Please check gitlab-tutorial

Skip to content

Merge request analytics with ClickHouse

The merge request analytics feature shows statistics about the merged merge requests in the project and also exposes record-level metadata. Aggregations include:

  • Average time to merge: The duration between the creation time and the merge time.
  • Monthly aggregations: A chart of 12 months of the merged merge requests.

Under the chart, the user can see the paginated list of merge requests, 12 months per page.

You can filter by:

  • Author
  • Assignee
  • Labels
  • Milestone
  • Source branch
  • Target branch

Current performance problems

  • The aggregation queries require specialized indexes, which cost additional disk space (index-only scans).
  • Querying the whole 12 months is slow (statement timeout). Instead, the frontend requests data per month (12 database queries).
  • Even with specialized indexes, making the feature available on the group level would not be feasible due to the large volume of merge requests.

Example queries

Get the number of merge requests merged in a given month:

SELECT COUNT(*)
FROM "merge_requests"
INNER JOIN "merge_request_metrics" ON "merge_request_metrics"."merge_request_id" = "merge_requests"."id"
WHERE (NOT EXISTS
         (SELECT 1
          FROM "banned_users"
          WHERE (merge_requests.author_id = banned_users.user_id)))
  AND "merge_request_metrics"."target_project_id" = 278964
  AND "merge_request_metrics"."merged_at" >= '2022-12-01 00:00:00'
  AND "merge_request_metrics"."merged_at" <= '2023-01-01 00:00:00'

The merge_request_metrics table was de-normalized (by adding target_project_id) to improve the first-page load time. The query itself works well for smaller date ranges, however, it can time out as the date range increases.

After an extra filter is added, the query becomes more complex because it must also filter the merge_requests table:

SELECT COUNT(*)
FROM "merge_requests"
INNER JOIN "merge_request_metrics" ON "merge_request_metrics"."merge_request_id" = "merge_requests"."id"
WHERE (NOT EXISTS
         (SELECT 1
          FROM "banned_users"
          WHERE (merge_requests.author_id = banned_users.user_id)))
  AND "merge_requests"."author_id" IN
    (SELECT "users"."id"
     FROM "users"
     WHERE (LOWER("users"."username") IN (LOWER('ahegyi'))))
  AND "merge_request_metrics"."target_project_id" = 278964
  AND "merge_request_metrics"."merged_at" >= '2022-12-01 00:00:00'
  AND "merge_request_metrics"."merged_at" <= '2023-01-01 00:00:00'

To calculate mean time to merge, we also query the total time between the merge request creation time and merge time.

SELECT EXTRACT(epoch
               FROM SUM(AGE(merge_request_metrics.merged_at, merge_request_metrics.created_at)))
FROM "merge_requests"
INNER JOIN "merge_request_metrics" ON "merge_request_metrics"."merge_request_id" = "merge_requests"."id"
WHERE (NOT EXISTS
         (SELECT 1
          FROM "banned_users"
          WHERE (merge_requests.author_id = banned_users.user_id)))
  AND "merge_requests"."author_id" IN
    (SELECT "users"."id"
     FROM "users"
     WHERE (LOWER("users"."username") IN (LOWER('ahegyi'))))
  AND "merge_request_metrics"."target_project_id" = 278964
  AND "merge_request_metrics"."merged_at" >= '2022-08-01 00:00:00'
  AND "merge_request_metrics"."merged_at" <= '2022-09-01 00:00:00'
  AND "merge_request_metrics"."merged_at" > "merge_request_metrics"."created_at"
LIMIT 1

Store merge request data in ClickHouse

Several other use cases exist for storing and querying merge request data in ClickHouse. In this document, we focus on this particular feature.

The core data exists in the merge_request_metrics and in the merge_requests database tables. Some filters require extra tables to be joined:

  • banned_users: Filter out merge requests created by banned users.
  • labels: A merge request can have one or more assigned labels.
  • assignees: A merge request can have one or more assignees.
  • merged_at: The merged_at column is located in the merge_request_metrics table.

