Multiple Databases

To allow GitLab to scale further we decomposed the GitLab application database into multiple databases. The two databases are main and ci. GitLab supports being run with either one database or two databases. On GitLab.com we are using two separate databases.

GitLab Schema

For properly discovering allowed patterns between different databases the GitLab application implements the database dictionary.

The database dictionary provides a virtual classification of tables into a gitlab_schema which conceptually is similar to PostgreSQL Schema. We decided as part of using database schemas to better isolated CI decomposed features that we cannot use PostgreSQL schema due to complex migration procedures. Instead we implemented the concept of application-level classification. Each table of GitLab needs to have a gitlab_schema assigned:

gitlab_main: describes all tables that are being stored in the main: database (for example, like projects, users).
gitlab_ci: describes all CI tables that are being stored in the ci: database (for example, ci_pipelines, ci_builds).
gitlab_geo: describes all Geo tables that are being stored in the geo: database (for example, like project_registry, secondary_usage_data).
gitlab_shared: describes all application tables that contain data across all decomposed databases (for example, loose_foreign_keys_deleted_records) for models that inherit from Gitlab::Database::SharedModel.
gitlab_internal: describes all internal tables of Rails and PostgreSQL (for example, ar_internal_metadata, schema_migrations, pg_*).
gitlab_pm: describes all tables that store package_metadata (it is an alias for gitlab_main).
...: more schemas to be introduced with additional decomposed databases

The usage of schema enforces the base class to be used:

ApplicationRecord for gitlab_main
Ci::ApplicationRecord for gitlab_ci
Geo::TrackingBase for gitlab_geo
Gitlab::Database::SharedModel for gitlab_shared
PackageMetadata::ApplicationRecord for gitlab_pm

The impact of `gitlab_schema`

The usage of gitlab_schema has a significant impact on the application. The gitlab_schema primary purpose is to introduce a barrier between different data access patterns.

This is used as a primary source of classification for:

Discovering cross-joins across tables from different schemas
Discovering cross-database transactions across tables from different schemas

The special purpose of `gitlab_shared`

gitlab_shared is a special case that describes tables or views that, by design, contain data across all decomposed databases. This classification describes application-defined tables (like loose_foreign_keys_deleted_records).

Be careful to use gitlab_shared as it requires special handling while accessing data. Since gitlab_shared shares not only structure but also data, the application needs to be written in a way that traverses all data from all databases in sequential manner.

Gitlab::Database::EachDatabase.each_model_connection([MySharedModel]) do |connection, connection_name|
  MySharedModel.select_all_data...
end

As such, migrations modifying data of gitlab_shared tables are expected to run across all decomposed databases.

The special purpose of `gitlab_internal`

gitlab_internal describes Rails-defined tables (like schema_migrations or ar_internal_metadata), as well as internal PostgreSQL tables (for example, pg_attribute). Its primary purpose is to support other databases, like Geo, that might be missing some of those application-defined gitlab_shared tables (like loose_foreign_keys_deleted_records), but are valid Rails databases.

The special purpose of `gitlab_pm`

gitlab_pm stores package metadata describing public repositories. This data is used for the License Compliance and Dependency Scanning product categories and is maintained by the Composition Analysis Group. It is an alias for gitlab_main intended to make it easier to route to a different database in the future.

Migrations

Read Migrations for Multiple Databases.

CI/CD Database

Configure single database

By default, GDK is configured to run with multiple databases.

WARNING: Switching back-and-forth between single and multiple databases in the same development instance is discouraged. Any data in the ci database will not be accessible in single database mode. For single database, you should use a separate development instance.

To configure GDK to use a single database:

On the GDK root directory, run:

gdk config set gitlab.rails.databases.ci.enabled false

Reconfigure GDK:
```
gdk reconfigure
```

To switch back to using multiple databases, set gitlab.rails.databases.ci.enabled to true and run gdk reconfigure.

Removing joins between `ci` and non `ci` tables

Queries that join across databases raise an error. Introduced in GitLab 14.3, for new queries only. Pre-existing queries do not raise an error.

Because GitLab can be run with multiple separate databases, referencing ci tables with non ci tables in a single query is not possible. Therefore, using any kind of JOIN in SQL queries will not work.

