Loose foreign keys

Problem statement

In relational databases (including PostgreSQL), foreign keys provide a way to link two database tables together, and ensure data-consistency between them. In GitLab, foreign keys are vital part of the database design process. Most of our database tables have foreign keys.

With the ongoing database decomposition work, linked records might be present on two different database servers. Ensuring data consistency between two databases is not possible with standard PostgreSQL foreign keys. PostgreSQL does not support foreign keys operating across multiple database servers.

Example:

Database "Main": projects table
Database "CI": ci_pipelines table

A project can have many pipelines. When a project is deleted, the associated ci_pipeline (via the project_id column) records must be also deleted.

With a multi-database setup, this cannot be achieved with foreign keys.

Asynchronous approach

Our preferred approach to this problem is eventual consistency. With the loose foreign keys feature, we can configure delayed association cleanup without negatively affecting the application performance.

How it works

In the previous example, a record in the projects table can have multiple ci_pipeline records. To keep the cleanup process separate from the actual parent record deletion, we can:

Create a DELETE trigger on the projects table. Record the deletions in a separate table (deleted_records).
A job checks the deleted_records table every minute or two.
For each record in the table, delete the associated ci_pipelines records using the project_id column.

NOTE: For this procedure to work, we must register which tables to clean up asynchronously.

The `scripts/decomposition/generate-loose-foreign-key`

We built an automation tool to aid migration of foreign keys into loose foreign keys as part of decomposition effort. It presents existing keys and allows chosen foreign keys to be automatically converted into loose foreign keys. This ensures consistency between foreign key and loose foreign key definitions, and ensures that they are properly tested.

WARNING: We strongly advise you to use the automation script for swapping any foreign key to a loose foreign key.

The tool ensures that all aspects of swapping a foreign key are covered. This includes:

Creating a migration to remove a foreign key.
Updating db/structure.sql with the new migration.
Updating config/gitlab_loose_foreign_keys.yml to add the new loose foreign key.
Creating or updating a model's specs to ensure that the loose foreign key is properly supported.

The tool is located at scripts/decomposition/generate-loose-foreign-key:

$ scripts/decomposition/generate-loose-foreign-key -h

Usage: scripts/decomposition/generate-loose-foreign-key [options] <filters...>
    -c, --cross-schema               Show only cross-schema foreign keys
    -n, --dry-run                    Do not execute any commands (dry run)
    -r, --[no-]rspec                 Create or not a rspecs automatically
    -h, --help                       Prints this help

For the migration of cross-schema foreign keys, we use the -c modifier to show the foreign keys yet to migrate:

$ scripts/decomposition/generate-loose-foreign-key -c
Re-creating current test database
Dropped database 'gitlabhq_test_ee'
Dropped database 'gitlabhq_geo_test_ee'
Created database 'gitlabhq_test_ee'
Created database 'gitlabhq_geo_test_ee'

Showing cross-schema foreign keys (20):
   ID | HAS_LFK |                                     FROM |                   TO |                         COLUMN |       ON_DELETE
    0 |       N |                                ci_builds |             projects |                     project_id |         cascade
    1 |       N |                         ci_job_artifacts |             projects |                     project_id |         cascade
    2 |       N |                             ci_pipelines |             projects |                     project_id |         cascade
    3 |       Y |                             ci_pipelines |       merge_requests |               merge_request_id |         cascade
    4 |       N |                   external_pull_requests |             projects |                     project_id |         cascade
    5 |       N |                     ci_sources_pipelines |             projects |                     project_id |         cascade
    6 |       N |                                ci_stages |             projects |                     project_id |         cascade
    7 |       N |                    ci_pipeline_schedules |             projects |                     project_id |         cascade
    8 |       N |                       ci_runner_projects |             projects |                     project_id |         cascade
    9 |       Y |             dast_site_profiles_pipelines |         ci_pipelines |                 ci_pipeline_id |         cascade
   10 |       Y |                   vulnerability_feedback |         ci_pipelines |                    pipeline_id |         nullify
   11 |       N |                             ci_variables |             projects |                     project_id |         cascade
   12 |       N |                                  ci_refs |             projects |                     project_id |         cascade
   13 |       N |                       ci_builds_metadata |             projects |                     project_id |         cascade
   14 |       N |                ci_subscriptions_projects |             projects |          downstream_project_id |         cascade
   15 |       N |                ci_subscriptions_projects |             projects |            upstream_project_id |         cascade
   16 |       N |                      ci_sources_projects |             projects |              source_project_id |         cascade
   17 |       N |         ci_job_token_project_scope_links |             projects |              source_project_id |         cascade
   18 |       N |         ci_job_token_project_scope_links |             projects |              target_project_id |         cascade
   19 |       N |                ci_project_monthly_usages |             projects |                     project_id |         cascade

