debian-mirror-gitlab/doc/development/background_migrations.md

319 lines
12 KiB
Markdown
Raw Normal View History

2017-09-10 17:25:29 +05:30
# Background Migrations
Background migrations can be used to perform data migrations that would
otherwise take a very long time (hours, days, years, etc) to complete. For
example, you can use background migrations to migrate data so that instead of
storing data in a single JSON column the data is stored in a separate table.
2018-11-18 11:00:15 +05:30
If the database cluster is considered to be in an unhealthy state, background
migrations automatically reschedule themselves for a later point in time.
2017-09-10 17:25:29 +05:30
## When To Use Background Migrations
2018-11-20 20:47:30 +05:30
> **Note:**
> When adding background migrations _you must_ make sure they are announced in the
> monthly release post along with an estimate of how long it will take to complete
> the migrations.
2017-09-10 17:25:29 +05:30
In the vast majority of cases you will want to use a regular Rails migration
instead. Background migrations should _only_ be used when migrating _data_ in
tables that have so many rows this process would take hours when performed in a
regular Rails migration.
Background migrations _may not_ be used to perform schema migrations, they
should only be used for data migrations.
Some examples where background migrations can be useful:
* Migrating events from one table to multiple separate tables.
* Populating one column based on JSON stored in another column.
2018-10-15 14:42:47 +05:30
* Migrating data that depends on the output of external services (e.g. an API).
2017-09-10 17:25:29 +05:30
## Isolation
Background migrations must be isolated and can not use application code (e.g.
models defined in `app/models`). Since these migrations can take a long time to
run it's possible for new versions to be deployed while they are still running.
It's also possible for different migrations to be executed at the same time.
This means that different background migrations should not migrate data in a
way that would cause conflicts.
## Idempotence
Background migrations are executed in a context of a Sidekiq process.
Usual Sidekiq rules apply, especially the rule that jobs should be small
and idempotent.
See [Sidekiq best practices guidelines](https://github.com/mperham/sidekiq/wiki/Best-Practices)
for more details.
Make sure that in case that your migration job is going to be retried data
2018-10-15 14:42:47 +05:30
integrity is guaranteed.
2017-09-10 17:25:29 +05:30
## How It Works
Background migrations are simple classes that define a `perform` method. A
Sidekiq worker will then execute such a class, passing any arguments to it. All
migration classes must be defined in the namespace
`Gitlab::BackgroundMigration`, the files should be placed in the directory
`lib/gitlab/background_migration/`.
## Scheduling
Scheduling a migration can be done in either a regular migration or a
post-deployment migration. To do so, simply use the following code while
replacing the class name and arguments with whatever values are necessary for
your migration:
```ruby
BackgroundMigrationWorker.perform_async('BackgroundMigrationClassName', [arg1, arg2, ...])
```
Usually it's better to enqueue jobs in bulk, for this you can use
2018-03-17 18:26:18 +05:30
`BackgroundMigrationWorker.bulk_perform_async`:
2017-09-10 17:25:29 +05:30
```ruby
2018-03-17 18:26:18 +05:30
BackgroundMigrationWorker.bulk_perform_async(
2017-09-10 17:25:29 +05:30
[['BackgroundMigrationClassName', [1]],
['BackgroundMigrationClassName', [2]]]
)
```
You'll also need to make sure that newly created data is either migrated, or
saved in both the old and new version upon creation. For complex and time
consuming migrations it's best to schedule a background job using an
`after_create` hook so this doesn't affect response timings. The same applies to
updates. Removals in turn can be handled by simply defining foreign keys with
cascading deletes.
If you would like to schedule jobs in bulk with a delay, you can use
2018-03-17 18:26:18 +05:30
`BackgroundMigrationWorker.bulk_perform_in`:
2017-09-10 17:25:29 +05:30
```ruby
jobs = [['BackgroundMigrationClassName', [1]],
['BackgroundMigrationClassName', [2]]]
2018-03-17 18:26:18 +05:30
BackgroundMigrationWorker.bulk_perform_in(5.minutes, jobs)
2017-09-10 17:25:29 +05:30
```
2018-03-17 18:26:18 +05:30
### Rescheduling background migrations
If one of the background migrations contains a bug that is fixed in a patch
release, the background migration needs to be rescheduled so the migration would
be repeated on systems that already performed the initial migration.
When you reschedule the background migration, make sure to turn the original
scheduling into a no-op by clearing up the `#up` and `#down` methods of the
migration performing the scheduling. Otherwise the background migration would be
scheduled multiple times on systems that are upgrading multiple patch releases at
once.
2017-09-10 17:25:29 +05:30
## Cleaning Up
>**Note:**
Cleaning up any remaining background migrations _must_ be done in either a major
or minor release, you _must not_ do this in a patch release.
Because background migrations can take a long time you can't immediately clean
things up after scheduling them. For example, you can't drop a column that's
used in the migration process as this would cause jobs to fail. This means that
you'll need to add a separate _post deployment_ migration in a future release
that finishes any remaining jobs before cleaning things up (e.g. removing a
column).
As an example, say you want to migrate the data from column `foo` (containing a
big JSON blob) to column `bar` (containing a string). The process for this would
roughly be as follows:
1. Release A:
2018-11-20 20:47:30 +05:30
1. Create a migration class that perform the migration for a row with a given ID.
1. Deploy the code for this release, this should include some code that will
schedule jobs for newly created data (e.g. using an `after_create` hook).
1. Schedule jobs for all existing rows in a post-deployment migration. It's
possible some newly created rows may be scheduled twice so your migration
should take care of this.
2017-09-10 17:25:29 +05:30
1. Release B:
2018-11-20 20:47:30 +05:30
1. Deploy code so that the application starts using the new column and stops
scheduling jobs for newly created data.
1. In a post-deployment migration you'll need to ensure no jobs remain.
1. Use `Gitlab::BackgroundMigration.steal` to process any remaining
jobs in Sidekiq.
1. Reschedule the migration to be run directly (i.e. not through Sidekiq)
on any rows that weren't migrated by Sidekiq. This can happen if, for
instance, Sidekiq received a SIGKILL, or if a particular batch failed
enough times to be marked as dead.
1. Remove the old column.
2017-09-10 17:25:29 +05:30
2018-10-15 14:42:47 +05:30
This may also require a bump to the [import/export version][import-export], if
importing a project from a prior version of GitLab requires the data to be in
the new format.
2017-09-10 17:25:29 +05:30
## Example
To explain all this, let's use the following example: the table `services` has a
field called `properties` which is stored in JSON. For all rows you want to
extract the `url` key from this JSON object and store it in the `services.url`
column. There are millions of services and parsing JSON is slow, thus you can't
do this in a regular migration.
To do this using a background migration we'll start with defining our migration
class:
```ruby
class Gitlab::BackgroundMigration::ExtractServicesUrl
class Service < ActiveRecord::Base
self.table_name = 'services'
end
def perform(service_id)
# A row may be removed between scheduling and starting of a job, thus we
# need to make sure the data is still present before doing any work.
service = Service.select(:properties).find_by(id: service_id)
return unless service
begin
json = JSON.load(service.properties)
rescue JSON::ParserError
# If the JSON is invalid we don't want to keep the job around forever,
# instead we'll just leave the "url" field to whatever the default value
# is.
return
end
service.update(url: json['url']) if json['url']
end
end
```
Next we'll need to adjust our code so we schedule the above migration for newly
created and updated services. We can do this using something along the lines of
the following:
```ruby
class Service < ActiveRecord::Base
after_commit :schedule_service_migration, on: :update
after_commit :schedule_service_migration, on: :create
def schedule_service_migration
BackgroundMigrationWorker.perform_async('ExtractServicesUrl', [id])
end
end
```
We're using `after_commit` here to ensure the Sidekiq job is not scheduled
before the transaction completes as doing so can lead to race conditions where
the changes are not yet visible to the worker.
Next we'll need a post-deployment migration that schedules the migration for
existing data. Since we're dealing with a lot of rows we'll schedule jobs in
batches instead of doing this one by one:
```ruby
class ScheduleExtractServicesUrl < ActiveRecord::Migration
disable_ddl_transaction!
class Service < ActiveRecord::Base
self.table_name = 'services'
end
def up
Service.select(:id).in_batches do |relation|
jobs = relation.pluck(:id).map do |id|
['ExtractServicesUrl', [id]]
end
2018-03-17 18:26:18 +05:30
BackgroundMigrationWorker.bulk_perform_async(jobs)
2017-09-10 17:25:29 +05:30
end
end
def down
end
end
```
Once deployed our application will continue using the data as before but at the
same time will ensure that both existing and new data is migrated.
In the next release we can remove the `after_commit` hooks and related code. We
will also need to add a post-deployment migration that consumes any remaining
2018-03-17 18:26:18 +05:30
jobs and manually run on any un-migrated rows. Such a migration would look like
this:
2017-09-10 17:25:29 +05:30
```ruby
class ConsumeRemainingExtractServicesUrlJobs < ActiveRecord::Migration
disable_ddl_transaction!
2018-03-17 18:26:18 +05:30
class Service < ActiveRecord::Base
include ::EachBatch
self.table_name = 'services'
end
2017-09-10 17:25:29 +05:30
def up
2018-03-17 18:26:18 +05:30
# This must be included
2017-09-10 17:25:29 +05:30
Gitlab::BackgroundMigration.steal('ExtractServicesUrl')
2018-03-17 18:26:18 +05:30
# This should be included, but can be skipped - see below
Service.where(url: nil).each_batch(of: 50) do |batch|
range = batch.pluck('MIN(id)', 'MAX(id)').first
Gitlab::BackgroundMigration::ExtractServicesUrl.new.perform(*range)
end
2017-09-10 17:25:29 +05:30
end
def down
end
end
```
2018-03-17 18:26:18 +05:30
The final step runs for any un-migrated rows after all of the jobs have been
processed. This is in case a Sidekiq process running the background migrations
received SIGKILL, leading to the jobs being lost. (See
[more reliable Sidekiq queue][reliable-sidekiq] for more information.)
If the application does not depend on the data being 100% migrated (for
instance, the data is advisory, and not mission-critical), then this final step
can be skipped.
2017-09-10 17:25:29 +05:30
This migration will then process any jobs for the ExtractServicesUrl migration
and continue once all jobs have been processed. Once done you can safely remove
the `services.properties` column.
## Testing
It is required to write tests for background migrations' scheduling migration
(either a regular migration or a post deployment migration), background
migration itself and a cleanup migration. You can use the `:migration` RSpec
tag when testing a regular / post deployment migration.
See [README][migrations-readme].
When you do that, keep in mind that `before` and `after` RSpec hooks are going
to migrate you database down and up, which can result in other background
migrations being called. That means that using `spy` test doubles with
`have_received` is encouraged, instead of using regular test doubles, because
your expectations defined in a `it` block can conflict with what is being
called in RSpec hooks. See [gitlab-org/gitlab-ce#35351][issue-rspec-hooks]
for more details.
## Best practices
2018-11-18 11:00:15 +05:30
1. Make sure to know how much data you're dealing with
2017-09-10 17:25:29 +05:30
1. Make sure that background migration jobs are idempotent.
1. Make sure that tests you write are not false positives.
2018-03-17 18:26:18 +05:30
1. Make sure that if the data being migrated is critical and cannot be lost, the
clean-up migration also checks the final state of the data before completing.
2018-11-18 11:00:15 +05:30
1. Make sure to know how much time it'll take to run all scheduled migrations
1. When migrating many columns, make sure it won't generate too many
dead tuples in the process (you may need to directly query the number of dead tuples
and adjust the scheduling according to this piece of data)
1. Make sure to discuss the numbers with a database specialist, the migration may add
more pressure on DB than you expect (measure on staging,
or ask someone to measure on production)
2017-09-10 17:25:29 +05:30
[migrations-readme]: https://gitlab.com/gitlab-org/gitlab-ce/blob/master/spec/migrations/README.md
[issue-rspec-hooks]: https://gitlab.com/gitlab-org/gitlab-ce/issues/35351
2018-03-17 18:26:18 +05:30
[reliable-sidekiq]: https://gitlab.com/gitlab-org/gitlab-ce/issues/36791
2018-10-15 14:42:47 +05:30
[import-export]: ../user/project/settings/import_export.md