Suggestions for removing cross-database joins

The following sections are some real examples that were identified as joining across databases, along with possible suggestions on how to fix them.

Remove the code

The simplest solution we've seen several times now has been an existing scope that is unused. This is the easiest example to fix. So the first step is to investigate if the code is unused and then remove it. These are some real examples:

There may be more examples where the code is used, but we can evaluate if we need it or if the feature should behave this way. Before complicating things by adding new columns and tables, consider if you can simplify the solution and still meet the requirements. One case being evaluated involves changing how certain UsageData is calculated to remove a join query in https://gitlab.com/gitlab-org/gitlab/-/issues/336170. This is a good candidate to evaluate, because UsageData is not critical to users and it may be possible to get a similarly useful metric with a simpler approach. Alternatively we may find that nobody is using these metrics, so we can remove them.

Use `preload` instead of `includes`

The includes and preload methods in Rails are both ways to avoid an N+1 query. The includes method in Rails uses a heuristic approach to determine if it needs to join to the table, or if it can load all of the records in a separate query. This method assumes it needs to join if it thinks you need to query the columns from the other table, but sometimes this method gets it wrong and executes a join even when not needed. In this case using preload to explicitly load the data in a separate query allows you to avoid the join, while still avoiding the N+1 query.

You can see a real example of this solution being used in https://gitlab.com/gitlab-org/gitlab/-/merge_requests/67655.

Remove a redundant join

Sometimes there are cases where a query is doing excess (or redundant) joins.

A common example occurs where a query is joining from A to C, via some table with both foreign keys, B. When you only care about counting how many rows there are in C and if there are foreign keys and NOT NULL constraints on the foreign keys in B, then it might be enough to count those rows. For example, in MR 71811, it was previously doing project.runners.count, which would produce a query like:

select count(*) from projects
inner join ci_runner_projects on ci_runner_projects.project_id = projects.id
where ci_runner_projects.runner_id IN (1, 2, 3)

This was changed to avoid the cross-join by changing the code to project.runner_projects.count. It produces the same response with the following query:

select count(*) from ci_runner_projects
where ci_runner_projects.runner_id IN (1, 2, 3)

Another common redundant join is joining all the way to another table, then filtering by primary key when you could have instead filtered on a foreign key. See an example in MR 71614. The previous code was joins(scan: :build).where(ci_builds: { id: build_ids }), which generated a query like:

select ...
inner join security_scans
inner join ci_builds on security_scans.build_id = ci_builds.id
where ci_builds.id IN (1, 2, 3)

However, as security_scans already has a foreign key build_id, the code can be changed to joins(:scan).where(security_scans: { build_id: build_ids }), which produces the same response with the following query:

select ...
inner join security_scans
where security_scans.build_id IN (1, 2, 3)

Both of these examples of removing redundant joins remove the cross-joins, but they have the added benefit of producing simpler and faster queries.

Limited pluck followed by a find

Using pluck or pick to get an array of ids is not advisable unless the returned array is guaranteed to be bounded in size. Usually this is a good pattern where you know the result will be at most 1, or in cases where you have a list of in memory ids (or usernames) that need to be mapped to another list of equal size. It would not be suitable when mapping a list of ids in a one-to-many relationship as the result will be unbounded. We can then use the returned ids to obtain the related record:

allowed_user_id = board_user_finder
  .where(user_id: params['assignee_id'])
  .pick(:user_id)

User.find_by(id: allowed_user_id)

You can see an example where this was used in https://gitlab.com/gitlab-org/gitlab/-/merge_requests/126856

Sometimes it might seem easy to convert a join into a pluck but often this results in loading an unbounded amount of ids into memory and then re-serializing those in a following query back to Postgres. These cases do not scale and we recommend attempting one of the other options. It might seem like a good idea to just apply some limit to the plucked data to have bounded memory but this introduces unpredictable results for users and often is most problematic for our largest customers (including ourselves), and as such we advise against it.

De-normalize some foreign key to the table

De-normalization refers to adding redundant precomputed (duplicated) data to a table to simplify certain queries or to improve performance. In this case, it can be useful when you are doing a join that involves three tables, where you are joining through some intermediate table.

