Skip to content

Geo proxying

With Geo proxying, secondaries now proxy web requests through Workhorse to the primary, so users navigating to the secondary see a read-write UI, and are able to do all operations that they can do on the primary.

Request life cycle

Top-level view

The proxying interaction can be explained at a high level through the following diagram:

sequenceDiagram
actor client
participant secondary
participant primary

client->>secondary: GET /explore
secondary-->>primary: GET /explore (proxied)
primary-->>secondary: HTTP/1.1 200 OK [..]
secondary->>client: HTTP/1.1 200 OK [..]

Proxy detection mechanism

To know whether or not it should proxy requests to the primary, and the URL of the primary (as it is stored in the database), Workhorse polls the internal API when Geo is enabled. When proxying should be enabled, the internal API responds with the primary URL and JWT-signed data that is passed on to the primary for every request.

sequenceDiagram
    participant W as Workhorse (secondary)
    participant API as Internal Rails API
    W->API: GET /api/v4/geo/proxy (internal)
    loop Poll every 10 seconds
        API-->W: {geo_proxy_primary_url, geo_proxy_extra_data}, update config
    end

In-depth request flow and local data acceleration compared with proxying

Detailing implementation, Workhorse on the secondary (requested) site decides whether to proxy the data or not. If it can "accelerate" the data type (that is, can serve locally to save a roundtrip request), it returns the data immediately. Otherwise, traffic is sent to the primary's internal URL, served by Workhorse on the primary exactly as a direct request would. The response is then be proxied back to the user through the secondary Workhorse in the same connection.

flowchart LR
  A[Client]--->W1["Workhorse (secondary)"]
  W1 --> W1C[Serve data locally?]
  W1C -- "Yes" ----> W1
  W1C -- "No (proxy)" ----> W2["Workhorse (primary)"]
  W2 --> W1 ----> A

Sign-in

Requests proxied to the primary requiring authorization

sequenceDiagram
autoNumber
participant Client
participant Secondary
participant Primary

Client->>Secondary: `/group/project` request
Secondary->>Primary: proxy /group/project
opt primary not signed in
Primary-->>Secondary: 302 redirect
Secondary-->>Client: proxy 302 redirect
Client->>Secondary: /users/sign_in
Secondary->>Primary: proxy /users/sign_in
Note right of Primary: authentication happens, POST to same URL etc
Primary-->>Secondary: 302 redirect
Secondary-->>Client: proxy 302 redirect
Client->>Secondary: /group/project
Secondary->>Primary: proxy /group/project
end
Primary-->>Secondary: /group/project logged in response (session on primary created)
Secondary-->>Client: proxy full response

Git pull

For historical reasons, the push_from_secondary path is used to forward a Git pull. There is an issue proposing to rename this route to avoid confusion.

Git pull over HTTP(s)

Accelerated repositories

When a repository exists on the secondary and we detect is up to date with the primary, we serve it directly instead of proxying.

sequenceDiagram
participant C as Git client
participant Wsec as "Workhorse (secondary)"
participant Rsec as "Rails (secondary)"
participant Gsec as "Gitaly (secondary)"
C->>Wsec: GET /foo/bar.git/info/refs/?service=git-upload-pack
Wsec->>Rsec: <internal API check>
note over Rsec: decide that the repo is synced and up to date
Rsec-->>Wsec: 401 Unauthorized
Wsec-->>C: <response>
C->>Wsec: GET /foo/bar.git/info/refs/?service=git-upload-pack
Wsec->>Rsec: <internal API check>
Rsec-->>Wsec: Render Workhorse OK
Wsec-->>C: 200 OK
C->>Wsec: POST /foo/bar.git/git-upload-pack
Wsec->>Rsec: GitHttpController#git_receive_pack
Rsec-->>Wsec: Render Workhorse OK
Wsec->>Gsec: Workhorse gets the connection details from Rails, connects to Gitaly: SmartHTTP Service, UploadPack RPC (check the proto for details)
Gsec-->>Wsec: Return a stream of Proto messages
Wsec-->>C: Pipe messages to the Git client

Proxied repositories

If a requested repository isn't synced, or we detect is not up to date, the request will be proxied to the primary, in order to get the latest version of the changes.

sequenceDiagram
participant C as Git client
participant Wsec as "Workhorse (secondary)"
participant Rsec as "Rails (secondary)"
participant W as "Workhorse (primary)"
participant R as "Rails (primary)"
participant G as "Gitaly (primary)"
C->>Wsec: GET /foo/bar.git/info/refs/?service=git-upload-pack
Wsec->>Rsec: <response>
note over Rsec: decide that the repo is out of date
Rsec-->>Wsec: 302 Redirect to /-/push_from_secondary/2/foo/bar.git/info/refs?service=git-upload-pack
Wsec-->>C: <response>
C->>Wsec: GET /-/push_from_secondary/2/foo/bar.git/info/refs/?service=git-upload-pack
Wsec->>W: <proxied request>
W->>R: <data>
R-->>W: 401 Unauthorized
W-->>Wsec: <proxied response>
Wsec-->>C: <response>
C->>Wsec: GET /-/push_from_secondary/2/foo/bar.git/info/refs/?service=git-upload-pack
note over W: proxied
Wsec->>W: <proxied request>
W->>R: <data>
R-->>W: Render Workhorse OK
W-->>Wsec: <proxied response>
Wsec-->>C: <response>
C->>Wsec: POST /-/push_from_secondary/2/foo/bar.git/git-upload-pack
Wsec->>W: <proxied request>
W->>R: GitHttpController#git_receive_pack
R-->>W: Render Workhorse OK
W->>G: Workhorse gets the connection details from Rails, connects to Gitaly: SmartHTTP Service, UploadPack RPC (check the proto for details)
G-->>W: Return a stream of Proto messages
W-->>Wsec: Pipe messages to the Git client
Wsec-->>C: Return piped messages from Git

Git pull over SSH

As SSH operations go through GitLab Shell instead of Workhorse, they are not proxied through the mechanism used for Workhorse requests. With SSH operations, they are proxied as Git HTTP requests to the primary site by the secondary Rails internal API.

