GitHub importer developer documentation

The GitHub importer offers two different types of importers:

  • A sequential importer. Used by the import:github Rake task.
  • A parallel importer. Used by everything else.

The difference between these two importers is:

  • The sequential importer does all the work in a single thread, so it’s more suited for debugging purposes or Rake tasks.
  • The parallel importer uses Sidekiq.

Prerequisites

  • Sidekiq workers that process the github_importer and github_importer_advance_stage queues (enabled by default).
  • Octokit (used for interacting with the GitHub API).

Code structure

The importer’s codebase is broken up into the following directories:

  • lib/gitlab/github_import: this directory contains most of the code such as the classes used for importing resources.
  • app/workers/gitlab/github_import: this directory contains the Sidekiq workers.
  • app/workers/concerns/gitlab/github_import: this directory contains a few modules reused by the various Sidekiq workers.

Architecture overview

When a GitHub project is imported, work is divided into separate stages, with each stage consisting of a set of Sidekiq jobs that are executed. Between every stage a job is scheduled that periodically checks if all work of the current stage is completed, advancing the import process to the next stage when this is the case. The worker handling this is called Gitlab::GithubImport::AdvanceStageWorker.

Stages

1. RepositoryImportWorker

This worker calls Projects::ImportService.new.execute, which calls importer.execute.

In this context, importer is an instance of Gitlab::ImportSources.importer(project.import_type), which for github import types maps to ParallelImporter.

ParallelImporter schedules a job for the next worker.

2. Stage::ImportRepositoryWorker

This worker imports the repository and wiki, scheduling the next stage when done.

3. Stage::ImportBaseDataWorker

This worker imports base data such as labels, milestones, and releases. This work is done in a single thread because it can be performed fast enough that we don’t need to perform this work in parallel.

4. Stage::ImportPullRequestsWorker

This worker imports all pull requests. For every pull request a job for the Gitlab::GithubImport::ImportPullRequestWorker worker is scheduled.

5. Stage::ImportCollaboratorsWorker

This worker imports only direct repository collaborators who are not outside collaborators. For every collaborator, we schedule a job for the Gitlab::GithubImport::ImportCollaboratorWorker worker.

note
This stage is optional (controlled by Gitlab::GithubImport::Settings) and is selected by default.

6. Stage::ImportPullRequestsMergedByWorker

This worker imports the pull requests’ merged-by user information. The List pull requests API doesn’t provide this information. Therefore, this stage must fetch each merged pull request individually to import this information. A Gitlab::GithubImport::PullRequests::ImportMergedByWorker job is scheduled for each fetched pull request.

7. Stage::ImportPullRequestsReviewRequestsWorker

This worker imports assigned reviewers of pull requests. For each pull request, this worker:

  • Fetches all assigned review requests.
  • Schedules a Gitlab::GithubImport::PullRequests::ImportReviewRequestWorker job for each fetched review request.

8. Stage::ImportPullRequestsReviewsWorker

This worker imports reviews of pull requests. For each pull request, this worker:

  • Fetches all the pages of reviews.
  • Schedules a Gitlab::GithubImport::PullRequests::ImportReviewWorker job for each fetched review.

9. Stage::ImportIssuesAndDiffNotesWorker

This worker imports all issues and pull request comments. For every issue, we schedule a job for the Gitlab::GithubImport::ImportIssueWorker worker. For pull request comments, we instead schedule jobs for the Gitlab::GithubImport::DiffNoteImporter worker.

This worker processes both issues and diff notes in parallel so we don’t need to schedule a separate stage and wait for the previous one to complete.

Issues are imported separately from pull requests because only the “issues” API includes labels for both issue and pull requests. Importing issues and setting label links in the same worker removes the need for performing a separate crawl through the API data, reducing the number of API calls necessary to import a project.

10. Stage::ImportIssueEventsWorker

This worker imports all issues and pull request events. For every event, we schedule a job for the Gitlab::GithubImport::ImportIssueEventWorker worker.

We can import both issues and pull request events by single stage because of a specific aspect of the GitHub API. It looks like that under the hood, issues and pull requests GitHub are stored in a single table. Therefore, they have globally-unique IDs and so:

  • Every pull request is an issue.
  • Issues aren’t pull requests.

Therefore, both issues and pull requests have a common API for most related things.

note
This stage is optional and can consume significant extra import time (controlled by Gitlab::GithubImport::Settings).