Generally when modeling a database schema, a "normalized" structure is preferred because of the following reasons:

Duplicate data uses extra storage.
Duplicate data needs to be kept in sync.

Sometimes normalized data is less performant so de-normalization has been a common technique GitLab has used to improve the performance of database queries for a while. The above problems are mitigated when the following conditions are met:

There isn't much data (for example, it's just an integer column).
The data does not update often (for example, the project_id column is almost never updated for most tables).

One example we found was the security_scans table. This table has a foreign key security_scans.build_id which allows you to join to the build. Therefore you could join to the project like so:

select projects.* from security_scans
inner join ci_builds on security_scans.build_id = ci_builds.id
inner join projects on ci_builds.project_id = projects.id

The problem with this query is that ci_builds is in a different database from the other two tables.

The solution in this case is to add the project_id column to security_scans. This doesn't use much extra storage, and due to the way these features work, it's never updated (a build never moves projects).

This simplified the query to:

select projects.* from security_scans
inner join projects on security_scans.project_id = projects.id

This also improves performance because you don't need to join through an extra table.

You can see this approach implemented in https://gitlab.com/gitlab-org/gitlab/-/merge_requests/66963 . This MR also de-normalizes pipeline_id to fix a similar query.

De-normalize into an extra table

Sometimes the previous de-normalization (adding an extra column) doesn't work for your specific case. This may be due to the fact that your data is not 1:1, or because the table you're adding to is already too wide (for example, the projects table shouldn't have more columns added).

In this case you may decide to just store the extra data in a separate table.

One example where this approach is being used was to implement the Project.with_code_coverage scope. This scope was essentially used to narrow down a list of projects to only those that have at one point in time used code coverage features. This query (simplified) was:

select projects.* from projects
inner join ci_daily_build_group_report_results on ci_daily_build_group_report_results.project_id = projects.id
where ((data->'coverage') is not null)
and ci_daily_build_group_report_results.default_branch = true
group by projects.id

This work is still in progress but the current plan is to introduce a new table called projects_with_ci_feature_usage which has 2 columns project_id and ci_feature. This table would be written to the first time a project creates a ci_daily_build_group_report_results for code coverage. Therefore the new query would be:

select projects.* from projects
inner join projects_with_ci_feature_usage on projects_with_ci_feature_usage.project_id = projects.id
where projects_with_ci_feature_usage.ci_feature = 'code_coverage'

The above example uses as a text column for simplicity but we should probably use an enum to save space.

The downside of this new design is that this may need to be updated (removed if the ci_daily_build_group_report_results is deleted). Depending on your domain, however, this may not be necessary because deletes are edge cases or impossible, or because the user impact of seeing the project on the list page may not be problematic. It's also possible to implement the logic to delete these rows if or whenever necessary in your domain.

Finally, this de-normalization and new query also improves performance because it does less joins and needs less filtering.

Use `disable_joins` for `has_one` or `has_many` `through:` relations

Sometimes a join query is caused by using has_one ... through: or has_many ... through: across tables that span the different databases. These joins sometimes can be solved by adding disable_joins:true. This is a Rails feature which we backported. We also extended the feature to allow a lambda syntax for enabling disable_joins with a feature flag. If you use this feature we encourage using a feature flag as it mitigates risk if there is some serious performance regression.

You can see an example where this was used in https://gitlab.com/gitlab-org/gitlab/-/merge_requests/66709/diffs.

With any change to DB queries it is important to analyze and compare the SQL before and after the change. disable_joins can introduce very poorly performing code depending on the actual logic of the has_many or has_one relationship. The key thing to look for is whether any of the intermediate result sets used to construct the final result set have an unbounded amount of data loaded. The best way to tell is by looking at the SQL generated and confirming that each one is limited in some way. You can tell by either a LIMIT 1 clause or by WHERE clause that is limiting based on a unique column. Any unbounded intermediate dataset could lead to loading too many IDs into memory.

