Database Migrations
Django migrations are how we handle changes to the database in Sentry.
Django migration official docs: https://docs.djangoproject.com/en/5.1/topics/migrations/ . These will cover most things you need to understand what a migration is doing.
Note that for all of these commands you can substitute getsentry
for sentry
if in the getsentry
repo.
sentry upgrade
will automatically bring your migrations up to date. You can also run sentry django migrate
to access the migration command directly.
This can be helpful for when you want to test a migration.
sentry django migrate <app_name> <migration_name>
- Note that migration_name
can be a partial match, often the number is all you need.
eg: sentry django migrate sentry 0005
This can be used to roll a migration back as well. Useful in dev if you make a mistake.
A GitHub action will automatically comment on your PR with the SQL for your migration, and the comment will stay updated with any future changes. You can also manually generate SQL with this command.
sentry django sqlmigrate <app_name> <migration_name>
eg sentry django sqlmigrate sentry 0003
This generates migrations for you automatically based on changes you've made to models.
sentry django makemigrations
or
sentry django makemigrations <app_name>
for a specific app.
eg sentry django makemigrations sentry
You can also generate an empty migration with sentry django makemigrations <app_name> --empty
. This is useful for data migrations and other custom work.
Note that if you have added a new model, you also need to import the model in __init__.py
, or the model will not be recognized in testing.
When merging to master you might notice a conflict with migrations_lockfile.txt
. This file is in place to help us avoid merging two migrations with the same migration number to master, and if you're conflicting with it then it's likely someone has committed a migration ahead of you.
To resolve this, rebase against latest master, delete your current migration and then regenerate it. If your migration was custom, just save the operations in a text file somewhere so that you can reapply them on the regenerated migration.
Always commit the changes to migrations_lockfile.txt
with your migration.
There are some things we need to be careful about when running migrations.
Database migrations are risky operations that can lead to irreversible data loss or corruption. This is especially true for data migrations. For this reason, every migration should have a corresponding integration test.
To test your migration, derive a test case from TestMigrations
and add it to tests/sentry/migrations
. For example:
class MyMigrationTest(TestMigrations):
migrate_from = "0123_previous_migration"
migrate_to = "0124_my_new_migration"
def setup_before_migration(self, apps):
# Create your db state here
Project = apps.get_model("sentry", "Project")
self.project = Project.objects.create(organization_id=self.organization.id, name="my_project")
def test(self):
# Test state after migration
self.project.refresh_from_db()
assert self.project.name == "MyProject"
To run the test locally, run pytest
with --migrations
flag. For example, pytest -v --migrations tests/getsentry/migrations/test_0XXX_migration_name.py
.
If you would like to speed up the migration tests and do not require rebuilding the databases on each test run, supply --reuse-db
as an additional option to the test command.
When you add or change a model, an error message in CI may appear explaining that one or multiple tests "produced an export.json
backup file that was missing the above models". In order to resolve this, there are two steps:
- Add the new or modified model to the exhaustive organization in testutils/helpers/backups.py by creating an instance of your model, for example by invoking MyModel.objects.create(). This ensures the presence of the new model when creating the snapshot and during testing.
- The snapshot files can be regenerated using the following command:
SENTRY_SNAPSHOTS_WRITEBACK=1 pytest tests/sentry/backup/test_sanitize.py
There are also tests for model dependencies that make use of automatically generated fixtures in tests/sentry/backup/test_dependencies.py. These tests will fail if they are not updated when a new model with dependencies on other models is added, or dependencies are modified. In order to re-generate the model dependency graphs, you can run bin/generate-model-dependency-fixtures.
- There is a known issue with the
django-pg-zero-downtime-migrations
package which causes the roll back of aNOT NULL
constraint to fail. If this happens with an old migration test, it's ok to delete the test rather than trying to fix the issue. - If you want to use existing
create_*
helper functions to create model instances, overridesetup_initial_state
rather thansetup_before_migration
. This function will run before the database is rolled back tomigration_from
.