The merge_requests table contains data that can be filtered directly:

  • Author: via the author_id column.
  • Milestone: via the milestone_id column.
  • Source branch.
  • Target branch.
  • Project: via the project_id column.

Keep ClickHouse data up to date

Replicating or syncing the merge_requests table is unfortunately not enough. Separate queries to associated tables are required to insert one de-normalized merge_requests row into the ClickHouse database.

Change detection is non-trivial to implement. A few corners we could cut:

  • The feature is available for GitLab Premium and GitLab Ultimate customers. We don't have to sync all the data, but instead sync only the merge_requests records which are part of licensed groups.
  • Data changes (often) happen via the MergeRequest services, where bumping the updated_at timestamp column is mostly consistent. Some sort of incremental synchronization process could be implemented.
  • We only need to query the merged merge requests. After the merge, the record rarely changes.

Database table structure

The database table structure uses de-normalization to make all required columns available in one database table. This eliminates the need for JOINs.

CREATE TABLE merge_requests
(
    `id` UInt64,
    `project_id` UInt64 DEFAULT 0 NOT NULL,
    `author_id` UInt64 DEFAULT 0 NOT NULL,
    `milestone_id` UInt64 DEFAULT 0 NOT NULL,
    `label_ids` Array(UInt64) DEFAULT [] NOT NULL,
    `assignee_ids` Array(UInt64) DEFAULT [] NOT NULL,
    `source_branch` String DEFAULT '' NOT NULL,
    `target_branch` String DEFAULT '' NOT NULL,
    `merged_at` DateTime64(6, 'UTC') NOT NULL,
    `created_at` DateTime64(6, 'UTC') DEFAULT now() NOT NULL,
    `updated_at` DateTime64(6, 'UTC') DEFAULT now() NOT NULL
)
ENGINE = ReplacingMergeTree(updated_at)
ORDER BY (project_id, merged_at, id);

Similarly to the activity data example, we use the ReplacingMergeTree engine. Several columns of the merge request record may change, so keeping the table up-to-date is important.

The database table is ordered by the project_id, merged_at, id columns. This ordering optimizes the table data for our use case: querying the merged_at column in a project.

Rewrite the count query

First, let's generate some data for the table.

INSERT INTO merge_requests (id, project_id, author_id, milestone_id, label_ids, merged_at, created_at)
SELECT id, project_id, author_id, milestone_id, label_ids, merged_at, created_at
FROM generateRandom('id UInt64, project_id UInt8, author_id UInt8, milestone_id UInt8, label_ids Array(UInt8), merged_at DateTime64(6, \'UTC\'), created_at DateTime64(6, \'UTC\')')
LIMIT 1000000;

NOTE: Some integer data types were cast as UInt8 so it is highly probable that they have same values across different rows.

The original count query only aggregated data for one month. With ClickHouse, we can attempt aggregating the data for the whole year.

PostgreSQL-based count query:

SELECT COUNT(*)
FROM "merge_requests"
INNER JOIN "merge_request_metrics" ON "merge_request_metrics"."merge_request_id" = "merge_requests"."id"
WHERE (NOT EXISTS
         (SELECT 1
          FROM "banned_users"
          WHERE (merge_requests.author_id = banned_users.user_id)))
  AND "merge_request_metrics"."target_project_id" = 278964
  AND "merge_request_metrics"."merged_at" >= '2022-12-01 00:00:00'
  AND "merge_request_metrics"."merged_at" <= '2023-01-01 00:00:00'

ClickHouse query:

SELECT
  toYear(merged_at) AS year,
  toMonth(merged_at) AS month,
  COUNT(*)
FROM merge_requests
WHERE
  project_id = 200
  AND merged_at BETWEEN '2022-01-01 00:00:00'
  AND '2023-01-01 00:00:00'
GROUP BY year, month

The query processed a significantly lower number of rows compared to the generated data. The ORDER BY clause (primary key) is helping the query execution:

11 rows in set. Elapsed: 0.010 sec.
Processed 8.19 thousand rows, 131.07 KB (783.45 thousand rows/s., 12.54 MB/s.)