11. Stage::ImportNotesWorker

This worker imports regular comments for both issues and pull requests. For every comment, we schedule a job for the Gitlab::GithubImport::ImportNoteWorker worker.

Regular comments have to be imported at the end because the GitHub API used returns comments for both issues and pull requests. This means we have to wait for all issues and pull requests to be imported before we can import regular comments.

12. Stage::ImportAttachmentsWorker

This worker imports note attachments that are linked inside Markdown. For each entity with Markdown text in the project, we schedule a job of:

  • Gitlab::GithubImport::Importer::Attachments::ReleasesImporter for every release.
  • Gitlab::GithubImport::Importer::Attachments::NotesImporter for every note.
  • Gitlab::GithubImport::Importer::Attachments::IssuesImporter for every issue.
  • Gitlab::GithubImport::Importer::Attachments::MergeRequestsImporter for every merge request.

Each job:

  1. Iterates over all attachment links inside of a specific record.
  2. Downloads the attachment.
  3. Replaces the old link with a newly-generated link to GitLab.
note
It’s an optional stage that could consume significant extra import time (controlled by Gitlab::GithubImport::Settings).

13. Stage::ImportProtectedBranchesWorker

This worker imports protected branch rules. For every rule that exists on GitHub, we schedule a job of Gitlab::GithubImport::ImportProtectedBranchWorker.

Each job compares the branch protection rules from GitHub and GitLab and applies the strictest of the rules to the branches in GitLab.

14. Stage::FinishImportWorker

This worker completes the import process by performing some housekeeping (such as flushing any caches) and by marking the import as completed.

Advancing stages

Advancing stages is done in one of two ways:

  • Scheduling the worker for the next stage directly.
  • Scheduling a job for Gitlab::GithubImport::AdvanceStageWorker which will advance the stage when all work of the current stage has been completed.

The first approach should only be used by workers that perform all their work in a single thread, while AdvanceStageWorker should be used for everything else.

When you schedule a job, AdvanceStageWorker is given a project ID, a list of Redis keys, and the name of the next stage. The Redis keys (produced by Gitlab::JobWaiter) are used to check if the running stage has been completed or not. If the stage has not yet been completed AdvanceStageWorker reschedules itself. After a stage finishes, or if more jobs have been finished after the last invocation. AdvanceStageworker refreshes the import JID (more on this below) and schedule the worker of the next stage.

To reduce the number of AdvanceStageWorker jobs scheduled this worker briefly waits for jobs to complete before deciding what the next action should be. For small projects, this may slow down the import process a bit, but it also reduces pressure on the system as a whole.

Refreshing import job IDs

GitLab includes a worker called Gitlab::Import::StuckProjectImportJobsWorker that periodically runs and marks project imports as failed if they have not been refreshed for more than 24 hours. For GitHub projects, this poses a bit of a problem: importing large projects could take several days depending on how often we hit the GitHub rate limit (more on this below), but we don’t want Gitlab::Import::StuckProjectImportJobsWorker to mark our import as failed because of this.

To prevent this from happening we periodically refresh the expiration time of the import. This works by storing the JID of the import job in the database, then refreshing this JID TTL at various stages throughout the import process. This is done either by calling ProjectImportState#refresh_jid_expiration, or by using the RefreshImportJidWorker and passing in the current worker’s jid. By refreshing this TTL we can ensure our import does not get marked as failed so long as we’re still performing work.

GitHub rate limit

GitHub has a rate limit of 5,000 API calls per hour. The number of requests necessary to import a project is largely dominated by the number of unique users involved in a project (for example, issue authors), because we need the email address of users to map them to GitLab users. Other data such as issue pages and comments typically only requires a few dozen requests to import.

We handle the rate limit by doing the following:

  1. After we hit the rate limit, we automatically reschedule jobs in such a way that they are not executed until the rate limit has been reset.
  2. We cache the mapping of GitHub users to GitLab users in Redis.

More information on user caching can be found below.

Caching user lookups

When mapping GitHub users to GitLab users we need to (in the worst case) perform:

  1. One API call to get the user’s Email address.
  2. Two database queries to see if a corresponding GitLab user exists. One query tries to find the user based on the GitHub user ID, while the second query is used to find the user using their GitHub Email address.

To avoid mismatching users, the search by GitHub user ID is not done when importing from GitHub Enterprise.