If a (data) migration involves large tables, or columns that aren't indexed it is better to iterate over the entire table instead of using a filter. For example:
EnvironmentProject.objects.filter(environment__name="none")
Because there are too many EnvironmentProject
rows, this will bring too many rows into memory at once. Instead we should iterate over all the EnvironmentProject
rows using RangeQuerySetWrapperWithProgressBar
since it will do it in chunks. For example:
for env in RangeQuerySetWrapperWithProgressBar(EnvironmentProject.objects.all()):
if env.name == 'none':
# Do what you need
We generally prefer to avoid using .filter
with RangeQuerySetWrapperWithProgressBar
. Since it already orders by the id to iterate through the table, we can't take advantage of any indexes on the fields, and could potentially scan a large number of rows for each chunk. This will run slower, but we generally prefer that, since it averages the load out over a longer period of time, and makes each query to fetch each chunk fairly cheap.
We prefer to create indexes on large existing tables with CREATE INDEX CONCURRENTLY
. Our migration framework will do this automatically when creating a new index. Note that when CONCURRENTLY
is used we can't run the migration in a transaction, so it's important to use atomic = False
to run these.
When adding indexes to large tables you should use a is_post_deployment
migration as creating the index could take longer than the migration statement timeout of 5s.
This is complicated due to our deploy process. When we deploy, we run migrations, and then push out the application code, which takes a while. This means that if we just delete a column or model, then code in sentry will be looking for those columns/tables and erroring until the deploy completes. In some cases, this can mean Sentry is hard down until the deploy is finished.
To avoid this, follow these steps:
- (Optional, but ideal) Make a PR to remove all uses of the column in the codebase in a separate PR. This mostly helps with code cleanliness. This should be merged ahead of the migration prs, but we don't need to worry about whether it is deployed first.
- Make another PR that:
- Checks if the column is either not nullable, or doesn't have a db_default set. If either of these is true, then make it nullable via
null=True
. - If the column is a foreign key, remove the database level foreign key constraint it by setting
db_constraint=False
. - Remove the column and in the generated migration use
SafeRemoveField(..., deletion_action=DeletionAction.MOVE_TO_PENDING)
to replaceRemoveField(...)
. This only marks the state for the column as removed. - Combine these migrations together to save making multiple deploys
- Checks if the column is either not nullable, or doesn't have a db_default set. If either of these is true, then make it nullable via
- Deploy your migration changes. It's important that all previous pull requests are in production before we remove the actual column from the table.
- Make a pull request that create a new migration that has the same
SafeRemoveField
operation as before, but setdeletion_action=DeletionAction.DELETE
instead. This deletes the actual column from the table in Postgres. - Deploy the drop column migration.
Here's an example of removing the project
column from this model. It is both a foreign key and not null:
@region_silo_model
class TestModel(Model):
__relocation_scope__ = RelocationScope.Excluded
project = FlexibleForeignKey("sentry.Project")
name = models.TextField()
class Meta:
app_label = "uptime"
db_table = "uptime_testmodel"
First we remove the constraint and make the column nullable:
# Model change
...
project = FlexibleForeignKey("sentry.Project", db_constraint=False, null=True)
...
# Migration operations
operations = [
migrations.AlterField(
model_name="testmodel",
name="project",
field=sentry.db.models.fields.foreignkey.FlexibleForeignKey(
db_constraint=False,
null=True,
on_delete=django.db.models.deletion.CASCADE,
to="sentry.project",
),
),
]
Once we've done this, we can now remove the column from the model and generate the migration to remove it. The generated migration looks like this:
operations = [
migrations.RemoveField(model_name="testmodel", name="project"),
]
Deleting this way is unsafe, so we want to replace this with SafeDeleteModel
instead. So this becomes
operations = [
SafeRemoveField(model_name="testmodel", name="project", deletion_action=DeletionAction.MOVE_TO_PENDING),
]
Using SafeRemoveField
allows us to remove all the state related to the column, but defer deleting it until a later migration. So now, we can combine the migration to remove the constraints and delete the column state together like so
operations = [
migrations.AlterField(
model_name="testmodel",
name="project",
field=sentry.db.models.fields.foreignkey.FlexibleForeignKey(
db_constraint=False,
null=True,
on_delete=django.db.models.deletion.CASCADE,
to="sentry.project",
),
),
SafeRemoveField(model_name="testmodel", name="project", deletion_action=DeletionAction.MOVE_TO_PENDING),
]
So now we deploy this migration and move onto the final step.