Rewrite the Mean time to merge query

The query calculates the mean time to merge as: duration(created_at, merged_at) / merge_request_count. The calculation is done in two separate steps:

  1. Request the monthly counts and the monthly duration values.
  2. Sum the counts to get the yearly count.
  3. Sum the durations to get the yearly duration.
  4. Divide the durations by the count.

In ClickHouse, we can calculate the mean time to merge with one query:

SELECT
  SUM(
    dateDiff('second', merged_at, created_at) / 3600 / 24
  ) / COUNT(*) AS mean_time_to_merge -- mean_time_to_merge is in days
FROM merge_requests
WHERE
  project_id = 200
  AND merged_at BETWEEN '2022-01-01 00:00:00'
  AND '2023-01-01 00:00:00'

Filtering

The database queries above can be used as base queries. You can add more filters. For example, filtering for a label and a milestone:

SELECT
  toYear(merged_at) AS year,
  toMonth(merged_at) AS month,
  COUNT(*)
FROM merge_requests
WHERE
  project_id = 200
  AND milestone_id = 15
  AND has(label_ids, 118)
  AND -- array includes 118
  merged_at BETWEEN '2022-01-01 00:00:00'
  AND '2023-01-01 00:00:00'
GROUP BY year, month

Optimizing a particular filter is usually done with a database index. This particular query reads 8000 rows:

1 row in set. Elapsed: 0.016 sec.
Processed 8.19 thousand rows, 589.99 KB (505.38 thousand rows/s., 36.40 MB/s.)

Adding an index on milestone_id:

ALTER TABLE merge_requests
ADD
  INDEX milestone_id_index milestone_id TYPE minmax GRANULARITY 10;
ALTER TABLE
  merge_requests MATERIALIZE INDEX milestone_id_index;

On the generated data, adding the index didn't improve the performance.

Banned users filter

A recently added feature in GitLab filters out merge requests where the author is banned by the admins. The banned users are tracked on the instance level in the banned_users database table.

Idea 1: Enumerate the banned user IDs

This would require no structural changes to the ClickHouse database schema. We could query the banned users in the project and filter the values out in query time.

Get the banned users (in PostgreSQL):

SELECT user_id FROM banned_users

In ClickHouse

SELECT
  toYear(merged_at) AS year,
  toMonth(merged_at) AS month,
  COUNT(*)
FROM merge_requests
WHERE
  author_id NOT IN (1, 2, 3, 4) AND -- banned users
  project_id = 200
  AND milestone_id = 15
  AND has(label_ids, 118) AND -- array includes 118
  merged_at BETWEEN '2022-01-01 00:00:00'
  AND '2023-01-01 00:00:00'
GROUP BY year, month

The problem with this approach is that the number of banned users could increase significantly which would make the query bigger and slower.

Idea 2: replicate the banned_users table

Assuming that the banned_users table doesn't grow to millions of rows, we could attempt to periodically sync the whole table to ClickHouse. With this approach, a mostly consistent banned_users table could be used in the ClickHouse database query:

SELECT
  toYear(merged_at) AS year,
  toMonth(merged_at) AS month,
  COUNT(*)
FROM merge_requests
WHERE
  author_id NOT IN (SELECT user_id FROM banned_users) AND
  project_id = 200 AND
  milestone_id = 15 AND
  has(label_ids, 118) AND -- array includes 118
  merged_at BETWEEN '2022-01-01 00:00:00' AND '2023-01-01 00:00:00'
GROUP BY year, month

Alternatively, the banned_users table could be stored as a dictionary to further improve the query performance.

Idea 3: Alter the feature

For analytical calculations, it might be acceptable to drop this particular filter. This approach assumes that including the merge requests of banned users doesn't skew the statistics significantly.