Internal allowed API

The internal/allowed endpoint assesses whether a user has permission to perform certain operations on the Git repository. It performs multiple checks, such as:

  • Ensuring the branch or tag name is acceptable.
  • Whether or not the user has the authority to perform that action.

Endpoint definition

The internal API endpoints are defined under lib/api/internal, and the /allowed path is in lib/api/internal/base.rb.

Use the endpoint

internal/allowed is called when you:

  • Push to the repository.
  • Perform actions on the repository through the GitLab user interface, such as applying suggestions or using the GitLab IDE.

Gitaly typically calls this endpoint. It is only called internally (by other parts of the application) rather than by external users of the API.

Push checks

A key part of the internal/allowed flow is the call to EE::Gitlab::Checks::PushRuleCheck, which can perform the following checks:

  • EE::Gitlab::Checks::PushRules::CommitCheck
  • EE::Gitlab::Checks::PushRules::TagCheck
  • EE::Gitlab::Checks::PushRules::BranchCheck
  • EE::Gitlab::Checks::PushRules::FileSizeCheck

Recursion

Some of the Gitaly RPCs called by internal/allowed then, themselves, make calls back to internal/allowed. These calls are now correlated with the original request. Gitaly relies on the Rails application for authorization, and maintains no permissions model itself.

These calls show up in the logs differently to the initial requests. {example}

Because this endpoint can be called recursively, slow performance on this endpoint can result in an exponential performance impact. This documentation is in fact adapted from the investigation into its performance.

Known performance issues

Refs

The number of refs on the Git repository have a notable effect on the performance of git commands called upon it. Gitaly RPCs are similarly affected. Certain git commands scan through all refs, causing a notable impact on the speed of those commands.

On the internal/allowed endpoint, the recursive nature of RPC calls mean the ref counts have an exponential effect on performance.

Environment refs

Stale environment refs are a common example of excessive refs causing performance issues. Stale environment refs can number into the tens of thousands on busy repositories, as they aren't cleared up automatically.

Dangling refs

Dangling refs are created to prevent accidental deletion of objects from object pools. Large numbers of these refs can exist, which may have potential performance implications. For existing discussion around this issue, read gitaly#1900. This issue appears to have less effect than stale environment refs.

Pool repositories

When a fork is created on GitLab, a central pool repository is created and the forks are linked to it. This pool repository prevents duplication of data by storing data common to other forks. However, the pool repository is not cleaned up in the same manner as the standard repositories, and is more prone to the refs issue.

Feature flags

Parallel push checks

FLAG: On self-managed GitLab, by default this feature is not available. To make it available, an administrator can enable the feature flag named parallel_push_checks. On GitLab.com, by default this feature is not available. To make it available per project, ask GitLab.com administrator to enable the feature flag named parallel_push_checks. You should not use this feature for production environments.

This experimental feature flag enables the endpoint to run multiple RPCs simultaneously, reducing the overall time taken by roughly half. This time savings is achieved through threading, and has potential side effects at large scale. On GitLab.com, this feature flag is enabled only for gitlab-org/gitlab and gitlab-com/www-gitlab-com projects. Without it, those projects routinely time out requests to the endpoint. When this feature was deployed to all of GitLab.com, some pushes failed, presumably due to exhausting resources like database connection pools.

We recommend you enable this feature flag only if you are experiencing timeouts, and only enable it for that specific project.