In this last step, we will use SafeRemoveField
again to finally remove the column from the table in Postgres. First, we generate an empty migration using sentry django makemigrations <your_app> --empty
to produce an empty migration, and then modify the operations to be like:
operations = [
SafeRemoveField(model_name="testmodel", name="project", deletion_action=DeletionAction.DELETE),
]
Then we deploy this and we're done. So to recap the steps here:
- Remove all references to the column in the code in a separate pull request and merge. Doesn't matter if this deploys before the next step or not.
- If the column has an foreign key constraint them remove it. If it's not null and has no db_default then mark it as nullable. Then delete the column using
SafeRemoveField(..., deletion_action=DeletionAction.MOVE_TO_PENDING)
. These operations can be in the same migration to save time. - Deploy all previous before continuing.
- Remove the column from the table in from Postgres using
SafeRemoveField(..., deletion_action=DeletionAction.DELETE),
Extra care is needed here if the table is referenced as a foreign key in other tables. In that case, first remove the foreign key columns in the other tables, then come back to this step.
- (Optional, but ideal) Make a pull request to remove all uses of the model in the codebase in a separate pull request. This mostly helps with code cleanliness. This should be merged ahead of the migration pull requests, but we don't need to worry about whether it is deployed first.
- Make another pull request to:
- Remove any database level foreign key constraints from this table to other tables by setting
db_constraint=False
on the columns. - Remove the model and in the generated migration use
SafeDeleteModel(..., deletion_action=DeletionAction.MOVE_TO_PENDING)
to replaceDeleteModel(...)
. This only marks the state for the model as removed.
- Remove any database level foreign key constraints from this table to other tables by setting
- Deploy. It's important that all previous pull requests are in production before we remove the actual table.
- Make a pull request that create a new migration that has the same
SafeDeleteModel
operation as before, but setdeletion_action=DeletionAction.DELETE
instead. This deletes the actual table from Postgres. - Deploy
Here's an example of removing this model:
@region_silo_model
class TestModel(Model):
__relocation_scope__ = RelocationScope.Excluded
project = FlexibleForeignKey("sentry.Project")
name = models.TextField()
class Meta:
app_label = "uptime"
db_table = "uptime_testmodel"
First, we remove all references to this model from the codebase, including making sure that it's not referenced by any other models. This is best done as a separate pr to keep things clean.
Next we need to remove any db level foreign key constraints. To do this, we change this column and generate a migration:
project = FlexibleForeignKey("sentry.Project", db_constraint=False)
The operations in the migration look like
operations = [
migrations.AlterField(
model_name="testmodel",
name="project",
field=sentry.db.models.fields.foreignkey.FlexibleForeignKey(
db_constraint=False,
on_delete=django.db.models.deletion.CASCADE,
to="sentry.project",
),
),
]
Once we've done this, we can now remove the model from code and generate the migration to remove it. The generated migration looks like this:
operations = [
migrations.DeleteModel(name="TestModel"),
]
Deleting this way is unsafe, so we want to replace this with SafeDeleteModel
instead. So this becomes
operations = [
SafeDeleteModel(name="TestModel", deletion_action=DeletionAction.MOVE_TO_PENDING),
]
Using SafeDeleteModel
allows us to remove all the state related to the model, but defer deleting it until a later migration. So now, we can combine the migration to remove the constraints and delete the model state together like so
operations = [
migrations.AlterField(
model_name="testmodel",
name="project",
field=sentry.db.models.fields.foreignkey.FlexibleForeignKey(
db_constraint=False,
on_delete=django.db.models.deletion.CASCADE,
to="sentry.project",
),
),
SafeDeleteModel(name="TestModel", deletion_action=DeletionAction.MOVE_TO_PENDING),
]
So now we deploy this migration and move onto the final step.
In this last step, we will use SafeDeleteModel
again to finally remove the table from Postgres. First, we generate an empty migration using sentry django makemigrations <your_app> --empty
to produce an empty migration, and then modify the operations to be like:
operations = [
SafeDeleteModel(name="TestModel", deletion_action=DeletionAction.DELETE),
]
Then we deploy this and we're done. So to recap the steps here:
- Remove all references to the model in the code in a separate pull request and merge. Doesn't matter if this deploys before the next step or not.