An example where you may see very poor performance is the following hypothetical code:

class Project
  has_many :pipelines
  has_many :builds, through: :pipelines
end

class Pipeline
  has_many :builds
end

class Build
  belongs_to :pipeline
end

def some_action
  @builds = Project.find(5).builds.order(created_at: :desc).limit(10)
end

In the above case some_action will generate a query like:

select * from builds
inner join pipelines on builds.pipeline_id = pipelines.id
where pipelines.project_id = 5
order by builds.created_at desc
limit 10

However, if you changed the relation to be:

class Project
  has_many :pipelines
  has_many :builds, through: :pipelines, disable_joins: true
end

Then you would get the following 2 queries:

select id from pipelines where project_id = 5;

select * from builds where pipeline_id in (...)
order by created_at desc
limit 10;

Because the first query does not limit by any unique column or have a LIMIT clause, it can load an unlimited number of pipeline IDs into memory, which are then sent in the following query. This can lead to very poor performance in the Rails application and the database. In cases like this, you might need to re-write the query or look at other patterns described above for removing cross-joins.

How to validate you have correctly removed a cross-join

RSpec is configured to automatically validate all SQL queries do not join across databases. If this validation is disabled in spec/support/database/cross-join-allowlist.yml then you can still validate an isolated code block using with_cross_joins_prevented.

You can use this method like so:

it 'does not join across databases' do
  with_cross_joins_prevented do
    ::Ci::Build.joins(:project).to_a
  end
end

This will raise an exception if the query joins across the two databases. The previous example is fixed by removing the join, like so:

it 'does not join across databases' do
  with_cross_joins_prevented do
    ::Ci::Build.preload(:project).to_a
  end
end

You can see a real example of using this method for fixing a cross-join in https://gitlab.com/gitlab-org/gitlab/-/merge_requests/67655.

Allowlist for existing cross-joins

A cross-join across databases can be explicitly allowed by wrapping the code in the ::Gitlab::Database.allow_cross_joins_across_databases helper method. Alternative way is to mark a given relation as relation.allow_cross_joins_across_databases.

This method should only be used:

For existing code.
If the code is required to help migrate away from a cross-join. For example, in a migration that backfills data for future use to remove a cross-join.

The allow_cross_joins_across_databases helper method can be used as follows:

# Scope the block executing a object from database
::Gitlab::Database.allow_cross_joins_across_databases(url: 'https://gitlab.com/gitlab-org/gitlab/-/issues/336590') do
  subject.perform(1, 4)
end

# Mark a relation as allowed to cross-join databases
def find_actual_head_pipeline
  all_pipelines
    .allow_cross_joins_across_databases(url: 'https://gitlab.com/gitlab-org/gitlab/-/issues/336891')
    .for_sha_or_source_sha(diff_head_sha)
    .first
end

The url parameter should point to an issue with a milestone for when we intend to fix the cross-join. If the cross-join is being used in a migration, we do not need to fix the code. See https://gitlab.com/gitlab-org/gitlab/-/issues/340017 for more details.

Removing cross-database transactions

When dealing with multiple databases, it's important to pay close attention to data modification that affects more than one database. Introduced GitLab 14.4, an automated check prevents cross-database modifications.

When at least two different databases are modified during a transaction initiated on any database server, the application triggers a cross-database modification error (only in test environment).

Example:

# Open transaction on Main DB
ApplicationRecord.transaction do
  ci_build.update!(updated_at: Time.current) # UPDATE on CI DB
  ci_build.project.update!(updated_at: Time.current) # UPDATE on Main DB
end
# raises error: Cross-database data modification of 'main, ci' were detected within
# a transaction modifying the 'ci_build, projects' tables

The code example above updates the timestamp for two records within a transaction. With the ongoing work on the CI database decomposition, we cannot ensure the schematics of a database transaction. If the second update query fails, the first update query will not be rolled back because the ci_build record is located on a different database server. For more information, look at the transaction guidelines page.

Fixing cross-database errors

Removing the transaction block

Without an open transaction, the cross-database modification check cannot raise an error. By making this change, we sacrifice consistency. In case of an application failure after the first UPDATE query, the second UPDATE query will never execute.

