GitLab Documentation

GitLab Geo

Geo feature requires that we orchestrate a lot of components together. For the Database we need to setup replication, writing operations that stores data directly to disk replicates asynchronously by sending Webhook requests from Primary to Secondary nodes, and assets are to be replicated in a future release using either a shared filesystem architecture or an object store setup with geographical replication.

Primary and Secondary

There can be only one Primary node, which is the one that can do writing operations. Both share the same codebase and are distinguished by a feature toggle mechanism (see Gitlab::Geo).

We use the values from gitlab.yml: host, port, relative_url_root and search in the database to identity which node we are in (see Gitlab::Geo.current_node).

Most of the Geo important methods are cached by the RequestStore, to reduce the performance impact of using the methods throghout the codebase.

To execute a piece of code in a Primary node use:

if Gitlab::Geo.primary?
  # code goes here

You can do the same thing for a Secondary node:

if Gitlab::Geo.secondary?
  # code goes here

.primary? and .secondary? are not mutually excludable, so you should never take for granted that when one of them returns false, other will be true.

Both methods check if Geo is .enabled?, so there is a "third" state where both will return false (when Geo is not enabled).


We consider Geo feature enabled when the user has a valid license with the feature included, and they have at least one node defined at the Geo Nodes screen.


Previous implementation (GitLab =< 8.6.x) used custom code to handle notification from Primary to Secondary by HTTP requests.

We decided to move away from custom code and integrate by using System Webhooks, as we have more people using them, so any improvements we make to this communication layer, many other will benefit from.

There is a specific internal endpoint in our api code (Grape), that receives all requests from this System Hooks: /api/v3/geo/receive_events.

We switch and filter from each event by the event_name field.


All Secondary nodes are read-only.

We have a Rails Middleware that filters any potentially writing operations and prevent user from trying to update the database and getting a 500 error (see Gitlab::Middleware::ReadonlyGeo).

Database will already be read-only in a replicated setup, so we don't need to take any extra step for that.

We do use our feature toggle .secondary? to coordinate Git operations and do the correct authorization (denying writing on any secondary node).

File Transfers

Secondary Geo Nodes need to transfer files, such as LFS objects, attachments, avatars, etc. from the primary. To do this, secondary nodes have a separate tracking database that records which objects it needs to transfer.

Files are copied via HTTP(s) and initiated via the /api/v4/geo/transfers/:type/:id endpoint.


To authenticate file transfers, each GeoNode has two fields:

  1. A public access key (access_key)
  2. A secret access key (secret_access_key)

The secondary authenticates itself via a JWT request. When the secondary wishes to download a file, it sends an HTTP request with the Authorization header:

Authorization: GL-Geo <access_key>:<JWT payload>

The primary uses the access_key to look up the corresponding Geo node and decrypt the JWT payload, which contains additional information to identify the file request. This ensures that the secondary downloads the right file for the right database ID. For example, for an LFS object, the request must also include the SHA256 of the file. An example JWT payload looks like:

{ "data": { sha256: "31806bb23580caab78040f8c45d329f5016b0115" }, iat: "1234567890" }

If the data checks out, then the Geo primary sends data via the XSendfile feature, which allows nginx to handle the file transfer without tying up Rails or Workhorse.

Geo Tracking Database

Secondary Geo nodes track data about what has been downloaded in a second PostgreSQL database that is distinct from the production GitLab database. The database configuration is set in config/database_geo.yml. db/geo contains the schema and migrations for this database.

To write a migration for the database, use the GeoMigrationGenerator:

rails g geo_generator [args] [options]

To migrate the tracking database, run:

bundle exec rake geo:db:migrate