- Remove any foreign key constraints and delete the model using
SafeDeleteModel(..., deletion_action=DeletionAction.MOVE_TO_PENDING)
. These operations can be in the same migration to save time. - Deploy all previous before continuing.
- Remove the table from Postgres using
SafeDeleteModel(..., deletion_action=DeletionAction.DELETE),
Creating foreign keys is mostly fine, but for some large/busy tables like Project
, Group
it can cause problems due to difficulties in acquiring a lock. You can still create a Django level foreign key though, without creating a database constraint. To do so, set db_constraint=False
when defining the key.
Renaming tables is dangerous and will result in downtime. The reason this occurs is that during the deploy a mix of old/new code will be running. So once we rename the table in Postgres, the old code will immediately start erroring if it attempts to access it. There are two ways to handle renaming a table:
- Don't rename the table in Postgres. Instead, just rename the model in Django, and make sure
Meta.db_table
is set to the current tablename so that nothing breaks. This is the preferred method. - If you absolutely want to rename the table, then the steps would be:
- Create a table with the new name
- Start dual-writing to both the old and new table, ideally in a transaction.
- Backfill the old rows into the new table.
- Change the model to start reading from the new table.
- Stop writing to the old table and remove references from the code.
- Drop the old table.
- Generally, this is not worth doing and a lot of risk/effort compared to the reward.
With postgres 14, columns can be added to tables of all sizes as deploy time migrations if you follow the guidelines on default values & allowing nulls. When creating new columns they should either be:
- Not null with a default. https://develop.sentry.dev/database-migrations/#adding-not-null-to-columns
- Created as nullable. If no default value can be set on the column, then it's best just to make it nullable.
Since we run Postgres >= 14 in production we are able to add columns with a default. To do so, instead of using default=<your_default>
, use db_default=<your_default>
. This tells Django to set a default at the database level and manage it there, rather than managing it in application code.
We can't use default
because Django's default behaviour for creating a new not null column with a default is dangerous. When using default, in the migration Django will add the default to backfill all fields, then immediately remove it so that it can handle them in the app layer. This means that during a deploy, the column is sitting in production without a default until all code rolls out, which means that inserts will fail for this table until the deploy completes.
It can be dangerous to add not null to columns, even if there is data in every row of the table for that column. This is because Postgres still needs to perform a not null check on all rows before it can add the constraint. On small tables this can be fine since the check will be quick, but on larger tables this can cause downtime. There are a few options here to make this safe:
ALTER TABLE tbl ADD CONSTRAINT cnstr CHECK (col IS NOT NULL) NOT VALID;
ALTER TABLE tbl VALIDATE CONSTRAINT cnstr;
One approach is to create the constraint as not valid. Then we validate it afterwards. We still need to scan the whole table to validate, but we only need to hold a SHARE UPDATE EXCLUSIVE
lock, which only blocks other ALTER TABLE
commands, but will allow reads/writes to continue. This works well, but has a slight performance penalty of 0.5-1%. After Postgres 12 we can extend this method to add a real NOT NULL
constraint.
Alternatively, if the table is small enough and has low enough volume it should be safe to just create a normal NOT NULL
constraint. Small being a few million rows or less.
Altering the type of a column is usually dangerous, since it will require a whole table rewrite. There are some exceptions:
- Altering a
varchar(<size>)
to avarchar
with a larger size. - Altering any
varchar
totext
- Altering a
numeric
to anumeric
where theprecision
is higher but thescale
is the same.
For any other types, the best path forward is usually:
- Create a column with the new type
- Start dual-writing to both the old and new column.
- Backfill and convert the old column values into the new column.
- Change the code to use the new field.
- Stop writing to the old column and remove references from the code.
- Drop the old column from the database.
Generally this can be worth a discussion in #discuss-backend.