To match foreign key (FK), write one or many filters to match against FROM/TO/COLUMN:
- scripts/decomposition/generate-loose-foreign-key (filters...)
- scripts/decomposition/generate-loose-foreign-key ci_job_artifacts project_id
- scripts/decomposition/generate-loose-foreign-key dast_site_profiles_pipelines

The command accepts a list of regular expressions to match from, to, or column for the purpose of the foreign key generation. For example, run this to swap all foreign keys for ci_job_token_project_scope_links for the decomposed database:

scripts/decomposition/generate-loose-foreign-key -c ci_job_token_project_scope_links

To swap only the source_project_id of ci_job_token_project_scope_links for the decomposed database, run:

scripts/decomposition/generate-loose-foreign-key -c ci_job_token_project_scope_links source_project_id

To match the exact name of a table or columns, you can make use of the regular expressions position anchors ^ and $. For example, this command matches only the foreign keys on the events table only, but not on the table incident_management_timeline_events.

scripts/decomposition/generate-loose-foreign-key -n ^events$

To swap all the foreign keys (all having _id appended), but not create a new branch (only commit the changes) and not create RSpec tests, run:

scripts/decomposition/generate-loose-foreign-key -c --no-branch --no-rspec _id

To swap all foreign keys referencing projects, but not create a new branch (only commit the changes), run:

scripts/decomposition/generate-loose-foreign-key -c --no-branch projects

Example migration and configuration

Configure the loose foreign key

Loose foreign keys are defined in a YAML file. The configuration requires the following information:

Parent table name (projects)
Child table name (ci_pipelines)
The data cleanup method (async_delete or async_nullify)

The YAML file is located at config/gitlab_loose_foreign_keys.yml. The file groups foreign key definitions by the name of the child table. The child table can have multiple loose foreign key definitions, therefore we store them as an array.

Example definition:

ci_pipelines:
  - table: projects
    column: project_id
    on_delete: async_delete

If the ci_pipelines key is already present in the YAML file, then a new entry can be added to the array:

ci_pipelines:
  - table: projects
    column: project_id
    on_delete: async_delete
  - table: another_table
    column: another_id
    on_delete: :async_nullify

Track record changes

To know about deletions in the projects table, configure a DELETE trigger using a post-deployment migration. The trigger needs to be configured only once. If the model already has at least one loose_foreign_key definition, then this step can be skipped:

class TrackProjectRecordChanges < Gitlab::Database::Migration[2.1]
  include Gitlab::Database::MigrationHelpers::LooseForeignKeyHelpers

  def up
    track_record_deletions(:projects)
  end

  def down
    untrack_record_deletions(:projects)
  end
end

Remove the foreign key

If there is an existing foreign key, then it can be removed from the database. This foreign key describes the link between the projects and ci_pipelines tables:

ALTER TABLE ONLY ci_pipelines
ADD CONSTRAINT fk_86635dbd80
FOREIGN KEY (project_id)
REFERENCES projects(id)
ON DELETE CASCADE;

