Deploy migration change to remove the column.

First deployment:

# First deploy, in the Ecto schema

defmodule MyApp.Post do
  schema "posts" do
-   column :no_longer_needed_column, :text
  end
end

Second deployment:

def change do
  alter table("posts") do
    remove :no_longer_needed_column
  end
end

Ask yourself: "Do I really need to rename a column?". Probably not, but if you must, read on and be aware it requires time and effort.

If Ecto is configured to read a column in any running instances of the application, then queries will fail when loading data into your structs. This can happen in multi-node deployments or if you start the application before running migrations.

There is a shortcut: Don't rename the database column, and instead rename the schema's field name and configure it to point to the database column.

# In your schema
schema "posts" do
  field :summary, :text
end


# In your migration
def change do
  rename table("posts"), :title, to: :summary
end

The time between your migration running and your application getting the new code may encounter trouble.

Strategy 1

Rename the field in the schema only, and configure it to point to the database column and keep the database column the same. Ensure all calling code relying on the old field name is also updated to reference the new field name.

defmodule MyApp.MySchema do
  use Ecto.Schema

  schema "weather" do
    field :temp_lo, :integer
    field :temp_hi, :integer
    field :precipitation, :float, source: :prcp
    field :city, :string

    timestamps(type: :naive_datetime_usec)
  end
end

## Update references in other parts of the codebase:
  my_schema = Repo.get(MySchema, "my_id")
-  my_schema.prcp
+  my_schema.precipitation

Strategy 2

Take a phased approach:

1. Create a new column
2. In application code, write to both columns
3. Backfill data from old column to new column
4. In application code, move reads from old column to the new column
5. In application code, remove old column from Ecto schemas.
6. Drop the old column.

Ask yourself: "Do I really need to rename a table?". Probably not, but if you must, read on and be aware it requires time and effort.

If Ecto is still configured to read a table in any running instances of the application, then queries will fail when loading data into your structs. This can happen in multi-node deployments or if you start the application before running migrations.

There is a shortcut: rename the schema only, and do not change the underlying database table name.

def change do
  rename table("posts"), to: table("articles")
end

Strategy 1

Rename the schema only and all calling code, and don’t rename the table:

- defmodule MyApp.Weather do
+ defmodule MyApp.Forecast do
  use Ecto.Schema

  schema "weather" do
    field :temp_lo, :integer
    field :temp_hi, :integer
    field :precipitation, :float, source: :prcp
    field :city, :string

    timestamps(type: :naive_datetime_usec)
  end
end

# and in calling code:
- weather = MyApp.Repo.get(MyApp.Weather, “my_id”)
+ forecast = MyApp.Repo.get(MyApp.Forecast, “my_id”)

Strategy 2

Take a phased approach:

1. Create the new table. This should include creating new constraints (checks and foreign keys) that mimic behavior of the old table.
2. In application code, write to both tables, continuing to read from the old table.
3. Backfill data from old table to new table
4. In application code, move reads from old table to the new table
5. In application code, remove the old table from Ecto schemas.
6. Drop the old table.

Adding a check constraint blocks reads and writes to the table in Postgres, and blocks writes in MySQL/MariaDB while every row is checked.

def change do
  create constraint("products", :price_must_be_positive, check: "price > 0")
  # Creating the constraint with validate: true (the default when unspecified)
  # will perform a full table scan and acquires a lock preventing updates
end

There are two operations occurring:

1. Creating a new constraint for new or updating records
2. Validating the new constraint for existing records

If these commands are happening at the same time, it obtains a lock on the table as it validates the entire table and fully scans the table. To avoid this full table scan, we can separate the operations.

In one migration:

def change do
  create constraint("products", :price_must_be_positive, check: "price > 0"), validate: false
  # Setting validate: false will prevent a full table scan, and therefore
  # commits immediately.
end

In the next migration:

def change do
  execute "ALTER TABLE products VALIDATE CONSTRAINT price_must_be_positive", ""
  # Acquires SHARE UPDATE EXCLUSIVE lock, which allows updates to continue
end

These can be in the same deployment, but ensure there are 2 separate migrations.

Setting NOT NULL on an existing column blocks reads and writes while every row is checked. Just like the Adding a check constraint scenario, there are two operations occurring:

1. Creating a new constraint for new or updating records
2. Validating the new constraint for existing records

To avoid the full table scan, we can separate these two operations.

def change do
  alter table("products") do
    modify :active, :boolean, null: false
  end
end

Add a check constraint without validating it, backfill data to satiate the constraint and then validate it. This will be functionally equivalent.