The same code without the transaction block:

ci_build.update!(updated_at: Time.current) # CI DB
ci_build.project.update!(updated_at: Time.current) # Main DB

Asynchronous processing

If we need more guarantee that an operation finishes the work consistently we can execute it within a background job. A background job is scheduled asynchronously and retried several times in case of an error. There is still a very small chance of introducing inconsistency.

Example:

current_time = Time.current

MyAsyncConsistencyJob.perform_async(cu_build.id)

ci_build.update!(updated_at: current_time)
ci_build.project.update!(updated_at: current_time)

The MyAsyncConsistencyJob would also attempt to update the timestamp if they differ.

Aiming for perfect consistency

At this point, we don't have the tooling (we might not even need it) to ensure similar consistency characteristics as we had with one database. If you think that the code you're working on requires these properties, then you can disable the cross-database modification check in your tests by wrapping the offending test code with a block and create a follow-up issue.

allow_cross_database_modification_within_transaction(url: 'gitlab issue URL') do
  ApplicationRecord.transaction do
    ci_build.update!(updated_at: Time.current) # UPDATE on CI DB
    ci_build.project.update!(updated_at: Time.current) # UPDATE on Main DB
  end
end

Don't hesitate to reach out to the Pods group for advice.

Avoid `dependent: :nullify` and `dependent: :destroy` across databases

There may be cases where we want to use dependent: :nullify or dependent: :destroy across databases. This is technically possible, but it's problematic because these hooks run in the context of an outer transaction from the call to #destroy, which creates a cross-database transaction and we are trying to avoid that. Cross-database transactions caused this way could lead to confusing outcomes when we switch to decomposed, because now you have some queries happening outside the transaction and they may be partially applied while the outer transaction fails, which could lead to surprising bugs.

For non-trivial objects that need to clean up data outside the database (for example, object storage), we recommend the setting dependent: :restrict_with_error. Such objects should be removed explicitly ahead of time. Using dependent: :restrict_with_error ensures that we forbid destroying the parent object if something is not cleaned up.

If all you need to do is clean up the child records themselves from PostgreSQL, consider using loose foreign keys.

Foreign keys that cross databases

There are many places where we use foreign keys that reference across the two databases. This is not possible to do with two separate PostgreSQL databases, so we need to replicate the behavior we get from PostgreSQL in a performant way. We can't, and shouldn't, try to replicate the data guarantees given by PostgreSQL which prevent creating invalid references, but we still need a way to replace cascading deletes so we don't end up with orphaned data or records that point to nowhere, which might lead to bugs. As such we created "loose foreign keys" which is an asynchronous process of cleaning up orphaned records.

Testing for multiple databases

In our testing CI pipelines, we test GitLab by default with multiple databases set up, using both main and ci databases. But in merge requests, for example when we modify some database-related code or add the label ~"pipeline:run-single-db" to the MR, we additionally run our tests in two other database modes: single-db and single-db-ci-connection.

To handle situations where our tests need to run in specific database modes, we have some RSpec helpers to limit the modes where tests can run, and skip them on any other modes.

Helper name	Test runs
`skip_if_shared_database(:ci)`	On multiple databases
`skip_if_database_exists(:ci)`	On single-db and single-db-ci-connection
`skip_if_multiple_databases_are_setup(:ci)`	Only on single-db
`skip_if_multiple_databases_not_setup(:ci)`	On single-db-ci-connection and multiple databases

Testing for multiple databases, including main_clusterwide

By default, we do not setup the main_clusterwide connection in CI pipelines. However, if you add the label ~"pipeline:run-clusterwide-db", the pipelines will run with 3 connections, main, ci and main_clusterwide.

NOTE: This setup is not completely ready yet, and running pipelines in the setup may fail some jobs. As of July 2023, this is only used by group::tenant scale to test out changes while building Cells.

Locking writes on the tables that don't belong to the database schemas

When the CI database is promoted and the two databases are fully split, as an extra safeguard against creating a split brain situation, run the Rake task gitlab:db:lock_writes. This command locks writes on:

The gitlab_main tables on the CI Database.
The gitlab_ci tables on the Main Database.

This Rake task adds triggers to all the tables, to prevent any INSERT, UPDATE, DELETE, or TRUNCATE statements from running against the tables that need to be locked.