The migration must run after the DELETE trigger is installed and the loose foreign key definition is deployed. As such, it must be a post-deployment migration dated after the migration for the trigger. If the foreign key is deleted earlier, there is a good chance of introducing data inconsistency which needs manual cleanup:

class RemoveProjectsCiPipelineFk < Gitlab::Database::Migration[2.1]
  disable_ddl_transaction!

  def up
    with_lock_retries do
      remove_foreign_key_if_exists(:ci_pipelines, :projects, name: "fk_86635dbd80")
    end
  end

  def down
    add_concurrent_foreign_key(:ci_pipelines, :projects, name: "fk_86635dbd80", column: :project_id, target_column: :id, on_delete: "cascade")
  end
end

At this point, the setup phase is concluded. The deleted projects records should be automatically picked up by the scheduled cleanup worker job.

Remove the loose foreign key

When the loose foreign key definition is no longer needed (parent table is removed, or FK is restored), we need to remove the definition from the YAML file and ensure that we don't leave pending deleted records in the database.

Remove the loose foreign key definition from the configuration (config/gitlab_loose_foreign_keys.yml).

The deletion tracking trigger needs to be removed only when the parent table no longer uses loose foreign keys. If the model still has at least one loose_foreign_key definition remaining, then these steps can be skipped:

Remove the trigger from the parent table (if the parent table is still there).
Remove leftover deleted records from the loose_foreign_keys_deleted_records table.

Migration for removing the trigger:

class UnTrackProjectRecordChanges < Gitlab::Database::Migration[2.1]
  include Gitlab::Database::MigrationHelpers::LooseForeignKeyHelpers

  def up
    untrack_record_deletions(:projects)
  end

  def down
    track_record_deletions(:projects)
  end
end

With the trigger removal, we prevent further records to be inserted in the loose_foreign_keys_deleted_records table however, there is still a chance for having leftover pending records in the table. These records must be removed with an inline data migration.

class RemoveLeftoverProjectDeletions < Gitlab::Database::Migration[2.1]
  disable_ddl_transaction!

  def up
    loop do
      result = execute <<~SQL
      DELETE FROM "loose_foreign_keys_deleted_records"
      WHERE
      ("loose_foreign_keys_deleted_records"."partition", "loose_foreign_keys_deleted_records"."id") IN (
        SELECT "loose_foreign_keys_deleted_records"."partition", "loose_foreign_keys_deleted_records"."id"
        FROM "loose_foreign_keys_deleted_records"
        WHERE
        "loose_foreign_keys_deleted_records"."fully_qualified_table_name" = 'public.projects' AND
        "loose_foreign_keys_deleted_records"."status" = 1
        LIMIT 100
      )
      SQL

      break if result.cmd_tuples == 0
    end
  end

  def down
    # no-op
  end
end

Testing

The "it has loose foreign keys" shared example can be used to test the presence of the ON DELETE trigger and the loose foreign key definitions.

Add to the model test file:

it_behaves_like 'it has loose foreign keys' do
  let(:factory_name) { :project }
end

After removing a foreign key, use the "cleanup by a loose foreign key" shared example to test a child record's deletion or nullification via the added loose foreign key:

it_behaves_like 'cleanup by a loose foreign key' do
  let!(:model) { create(:ci_pipeline, user: create(:user)) }
  let!(:parent) { model.user }
end

Caveats of loose foreign keys

Record creation

The feature provides an efficient way of cleaning up associated records after the parent record is deleted. Without foreign keys, it's the application's responsibility to validate if the parent record exists when a new associated record is created.

A bad example: record creation with the given ID (project_id comes from user input). In this example, nothing prevents us from passing a random project ID:

Ci::Pipeline.create!(project_id: params[:project_id])

A good example: record creation with extra check:

project = Project.find(params[:project_id])
Ci::Pipeline.create!(project_id: project.id)

Association lookup

Consider the following HTTP request:

GET /projects/5/pipelines/100

The controller action ignores the project_id parameter and finds the pipeline using the ID:

  def show
  # bad, avoid it
  pipeline = Ci::Pipeline.find(params[:id]) # 100
end

This endpoint still works when the parent Project model is deleted. This can be considered a a data leak which should not happen under typical circumstances:

def show
  # good
  project = Project.find(params[:project_id])
  pipeline = project.pipelines.find(params[:pipeline_id]) # 100
end

NOTE: This example is unlikely in GitLab, because we usually look up the parent models to perform permission checks.