In the first migration:

# Deployment 1
def change do
  create constraint("products", :active_not_null, check: "active IS NOT NULL"), validate: false
end

This will enforce the constraint in all new rows, but not care about existing rows until that row is updated.

You'll likely need a data migration at this point to ensure that the constraint is satisfied.

Then, in the next deployment's migration, we'll enforce the constraint on all rows:

# Deployment 2
def change do
  execute "ALTER TABLE products VALIDATE CONSTRAINT active_not_null", ""
end

If you're using Postgres 12+, you can add the NOT NULL to the column after validating the constraint. From the Postgres 12 docs:

SET NOT NULL may only be applied to a column provided none of the records in the table contain a NULL value for the column. Ordinarily this is checked during the ALTER TABLE by scanning the entire table; however, if a valid CHECK constraint is found which proves no NULL can exist, then the table scan is skipped.

# **Postgres 12+ only**

def change do
  execute "ALTER TABLE products VALIDATE CONSTRAINT active_not_null", ""

  alter table("products") do
    modify :active, :boolean, null: false
  end

  drop constraint("products", :active_not_null)
end

If your constraint fails, then you should consider backfilling data first to cover the gaps in your desired data integrity, then revisit validating the constraint.

In Postgres, there is no equality operator for the json column type, which can cause errors for existing SELECT DISTINCT queries in your application.

def change do
  alter table("posts") do
    add :extra_data, :json
  end
end

Use jsonb instead. Some say it’s like “json” but “better.”

def change do
  alter table("posts") do
    add :extra_data, :jsonb
  end
end

If you have a long list of migrations, sometimes it can take a while to migrate each of those files every time the project is reset or spun up by a new developer. Thankfully, Ecto comes with mix tasks to dump and load a database structure which will represent the state of the database up to a certain point in time, not including content.

• mix ecto.dump
• mix ecto.load

Schema dumping and loading is only supported by external binaries pg_dump and mysqldump, which are used by the Postgres, MyXQL, and MySQL Ecto adapters (not supported in MSSQL adapter).

For example:

20210101000000 - First Migration
20210201000000 - Second Migration
20210701000000 - Third Migration <-- we are here now. run `mix ecto.dump`

We can "squash" the migrations up to the current day which will effectively fast-forward migrations to that structure. The Ecto Migrator will detect that the database is already migrated to the third migration, and so it begins there and migrates forward.

Let's add a new migration:

20210101000000 - First Migration
20210201000000 - Second Migration
20210701000000 - Third Migration <-- `structure.sql` represents up to here
20210801000000 - New Migration <-- This is where migrations will begin

The new migration will still run, but the first-through-third migrations will not need to be run since the structure already represents the changes applied by those migrations. At this point, you can safely delete the first, second, and third migration files or keep them for historical auditing.

Let's make this work:

1. Run mix ecto.dump which will dump the current structure into priv/repo/structure.sql by default. Check the mix task for more options.
2. During project setup with an empty database, run mix ecto.load to load structure.sql.
3. Run mix ecto.migrate to run any additional migrations created after the structure was dumped.

To simplify these actions into one command, we can leverage mix aliases:

# mix.exs

defp aliases do
  [
    "ecto.reset": ["ecto.drop", "ecto.setup"],
    "ecto.setup": ["ecto.load", "ecto.migrate"],
    # ...
  ]
end

Now you can run mix ecto.setup and it will load the database structure and run remaining migrations. Or, run mix ecto.reset and it will drop and run setup. Of course, you can continue running mix ecto.migrate as you create them.

Created and written by David Bernheisel with recipes heavily inspired from Andrew Kane and his library strong_migrations.

PostgreSQL at Scale by James Coleman

Strong Migrations by Andrew Kane

Adding a NOT NULL CONSTRAINT on PG Faster with Minimal Locking by Christophe Escobar

Postgres Runtime Configuration

Automatic and Manual Ecto Migrations by Wojtek Mach

Special thanks for sponsorship:

• Fly.io

Special thanks for these reviewers:

• Steve Bussey
• Stephane Robino
• Dennis Beatty
• Wojtek Mach
• Mark Ericksen
• And all the contributors

Postgres Lock Conflicts

• SELECT acquires a ACCESS SHARE lock
• SELECT FOR UPDATE acquires a ROW SHARE lock
• UPDATE, DELETE, and INSERT will acquire a ROW EXCLUSIVE lock
• CREATE INDEX CONCURRENTLY and VALIDATE CONSTRAINT acquires SHARE UPDATE EXCLUSIVE
• CREATE INDEX acquires SHARE lock

Knowing this, let's re-think the above table: