Database table partitioning

Table partitioning is a powerful database feature that allows a table's data to be split into smaller physical tables that act as a single large table. If the application is designed to work with partitioning in mind, there can be multiple benefits, such as:

  • Query performance can be improved greatly, because the database can cheaply eliminate much of the data from the search space, while still providing full SQL capabilities.

  • Bulk deletes can be achieved with minimal impact on the database by dropping entire partitions. This is a natural fit for features that need to periodically delete data that falls outside the retention window.

  • Administrative tasks like VACUUM and index rebuilds can operate on individual partitions, rather than across a single massive table.

Unfortunately, not all models fit a partitioning scheme, and there are significant drawbacks if implemented incorrectly. Additionally, tables can only be partitioned at their creation, making it nontrivial to apply partitioning to a busy database. A suite of migration tools are available to enable backend developers to partition existing tables, but the migration process is rather heavy, taking multiple steps split across several releases. Due to the limitations of partitioning and the related migrations, you should understand how partitioning fits your use case before attempting to leverage this feature.

Determining when to use partitioning

While partitioning can be very useful when properly applied, it's imperative to identify if the data and workload of a table naturally fit a partitioning scheme. There are a few details you have to understand to decide if partitioning is a good fit for your particular problem.

First, a table is partitioned on a partition key, which is a column or set of columns which determine how the data is split across the partitions. The partition key is used by the database when reading or writing data, to decide which partitions must be accessed. The partition key should be a column that would be included in a WHERE clause on almost all queries accessing that table.

Second, it's necessary to understand the strategy the database uses to split the data across the partitions. The scheme supported by the GitLab migration helpers is date-range partitioning, where each partition in the table contains data for a single month. In this case, the partitioning key must be a timestamp or date column. In order for this type of partitioning to work well, most queries must access data in a certain date range.

For a more concrete example, the audit_events table can be used, which was the first table to be partitioned in the application database (scheduled for deployment with the GitLab 13.5 release). This table tracks audit entries of security events that happen in the application. In almost all cases, users want to see audit activity that occurs in a certain time frame. As a result, date-range partitioning was a natural fit for how the data would be accessed.

To look at this in more detail, imagine a simplified audit_events schema:

CREATE TABLE audit_events (
  id SERIAL NOT NULL PRIMARY KEY,
  author_id INT NOT NULL,
  details jsonb NOT NULL,
  created_at timestamptz NOT NULL);

Now imagine typical queries in the UI would display the data in a certain date range, like a single week:

SELECT *
FROM audit_events
WHERE created_at >= '2020-01-01 00:00:00'
  AND created_at < '2020-01-08 00:00:00'
ORDER BY created_at DESC
LIMIT 100

If the table is partitioned on the created_at column the base table would look like:

CREATE TABLE audit_events (
  id SERIAL NOT NULL,
  author_id INT NOT NULL,
  details jsonb NOT NULL,
  created_at timestamptz NOT NULL,
  PRIMARY KEY (id, created_at))
PARTITION BY RANGE(created_at);

NOTE: The primary key of a partitioned table must include the partition key as part of the primary key definition.

And we might have a list of partitions for the table, such as:

audit_events_202001 FOR VALUES FROM ('2020-01-01') TO ('2020-02-01')
audit_events_202002 FOR VALUES FROM ('2020-02-01') TO ('2020-03-01')
audit_events_202003 FOR VALUES FROM ('2020-03-01') TO ('2020-04-01')

Each partition is a separate physical table, with the same structure as the base audit_events table, but contains only data for rows where the partition key falls in the specified range. For example, the partition audit_events_202001 contains rows where the created_at column is greater than or equal to 2020-01-01 and less than 2020-02-01.

Now, if we look at the previous example query again, the database can use the WHERE to recognize that all matching rows are in the audit_events_202001 partition. Rather than searching all of the data in all of the partitions, it can search only the single month's worth of data in the appropriate partition. In a large table, this can dramatically reduce the amount of data the database needs to access. However, imagine a query that does not filter based on the partitioning key, such as:

SELECT *
FROM audit_events
WHERE author_id = 123
ORDER BY created_at DESC
LIMIT 100

In this example, the database can't prune any partitions from the search, because matching data could exist in any of them. As a result, it has to query each partition individually, and aggregate the rows into a single result set. Because author_id would be indexed, the performance impact could likely be acceptable, but on more complex queries the overhead can be substantial. Partitioning should only be leveraged if the access patterns of the data support the partitioning strategy, otherwise performance suffers.

Partitioning a table

Unfortunately, tables can only be partitioned at their creation, making it nontrivial to apply to a busy database. A suite of migration tools have been developed to enable backend developers to partition existing tables. This migration process takes multiple steps which must be split across several releases.

Caveats

The partitioning migration helpers work by creating a partitioned duplicate of the original table and using a combination of a trigger and a background migration to copy data into the new table. Changes to the original table schema can be made in parallel with the partitioning migration, but they must take care to not break the underlying mechanism that makes the migration work. For example, if a column is added to the table that is being partitioned, both the partitioned table and the trigger definition must be updated to match.

Step 1: Creating the partitioned copy (Release N)

The first step is to add a migration to create the partitioned copy of the original table. This migration creates the appropriate partitions based on the data in the original table, and install a trigger that syncs writes from the original table into the partitioned copy.

An example migration of partitioning the audit_events table by its created_at column would look like:

class PartitionAuditEvents < Gitlab::Database::Migration[1.0]
  include Gitlab::Database::PartitioningMigrationHelpers

  def up
    partition_table_by_date :audit_events, :created_at
  end

  def down
    drop_partitioned_table_for :audit_events
  end
end

After this has executed, any inserts, updates, or deletes in the original table are also duplicated in the new table. For updates and deletes, the operation only has an effect if the corresponding row exists in the partitioned table.

Step 2: Backfill the partitioned copy (Release N)

The second step is to add a post-deployment migration that schedules the background jobs that backfill existing data from the original table into the partitioned copy.

Continuing the above example, the migration would look like:

class BackfillPartitionAuditEvents < Gitlab::Database::Migration[1.0]
  include Gitlab::Database::PartitioningMigrationHelpers

  def up
    enqueue_partitioning_data_migration :audit_events
  end

  def down
    cleanup_partitioning_data_migration :audit_events
  end
end

This step uses the same mechanism as any background migration, so you may want to read the Background Migration guide for details on that process. Background jobs are scheduled every 2 minutes and copy 50_000 records at a time, which can be used to estimate the timing of the background migration portion of the partitioning migration.

Step 3: Post-backfill cleanup (Release N+1)

The third step must occur at least one release after the release that includes the background migration. This gives time for the background migration to execute properly in self-managed installations. In this step, add another post-deployment migration that cleans up after the background migration. This includes forcing any remaining jobs to execute, and copying data that may have been missed, due to dropped or failed jobs.

Once again, continuing the example, this migration would look like:

class CleanupPartitionedAuditEventsBackfill < Gitlab::Database::Migration[1.0]
  include Gitlab::Database::PartitioningMigrationHelpers

  def up
    finalize_backfilling_partitioned_table :audit_events
  end

  def down
    # no op
  end
end

After this migration has completed, the original table and partitioned table should contain identical data. The trigger installed on the original table guarantees that the data remains in sync going forward.

Step 4: Swap the partitioned and non-partitioned tables (Release N+1)

The final step of the migration makes the partitioned table ready for use by the application. This section will be updated when the migration helper is ready, for now development can be followed in the Tracking Issue.