A note on `dependent: :destroy` and `dependent: :nullify`

We considered using these Rails features as an alternative to foreign keys but there are several problems which include:

These run on a different connection in the context of a transaction which we do not allow.
These can lead to severe performance degradation as we load all records from PostgreSQL, loop over them in Ruby, and call individual DELETE queries.
These can miss data as they only cover the case when the destroy method is called directly on the model. There are other cases including delete_all and cascading deletes from another parent table that could mean these are missed.

For non-trivial objects that need to clean up data outside the database (for example, object storage) where you might wish to use dependent: :destroy, see alternatives in Avoid dependent: :nullify and dependent: :destroy across databases.

Update target column to a value

A loose foreign key might be used to update a target column to a value when an entry in parent table is deleted.

It's important to add an index (if it doesn't exist yet) on (column, target_column) to avoid any performance issues. Any index starting with these two columns will work.

The configuration requires additional information:

Column to be updated (target_column)
Value to be set in the target column (target_value)

Example definition:

packages:
  - table: projects
    column: project_id
    on_delete: update_column_to
    target_column: status
    target_value: 4

Risks of loose foreign keys and possible mitigations

In general, the loose foreign keys architecture is eventually consistent and the cleanup latency might lead to problems visible to GitLab users or operators. We consider the tradeoff as acceptable, but there might be cases where the problems are too frequent or too severe, and we must implement a mitigation strategy. A general mitigation strategy might be to have an "urgent" queue for cleanup of records that have higher impact with a delayed cleanup.

Below are some more specific examples of problems that might occur and how we might mitigate them. In all the listed cases we might still consider the problem described to be low risk and low impact, and in that case we would choose to not implement any mitigation.

The record should be deleted but it shows up in a view

This hypothetical example might happen with a foreign key like:

ALTER TABLE ONLY vulnerability_occurrence_pipelines
    ADD CONSTRAINT fk_rails_6421e35d7d FOREIGN KEY (pipeline_id) REFERENCES ci_pipelines(id) ON DELETE CASCADE;

In this example we expect to delete all associated vulnerability_occurrence_pipelines records whenever we delete the ci_pipelines record associated with them. In this case you might end up with some vulnerability page in GitLab which shows an occurrence of a vulnerability. However, when you try to select a link to the pipeline, you get a 404, because the pipeline is deleted. Then, when you navigate back you might find the occurrence has disappeared too.

Mitigation

When rendering the vulnerability occurrences on the vulnerability page we could try to load the corresponding pipeline and choose to skip displaying that occurrence if pipeline is not found.

The deleted parent record is needed to render a view and causes a `500` error

This hypothetical example might happen with a foreign key like:

ALTER TABLE ONLY vulnerability_occurrence_pipelines
    ADD CONSTRAINT fk_rails_6421e35d7d FOREIGN KEY (pipeline_id) REFERENCES ci_pipelines(id) ON DELETE CASCADE;

In this example we expect to delete all associated vulnerability_occurrence_pipelines records whenever we delete the ci_pipelines record associated with them. In this case you might end up with a vulnerability page in GitLab which shows an "occurrence" of a vulnerability. However, when rendering the occurrence we try to load, for example, occurrence.pipeline.created_at, which causes a 500 for the user.

Mitigation

When rendering the vulnerability occurrences on the vulnerability page we could try to load the corresponding pipeline and choose to skip displaying that occurrence if pipeline is not found.