If this task was run against a GitLab setup that uses only a single database for both gitlab_main and gitlab_ci tables, then no tables will be locked.

To undo the operation, run the opposite Rake task: gitlab:db:unlock_writes.

Truncating tables

When the databases main and ci are fully split, we can free up disk space by truncating tables. This results in a smaller data set: For example, the data in users table on CI database is no longer read and also no longer updated. So this data can be removed by truncating the tables.

For this purpose, GitLab provides two Rake tasks, one for each database:

gitlab:db:truncate_legacy_tables:main will truncate the CI tables in Main database.
gitlab:db:truncate_legacy_tables:ci will truncate the Main tables in CI database.

NOTE: These tasks can only be run when the tables in the database are locked for writes.

WARNING: The examples in this section use DRY_RUN=true. This ensures no data is actually truncated. GitLab highly recommends to have a backup available before you run any of these tasks without DRY_RUN=true.

These tasks have the option to see what they do without actually changing the data:

$ sudo DRY_RUN=true gitlab-rake gitlab:db:truncate_legacy_tables:main
I, [2023-07-14T17:08:06.665151 #92505]  INFO -- : DRY RUN:
I, [2023-07-14T17:08:06.761586 #92505]  INFO -- : Truncating legacy tables for the database main
I, [2023-07-14T17:08:06.761709 #92505]  INFO -- : SELECT set_config('lock_writes.ci_build_needs', 'false', false)
I, [2023-07-14T17:08:06.765272 #92505]  INFO -- : SELECT set_config('lock_writes.ci_build_pending_states', 'false', false)
I, [2023-07-14T17:08:06.768220 #92505]  INFO -- : SELECT set_config('lock_writes.ci_build_report_results', 'false', false)
[...]
I, [2023-07-14T17:08:06.957294 #92505]  INFO -- : TRUNCATE TABLE ci_build_needs, ci_build_pending_states, ci_build_report_results, ci_build_trace_chunks, ci_build_trace_metadata, ci_builds, ci_builds_metadata, ci_builds_runner_session, ci_cost_settings, ci_daily_build_group_report_results, ci_deleted_objects, ci_editor_ai_conversation_messages, ci_freeze_periods, ci_group_variables, ci_instance_variables, ci_job_artifact_states, ci_job_artifacts, ci_job_token_project_scope_links, ci_job_variables, ci_minutes_additional_packs, ci_namespace_mirrors, ci_namespace_monthly_usages, ci_partitions, ci_pending_builds, ci_pipeline_artifacts, ci_pipeline_chat_data, ci_pipeline_messages, ci_pipeline_metadata, ci_pipeline_schedule_variables, ci_pipeline_schedules, ci_pipeline_variables, ci_pipelines, ci_pipelines_config, ci_platform_metrics, ci_project_mirrors, ci_project_monthly_usages, ci_refs, ci_resource_groups, ci_resources, ci_runner_machines, ci_runner_namespaces, ci_runner_projects, ci_runner_versions, ci_runners, ci_running_builds, ci_secure_file_states, ci_secure_files, ci_sources_pipelines, ci_sources_projects, ci_stages, ci_subscriptions_projects, ci_trigger_requests, ci_triggers, ci_unit_test_failures, ci_unit_tests, ci_variables, external_pull_requests, p_ci_builds, p_ci_builds_metadata, p_ci_job_annotations, p_ci_runner_machine_builds, taggings, tags RESTRICT

The tasks will first find out the tables that need to be truncated. Truncation will happen in stages because we need to limit the amount of data removed in one database transaction. The tables are processed in a specific order depending on the definition of the foreign keys. The number of tables processed in one stage can be changed by adding a number when invoking the task. The default value is 5:

sudo DRY_RUN=true gitlab-rake gitlab:db:truncate_legacy_tables:main\[10\]

It is also possible to limit the number of tables to be truncated by setting the UNTIL_TABLE variable. For example in this case, the process will stop when ci_unit_test_failures has been truncated:

sudo DRY_RUN=true UNTIL_TABLE=ci_unit_test_failures gitlab-rake gitlab:db:truncate_legacy_tables:main