Because this process is quite expensive we cache the result of these lookups in Redis. For every user looked up we store five keys:

  • A Redis key mapping GitHub usernames to their Email addresses.
  • A Redis key mapping a GitHub Email addresses to a GitLab user ID.
  • A Redis key mapping a GitHub user ID to GitLab user ID.
  • A Redis key mapping a GitHub username to an ETAG header.
  • A Redis key indicating whether an email lookup has been done for a project.

We cache two types of lookups:

  • A positive lookup, meaning we found a GitLab user ID.
  • A negative lookup, meaning we didn’t find a GitLab user ID. Caching this prevents us from performing the same work for users that we know don’t exist in our GitLab database.

The expiration time of these keys is 24 hours. When retrieving the cache of a positive lookup, we refresh the TTL automatically. The TTL of false lookups is never refreshed.

If a lookup for email returns an empty or negative lookup, a Conditional Request is made with a cached ETAG in the header once for every project. Conditional Requests do not count towards the GitHub API rate limit.

Because of this caching layer, it’s possible newly registered GitLab accounts aren’t linked to their corresponding GitHub accounts. This, however, is resolved after the cached keys expire or if a new project is imported.

The user cache lookup is shared across projects. This means that the greater the number of projects that are imported, fewer GitHub API calls are needed.

The code for this resides in:

  • lib/gitlab/github_import/user_finder.rb
  • lib/gitlab/github_import/caching.rb

Increasing Sidekiq interrupts

When a Sidekiq process shut downs, it waits for a period of time for running jobs to finish before it then interrupts them. An interrupt terminates the job and requeues it again. Our vendored sidekiq-reliable-fetcher gem puts a limit of 3 interrupts before a job is no longer requeued and is permanently terminated. Jobs that have been interrupted log a json.interrupted_count in Kibana.

This limit offers protection from jobs that can never be completed in the time between Sidekiq restarts.

For large imports, our GitHub stage workers (namespaced in Stage::) take many hours to finish. By default, the import is at risk of failing because of sidekiq-reliable-fetcher permanently stopping these workers before they can complete.

Stage workers that pick up from where they left off when restarted can increase the interrupt limit of sidekiq-reliable-fetcher to 20 by calling .resumes_work_when_interrupted!: 

module Gitlab
  module GithubImport
    module Stage
      class MyWorker
        resumes_work_when_interrupted!

        # ...
      end
    end
  end
end

Stage workers that do not fully resume their work when restarted should not call this method. For example, a worker that skips already imported objects, but starts its loop from the beginning each time.

Examples of stage workers that do resume work fully are ones that execute services that:

Mapping labels and milestones

To reduce pressure on the database we do not query it when setting labels and milestones on issues and merge requests. Instead, we cache this data when we import labels and milestones, then we reuse this cache when assigning them to issues/merge requests. Similar to the user lookups these cache keys are expired automatically after 24 hours of not being used.

Unlike the user lookup caches, these label and milestone caches are scoped to the project that is being imported.

The code for this resides in:

  • lib/gitlab/github_import/label_finder.rb
  • lib/gitlab/github_import/milestone_finder.rb
  • lib/gitlab/cache/import/caching.rb

Logs

Version history
  • Introduced in GitLab 13.7.
  • Number of imported objects introduced in GitLab 14.1.
  • Gitlab::GithubImport::Logger introduced in GitLab 14.2.
  • import_source renamed to import_type in GitLab 14.2.

The import progress can be checked in the logs/importer.log file. Each relevant import is logged with "import_type": "github" and the "project_id".

The last log entry reports the number of objects fetched and imported: 

{
  "message": "GitHub project import finished",
  "duration_s": 347.25,
  "objects_imported": {
    "fetched": {
      "diff_note": 93,
      "issue": 321,
      "note": 794,
      "pull_request": 108,
      "pull_request_merged_by": 92,
      "pull_request_review": 81
    },
    "imported": {
      "diff_note": 93,
      "issue": 321,
      "note": 794,
      "pull_request": 108,
      "pull_request_merged_by": 92,
      "pull_request_review": 81
    }
  },
  "import_source": "github",
  "project_id": 47,
  "import_stage": "Gitlab::GithubImport::Stage::FinishImportWorker"
}

Metrics dashboards

To assess the GitHub importer health, the GitHub importer dashboard provides information about the total number of objects fetched vs. imported over time.