The deleted parent record is accessed in a Sidekiq worker and causes a failed job

This hypothetical example might happen with a foreign key like:

ALTER TABLE ONLY vulnerability_occurrence_pipelines
    ADD CONSTRAINT fk_rails_6421e35d7d FOREIGN KEY (pipeline_id) REFERENCES ci_pipelines(id) ON DELETE CASCADE;

In this example we expect to delete all associated vulnerability_occurrence_pipelines records whenever we delete the ci_pipelines record associated with them. In this case you might end up with a Sidekiq worker that is responsible for processing a vulnerability and looping over all occurrences causing a Sidekiq job to fail if it executes occurrence.pipeline.created_at.

Mitigation

When looping through the vulnerability occurrences in the Sidekiq worker, we could try to load the corresponding pipeline and choose to skip processing that occurrence if pipeline is not found.

Architecture

The loose foreign keys feature is implemented within the LooseForeignKeys Ruby namespace. The code is isolated from the core application code and theoretically, it could be a standalone library.

The feature is invoked solely in the LooseForeignKeys::CleanupWorker worker class. The worker is scheduled via a cron job where the schedule depends on the configuration of the GitLab instance.

Non-decomposed GitLab (1 database): invoked every minute.
Decomposed GitLab (2 databases, CI and Main): invoked every minute, cleaning up one database at a time. For example, the cleanup worker for the main database runs every two minutes.

To avoid lock contention and the processing of the same database rows, the worker does not run parallel. This behavior is ensured with a Redis lock.

Record cleanup procedure:

Acquire the Redis lock.
Determine which database to clean up.
Collect all database tables where the deletions are tracked (parent tables).
- This is achieved by reading the config/gitlab_loose_foreign_keys.yml file.
- A table is considered "tracked" when a loose foreign key definition exists for the table and the DELETE trigger is installed.
Cycle through the tables with an infinite loop.
For each table, load a batch of deleted parent records to clean up.
Depending on the YAML configuration, build DELETE or UPDATE (nullify) queries for the referenced child tables.
Invoke the queries.
Repeat until all child records are cleaned up or the maximum limit is reached.
Remove the deleted parent records when all child records are cleaned up.

Database structure

The feature relies on triggers installed on the parent tables. When a parent record is deleted, the trigger automatically inserts a new record into the loose_foreign_keys_deleted_records database table.

The inserted record stores the following information about the deleted record:

fully_qualified_table_name: name of the database table where the record was located.
primary_key_value: the ID of the record, the value is present in the child tables as the foreign key value. At the moment, composite primary keys are not supported, the parent table must have an id column.
status: defaults to pending, represents the status of the cleanup process.
consume_after: defaults to the current time.
cleanup_attempts: defaults to 0. The number of times the worker tried to clean up this record. A non-zero number would mean that this record has many child records and cleaning it up requires several runs.

Database decomposition

The loose_foreign_keys_deleted_records table exists on both database servers (ci and main) after the database decomposition. The worker ill determine which parent tables belong to which database by reading the lib/gitlab/database/gitlab_schemas.yml YAML file.

Example:

Main database tables
- projects
- namespaces
- merge_requests
Ci database tables
- ci_builds
- ci_pipelines

When the worker is invoked for the ci database, the worker loads deleted records only from the ci_builds and ci_pipelines tables. During the cleanup process, DELETE and UPDATE queries mostly run on tables located in the Main database. In this example, one UPDATE query nullifies the merge_requests.head_pipeline_id column.

Database partitioning

Due to the large volume of inserts the database table receives daily, a special partitioning strategy was implemented to address data bloat concerns. Originally, the time-decay strategy was considered for the feature but due to the large data volume we decided to implement a new strategy.

A deleted record is considered fully processed when all its direct children records have been cleaned up. When this happens, the loose foreign key worker updates the status column of the deleted record. After this step, the record is no longer needed.

The sliding partitioning strategy provides an efficient way of cleaning up old, unused data by adding a new database partition and removing the old one when certain conditions are met. The loose_foreign_keys_deleted_records database table is list partitioned where most of the time there is only one partition attached to the table.

                                                             Partitioned table "public.loose_foreign_keys_deleted_records"
           Column           |           Type           | Collation | Nullable |                            Default                             | Storage  | Stats target | Description
----------------------------+--------------------------+-----------+----------+----------------------------------------------------------------+----------+--------------+-------------
 id                         | bigint                   |           | not null | nextval('loose_foreign_keys_deleted_records_id_seq'::regclass) | plain    |              |
 partition                  | bigint                   |           | not null | 84                                                             | plain    |              |
 primary_key_value          | bigint                   |           | not null |                                                                | plain    |              |
 status                     | smallint                 |           | not null | 1                                                              | plain    |              |
 created_at                 | timestamp with time zone |           | not null | now()                                                          | plain    |              |
 fully_qualified_table_name | text                     |           | not null |                                                                | extended |              |
 consume_after              | timestamp with time zone |           |          | now()                                                          | plain    |              |
 cleanup_attempts           | smallint                 |           |          | 0                                                              | plain    |              |
Partition key: LIST (partition)
Indexes:
    "loose_foreign_keys_deleted_records_pkey" PRIMARY KEY, btree (partition, id)
    "index_loose_foreign_keys_deleted_records_for_partitioned_query" btree (partition, fully_qualified_table_name, consume_after, id) WHERE status = 1
Check constraints:
    "check_1a541f3235" CHECK (char_length(fully_qualified_table_name) <= 150)
Partitions: gitlab_partitions_dynamic.loose_foreign_keys_deleted_records_84 FOR VALUES IN ('84')

The partition column controls the insert direction, the partition value determines which partition gets the deleted rows inserted via the trigger. Notice that the default value of the partition table matches with the value of the list partition (84). In INSERT query within the trigger the value of the partition is omitted, the trigger always relies on the default value of the column.

Example INSERT query for the trigger:

INSERT INTO loose_foreign_keys_deleted_records
(fully_qualified_table_name, primary_key_value)
SELECT TG_TABLE_SCHEMA || '.' || TG_TABLE_NAME, old_table.id FROM old_table;

The partition "sliding" process is controlled by two, regularly executed callbacks. These callbacks are defined within the LooseForeignKeys::DeletedRecord model.

The next_partition_if callback controls when to create a new partition. A new partition is created when the current partition has at least one record older than 24 hours. A new partition is added by the PartitionManager using the following steps:

Create a new partition, where the VALUE for the partition is CURRENT_PARTITION + 1.
Update the default value of the partition column to CURRENT_PARTITION + 1.

With these steps, all new INSERT queries via the triggers end up in the new partition. At this point, the database table has two partitions.

The detach_partition_if callback determines if the old partitions can be detached from the table. A partition is detachable if there are no pending (unprocessed) records in the partition (status = 1). The detached partitions are available for some time, you can see the list detached partitions in the detached_partitions table:

select * from detached_partitions;

Cleanup queries

The LooseForeignKeys::CleanupWorker has its database query builder which depends on Arel. The feature doesn't reference any application-specific ActiveRecord models to avoid unexpected side effects. The database queries are batched, which means that several parent records are being cleaned up at the same time.

Example DELETE query:

DELETE
FROM "merge_request_metrics"
WHERE ("merge_request_metrics"."id") IN
  (SELECT "merge_request_metrics"."id"
    FROM "merge_request_metrics"
    WHERE "merge_request_metrics"."pipeline_id" IN (1, 2, 10, 20)
    LIMIT 1000 FOR UPDATE SKIP LOCKED)

The primary key values of the parent records are 1, 2, 10, and 20.

Example UPDATE (nullify) query:

UPDATE "merge_requests"
SET "head_pipeline_id" = NULL
WHERE ("merge_requests"."id") IN
    (SELECT "merge_requests"."id"
     FROM "merge_requests"
     WHERE "merge_requests"."head_pipeline_id" IN (3, 4, 30, 40)
     LIMIT 500 FOR UPDATE SKIP LOCKED)

These queries are batched, which means that in many cases, several invocations are needed to clean up all associated child records.

The batching is implemented with loops, the processing stops when all associated child records are cleaned up or the limit is reached.

loop do
  modification_count = process_batch_with_skip_locked

  break if modification_count == 0 || over_limit?
end

loop do
  modification_count = process_batch

  break if modification_count == 0 || over_limit?
end

The loop-based batch processing is preferred over EachBatch for the following reasons:

The records in the batch are modified, so the next batch contains different records.
There is always an index on the foreign key column however, the column is usually not unique. EachBatch requires a unique column for the iteration.
The record order doesn't matter for the cleanup.

Notice that we have two loops. The initial loop processes records with the SKIP LOCKED clause. The query skips rows that are locked by other application processes. This ensures that the cleanup worker is less likely to become blocked. The second loop executes the database queries without SKIP LOCKED to ensure that all records have been processed.

Processing limits

A constant, large volume of record updates or deletions can cause incidents and affect the availability of GitLab:

Increased table bloat.
Increased number of pending WAL files.
Busy tables, difficulty when acquiring locks.

To mitigate these issues, several limits are applied when the worker runs.

Each query has LIMIT, a query cannot process an unbounded number of rows.
The maximum number of record deletions and record updates is limited.
The maximum runtime (30 seconds) for the database queries is limited.

The limit rules are implemented in the LooseForeignKeys::ModificationTracker class. When one of the limits (record modification count, time limit) is reached the processing is stopped immediately. After some time, the next scheduled worker continues the cleanup process.

Performance characteristics

The database trigger on the parent tables decreases the record deletion speed. Each statement that removes rows from the parent table invokes the trigger to insert records into the loose_foreign_keys_deleted_records table.

The queries within the cleanup worker are fairly efficient index scans, with limits in place they're unlikely to affect other parts of the application.

The database queries are not running in transaction, when an error happens for example a statement timeout or a worker crash, the next job continues the processing.

Troubleshooting

Accumulation of deleted records

There can be cases where the workers need to process an unusually large amount of data. This can happen under typical usage, for example when a large project or group is deleted. In this scenario, there can be several million rows to be deleted or nullified. Due to the limits enforced by the worker, processing this data takes some time.

When cleaning up "heavy-hitters", the feature ensures fair processing by rescheduling larger batches for later. This gives time for other deleted records to be processed.

For example, a project with millions of ci_builds records is deleted. The ci_builds records is deleted by the loose foreign keys feature.

The cleanup worker is scheduled and picks up a batch of deleted projects records. The large project is part of the batch.
Deletion of the orphaned ci_builds rows has started.
The time limit is reached, but the cleanup is not complete.
The cleanup_attempts column is incremented for the deleted records.
Go to step 1. The next cleanup worker continues the cleanup.
When the cleanup_attempts reaches 3, the batch is re-scheduled 10 minutes later by updating the consume_after column.
The next cleanup worker processes a different batch.

We have Prometheus metrics in place to monitor the deleted record cleanup:

loose_foreign_key_processed_deleted_records: Number of processed deleted records. When large cleanup happens, this number would decrease.
loose_foreign_key_incremented_deleted_records: Number of deleted records which were not finished processing. The cleanup_attempts column was incremented.
loose_foreign_key_rescheduled_deleted_records: Number of deleted records that had to be rescheduled at a later time after 3 cleanup attempts.

Example PromQL query:

loose_foreign_key_rescheduled_deleted_records{env="gprd", table="ci_runners"}

Another way to look at the situation is by running a database query. This query gives the exact counts of the unprocessed records:

SELECT partition, fully_qualified_table_name, count(*)
FROM loose_foreign_keys_deleted_records
WHERE
status = 1
GROUP BY 1, 2;

Example output:

 partition | fully_qualified_table_name | count
-----------+----------------------------+-------
        87 | public.ci_builds           |   874
        87 | public.ci_job_artifacts    |  6658
        87 | public.ci_pipelines        |   102
        87 | public.ci_runners          |   111
        87 | public.merge_requests      |   255
        87 | public.namespaces          |    25
        87 | public.projects            |     6

The query includes the partition number which can be useful to detect if the cleanup process is significantly lagging behind. When multiple different partition values are present in the list that means the cleanup of some deleted records didn't finish in several days (1 new partition is added every day).

Steps to diagnose the problem:

Check which records are accumulating.
Try to get an estimate of the number of remaining records.
Looking into the worker performance stats (Kibana or Grafana).

Possible solutions:

Short-term: increase the batch sizes.
Long-term: invoke the worker more frequently. Parallelize the worker

For a one-time fix, we can run the cleanup worker several times from the rails console. The worker can run in parallel however, this can introduce lock contention and it could increase the worker runtime.

LooseForeignKeys::CleanupWorker.new.perform

When the cleanup is done, the older partitions are automatically detached by the PartitionManager.

PartitionManager bug

NOTE: This issue happened in the past on Staging and it has been mitigated.

When adding a new partition, the default value of the partition column is also updated. This is a schema change that is executed in the same transaction as the new partition creation. It's highly unlikely that the partition column goes outdated.

However, if this happens then this can cause application-wide incidents because the partition value points to a partition that doesn't exist. Symptom: deletion of records from tables where the DELETE trigger is installed fails.

\d+ loose_foreign_keys_deleted_records;

           Column           |           Type           | Collation | Nullable |                            Default                             | Storage  | Stats target | Description
----------------------------+--------------------------+-----------+----------+----------------------------------------------------------------+----------+--------------+-------------
 id                         | bigint                   |           | not null | nextval('loose_foreign_keys_deleted_records_id_seq'::regclass) | plain    |              |
 partition                  | bigint                   |           | not null | 4                                                              | plain    |              |
 primary_key_value          | bigint                   |           | not null |                                                                | plain    |              |
 status                     | smallint                 |           | not null | 1                                                              | plain    |              |
 created_at                 | timestamp with time zone |           | not null | now()                                                          | plain    |              |
 fully_qualified_table_name | text                     |           | not null |                                                                | extended |              |
 consume_after              | timestamp with time zone |           |          | now()                                                          | plain    |              |
 cleanup_attempts           | smallint                 |           |          | 0                                                              | plain    |              |
Partition key: LIST (partition)
Indexes:
    "loose_foreign_keys_deleted_records_pkey" PRIMARY KEY, btree (partition, id)
    "index_loose_foreign_keys_deleted_records_for_partitioned_query" btree (partition, fully_qualified_table_name, consume_after, id) WHERE status = 1
Check constraints:
    "check_1a541f3235" CHECK (char_length(fully_qualified_table_name) <= 150)
Partitions: gitlab_partitions_dynamic.loose_foreign_keys_deleted_records_3 FOR VALUES IN ('3')

Check the default value of the partition column and compare it with the available partitions (4 vs 3). The partition with the value of 4 does not exist. To mitigate the problem an emergency schema change is required:

ALTER TABLE loose_foreign_keys_deleted_records ALTER COLUMN partition SET DEFAULT 3;

Loose foreign keys

Problem statement

Asynchronous approach

How it works

The scripts/decomposition/generate-loose-foreign-key

Example migration and configuration

Configure the loose foreign key

Track record changes

Remove the foreign key

Remove the loose foreign key

Testing

Caveats of loose foreign keys

Record creation

Association lookup

A note on dependent: :destroy and dependent: :nullify

Update target column to a value

Risks of loose foreign keys and possible mitigations

The record should be deleted but it shows up in a view

The deleted parent record is needed to render a view and causes a 500 error

The deleted parent record is accessed in a Sidekiq worker and causes a failed job

Architecture

Database structure

Database decomposition

Database partitioning

Cleanup queries

Processing limits

Performance characteristics

Troubleshooting

Accumulation of deleted records

PartitionManager bug

The `scripts/decomposition/generate-loose-foreign-key`

A note on `dependent: :destroy` and `dependent: :nullify`

The deleted parent record is needed to render a view and causes a `500` error