Renaming columns is dangerous and will result in downtime. The reason this occurs is that during the deploy a mix of old/new code will be running. So once we rename the column in Postgres, the old code will immediately start erroring if it attempts to access it. There are two ways to handle renaming a column:
- Don't rename the column in Postgres. Instead, just rename the field in Django, and use
db_column
in the definition to set it to the existing column name so that nothing breaks. This is the preferred method. - If you absolutely want to rename the column, then the steps would be:
- Create a column with the new name
- Start dual-writing to both the old and new column.
- Backfill the old column values into the new column.
- Change the field to start reading from the new column.
- Stop writing to the old column and remove references from the code.
- Drop the old column from the database.
- Generally, this is not worth doing and a lot of risk/effort compared to the reward.
The local database for siloed servers is separate from the database used for monolith operations. The siloed databases are named region
and control
matching the silo modes. Within django, the default
connection is for the region database, and the control
connection is for the control
database. The same database names are used by both sentry and getsentry.
# Copy your existing application data into the split databases.
bin/split-silo-database --database sentry --reset
# When working with getsentry run the following from getsentry root directory
bin/split-silo-database --database getsentry --reset
There is a script in both sentry & getsentry that are functionally equivalent. If you are working on getsentry, you need to use the getsentry script to ensure that all of your tables end up in the correct siloed database.
The split-silo-database
scripts use silo annotations on models to selectively dump your monolith database into the siloed databases.
You might need to set SENTRY_MONOLITH_REGION="us"
in your sentry config in order to successfully update the organization mappings.
You have two options for maintaining siloed databases:
- Run migrations on your monolith database, and use
split-silo-database
to rebuild your siloed databases. - Run migrations for siloed databases.
To run migrations on the siloed databases, run migrations in region and control mode.
# Run migrations for region mode
SENTRY_SILO_DEVSERVER=1 SENTRY_SILO_MODE=REGION SENTRY_REGION=us getsentry upgrade
# Run migrations for control
SENTRY_SILO_DEVSERVER=1 SENTRY_SILO_MODE=CONTROL getsentry upgrade
Employees Only
The following section covers migration deployment in Sentry's SaaS products.
We support two kinds of migrations in our SaaS deployments:
- Deploy time migrations
- Post-deploy migrations
Deployment migrations are run in each region and tenant before code is deployed. Deployment migrations are expected to finish quickly and all statements must complete within 5 seconds. If a migration could take longer because a large number of rows is being operated on, it should be deployed as a post-deploy migration instead. Deployment migrations are ideal for:
- Adding new tables and columns.
- Adding indexes to most tables.
- Removing columns and tables - as long as you follow the processes outlined above.
Post-deploy migrations are run manually by engineers after a migration has been deployed to all regions. During deployment, post-deploy migrations are marked as complete (faked) with django's fake migration behavior. When a post-deploy migration is run, it is run against all regions and tenants. Post-deploy migrations are triggered manually by engineers.
Post-deploy migrations are ideal for:
- Adding indexes to large tables, where adding the index would take longer than 5 seconds in any given region.
- Doing data backfills or mutations on tables with more than 50,000 rows.
Post-deploy migrations should not be used for:
- Column additions, removals or renames.
- Table creation.
Using post-deploy migrations for these operations will cause an outage.
Post-deploy migrations (both data and schema) are now run through a GoCD pipeline named post-deploy-migrations
.
To run a post-deploy migration, first locate the post-deploy-migrations job
. Click on the play button:
Under Materials
, input the getsentry SHA you want to run migrations from. The sha you choose should be one that contains the migration, and has been deployed to all regions.
Then click Environment Variables
and fill in a value for django_app
and django_migration
. The django_app should match the app name containing the migration. e.g. sentry
, getsentry
. Next, input the name of the migration to run, e.g. 0233_pickle_to_json_admin_auditlogentry
.
Just cancel the GoCD stage.
We don't have tooling to reverse migrations, so generally we introduce another migration which reverse the migration we want to reverse and run that.
GoCD won’t always be able to kill a long-running query - so instead we'll need to find the query run in pg_stat_activity
and kill it using pg_terminate_backend(pid)
. This will require assistance from SRE.
Our documentation is open source and available on GitHub. Your contributions are welcome, whether fixing a typo (drat!) or suggesting an update ("yeah, this would be better").