Geo proxying for secondary sites

Version history

Use Geo proxying to:

  • Have secondary sites serve read-write traffic by proxying to the primary site.
  • Selectively accelerate replicated data types by directing read-only operations to the local site instead.

When enabled, users of the secondary site can use the WebUI as if they were accessing the primary site’s UI. This significantly improves the overall user experience of secondary sites.

With secondary proxying, web requests to secondary Geo sites are proxied directly to the primary, and appear to act as a read-write site.

Proxying is done by the gitlab-workhorse component. Traffic usually sent to the Rails application on the Geo secondary site is proxied to the internal URL of the primary Geo site instead.

Use secondary proxying for use-cases including:

  • Having all Geo sites behind a single URL.
  • Geographically load-balancing traffic without worrying about write access.
note
Geo proxying for secondary sites is separate from Geo proxying/redirecting for Git push operations.

For an overview, see: Secondary proxying using geographic load-balancer and AWS Route53.

Set up a unified URL for Geo sites

Secondary sites can transparently serve read-write traffic. You can use a single external URL so that requests can hit either the primary Geo site or any secondary Geo sites that use Geo proxying.

Configure an external URL to send traffic to both sites

Follow the Location-aware public URL steps to create a single URL used by all Geo sites, including the primary.

Update the Geo sites to use the same external URL

  1. On your Geo sites, SSH **into each node running Rails (Puma, Sidekiq, Log-Cursor) and change the external_url to that of the single URL:

    sudo editor /etc/gitlab/gitlab.rb
    
  2. Reconfigure the updated nodes for the change to take effect if the URL was different than the one already set:

    gitlab-ctl reconfigure
    
  3. To match the new external URL set on the secondary Geo sites, the primary database needs to reflect this change.

    In the Geo administration page of the primary site, edit each Geo secondary that is using the secondary proxying and set the URL field to the single URL. Make sure the primary site is also using this URL.

In Kubernetes, you can use the same domain under global.hosts.domain as for the primary site.

Geo proxying with Separate URLs

Geo secondary proxying for separate URLs is enabled by default in GitLab 15.1.

note
The feature flag described in this section is planned to be deprecated and removed in a future release. Support for read-only Geo secondary sites is proposed in issue 366810, you can upvote and share your use cases in that issue.

There are minor known issues linked in the “Geo secondary proxying with separate URLs” epic. You can also add feedback in the epic about any use-cases that are not possible anymore with proxying enabled.

If you run into issues, to disable this feature, disable the geo_secondary_proxy_separate_urls feature flag.

  1. SSH into one node running Rails on your primary Geo site and run:

    sudo gitlab-rails runner "Feature.disable(:geo_secondary_proxy_separate_urls)"
    
  2. Restart Puma on all of the nodes running Rails on your secondary Geo site:

    sudo gitlab-ctl restart puma
    

In Kubernetes, you can run the same command in the toolbox pod. Refer to the Kubernetes cheat sheet for details.

Limitations

  • When secondary proxying is used, the asynchronous Geo replication can cause unexpected issues for accelerated data types that may be replicated to the Geo secondaries with a delay.

    For example, we found a potential issue where replication lag introduces read-after-write inconsistencies. If the replication lag is high enough, this can result in Git reads receiving stale data when hitting a secondary.

  • Non-Rails requests are not proxied, so other services may need to use a separate, non-unified URL to ensure requests are always sent to the primary. These services include:

  • With a unified URL, Let’s Encrypt can’t generate certificates unless it can reach both IPs through the same domain. To use TLS certificates with Let’s Encrypt, you can manually point the domain to one of the Geo sites, generate the certificate, then copy it to all other sites.

  • Viewing projects and designs data from a primary site is not possible when using a unified URL.

  • When secondary proxying is used together with separate URLs, registering GitLab runners to clone from secondary sites is not supported. The runner registration succeeds, but the clone URL defaults to the primary site. The runner clone URL is configured per GitLab deployment and cannot be configured per Geo site. Therefore, all runners clone from the primary site (or configured clone URL) irrespective of which Geo site they register on. For information about GitLab CI using a specific Geo secondary to clone from, see issue 3294.

Behavior of secondary sites when the primary Geo site is down

Considering that web traffic is proxied to the primary, the behavior of the secondary sites differs when the primary site is inaccessible:

  • UI and API traffic return the same errors as the primary (or fail if the primary is not accessible at all), since they are proxied.
  • For repositories that already exist on the specific secondary site being accessed, Git read operations still work as expected, including authentication through HTTP(s) or SSH.
  • Git operations for repositories that are not replicated to the secondary site return the same errors as the primary site, since they are proxied.
  • All Git write operations return the same errors as the primary site, since they are proxied.

Features accelerated by secondary Geo sites

Most HTTP traffic sent to a secondary Geo site can be proxied to the primary Geo site. With this architecture, secondary Geo sites are able to support write requests. Certain read requests are handled locally by secondary sites for improved latency and bandwidth nearby. All write requests are proxied to the primary site.

The following table details the components currently tested through the Geo secondary site Workhorse proxy. It does not cover all data types.

Feature / component Accelerated reads?
Project, wiki, design repository (using the web UI) No
Project, wiki repository (using Git) Yes 1
Project, Personal Snippet (using the web UI) No
Project, Personal Snippet (using Git) Yes 1
Group wiki repository (using the web UI) No
Group wiki repository (using Git) Yes 1
User uploads No
LFS objects (using the web UI) No
LFS objects (using Git) Yes
Pages No 2
Advanced search (using the web UI) No
  1. Git reads are served from the local secondary while pushes get proxied to the primary. Selective sync or cases where repositories don’t exist locally on the Geo secondary throw a “not found” error.
  2. Pages can use the same URL (without access control), but must be configured separately and are not proxied.

Disable Geo proxying

Secondary proxying is enabled by default on a secondary site when it uses a unified URL, meaning, the same external_url as the primary site. Disabling proxying in this case tends to not be helpful due to completely different behavior being served at the same URL, depending on routing.

Secondary proxying is enabled by default in GitLab 15.1 on a secondary site even without a unified URL. If proxying needs to be disabled on a secondary site, it is much easier to disable the feature flag in Geo proxying with Separate URLs. However, if there are multiple secondary sites, then the instructions in this section can be used to disable secondary proxying per site.

Additionally, the gitlab-workhorse service polls /api/v4/geo/proxy every 10 seconds. In GitLab 15.2 and later, it is only polled once, if Geo is not enabled. Prior to GitLab 15.2, you can stop this polling by disabling secondary proxying.

You can disable the secondary proxying on each Geo site, separately, by following these steps with Omnibus-based packages:

  1. SSH into each application node (serving user traffic directly) on your secondary Geo site and add the following environment variable:

    sudo editor /etc/gitlab/gitlab.rb
    
    gitlab_workhorse['env'] = {
      "GEO_SECONDARY_PROXY" => "0"
    }
    
  2. Reconfigure the updated nodes for the change to take effect:

    gitlab-ctl reconfigure
    

In Kubernetes, you can use --set gitlab.webservice.extraEnv.GEO_SECONDARY_PROXY="0", or specify the following in your values file:

gitlab:
  webservice:
    extraEnv:
      GEO_SECONDARY_PROXY: "0"