Accelerated repositories

When a repository exists on the secondary and we detect is up to date with the primary, we serve it directly instead of proxying.

sequenceDiagram
participant C as Git client
participant S as GitLab Shell (secondary)
participant I as Internal API (secondary Rails)
participant G as Gitaly (secondary)
C->>S: git pull
S->>I: SSH key validation (api/v4/internal/authorized_keys?key=..)
I-->>S: HTTP/1.1 200 OK
S->>G: InfoRefs:UploadPack RPC
G-->>S: stream Git response back
S-->>C: stream Git response back
C-->>S: stream Git data to push
S->>G: UploadPack RPC
G-->>S: stream Git response back
S-->>C: stream Git response back

Proxied repositories

If a requested repository isn't synced, or we detect is not up to date, the request will be proxied to the primary, in order to get the latest version of the changes.

sequenceDiagram
participant C as Git client
participant S as GitLab Shell (secondary)
participant I as Internal API (secondary Rails)
participant P as Primary API
C->>S: git pull
S->>I: SSH key validation (api/v4/internal/authorized_keys?key=..)
I-->>S: HTTP/1.1 300 (custom action status) with {endpoint, msg, primary_repo}
S->>I: POST /api/v4/geo/proxy_git_ssh/info_refs_upload_pack
I->>P: POST $PRIMARY/foo/bar.git/info/refs/?service=git-upload-pack
P-->>I: HTTP/1.1 200 OK
I-->>S: <response>
S-->>C: return Git response from primary
C-->>S: stream Git data to push
S->>I: POST /api/v4/geo/proxy_git_ssh/upload_pack
I->>P: POST $PRIMARY/foo/bar.git/git-upload-pack
P-->>I: HTTP/1.1 200 OK
I-->>S: <response>
S-->>C: return Git response from primary

Git push

Git push over SSH

As SSH operations go through GitLab Shell instead of Workhorse, they are not proxied through the mechanism used for Workhorse requests. With SSH operations, they are proxied as Git HTTP requests to the primary site by the secondary Rails internal API.

sequenceDiagram
participant C as Git client
participant S as GitLab Shell (secondary)
participant I as Internal API (secondary Rails)
participant P as Primary API
C->>S: git push
S->>I: SSH key validation (api/v4/internal/authorized_keys?key=..)
I-->>S: HTTP/1.1 300 (custom action status) with {endpoint, msg, primary_repo}
S->>I: POST /api/v4/geo/proxy_git_ssh/info_refs_receive_pack
I->>P: POST $PRIMARY/foo/bar.git/info/refs/?service=git-receive-pack
P-->>I: HTTP/1.1 200 OK
I-->>S: <response>
S-->>C: return Git response from primary
C-->>S: stream Git data to push
S->>I: POST /api/v4/geo/proxy_git_ssh/receive_pack
I->>P: POST $PRIMARY/foo/bar.git/git-receive-pack
P-->>I: HTTP/1.1 200 OK
I-->>S: <response>
S-->>C: return Git response from primary

Git push over HTTP(S)

If a requested repository isn't synced, or we detect is not up to date, the request will be proxied to the primary, a push redirects to a local path formatted as /-/push_from_secondary/$SECONDARY_ID/*. Further, requests through this path are proxied to the primary, which will handle the push.

sequenceDiagram
participant C as Git client
participant Wsec as Workhorse (secondary)
participant W as Workhorse (primary)
participant R as Rails (primary)
participant G as Gitaly (primary)
C->>Wsec: GET /foo/bar.git/info/refs/?service=git-receive-pack
Wsec->>C: 302 Redirect to /-/push_from_secondary/2/foo/bar.git/info/refs?service=git-receive-pack
C->>Wsec: GET /-/push_from_secondary/2/foo/bar.git/info/refs/?service=git-receive-pack
Wsec->>W: <proxied request>
W->>R: <data>
R-->>W: 401 Unauthorized
W-->>Wsec: <proxied response>
Wsec-->>C: <response>
C->>Wsec: GET /-/push_from_secondary/2/foo/bar.git/info/refs/?service=git-receive-pack
Wsec->>W: <proxied request>
W->>R: <data>
R-->>W: Render Workhorse OK
W-->>Wsec: <proxied response>
Wsec-->>C: <response>
C->>Wsec: POST /-/push_from_secondary/2/foo/bar.git/git-receive-pack
Wsec->>W: <proxied request>
W->>R: GitHttpController:git_receive_pack
R-->>W: Render Workhorse OK
W->>G: Get connection details from Rails and connects to SmartHTTP Service, ReceivePack RPC
G-->>W: Return a stream of Proto messages
W-->>Wsec: Pipe messages to the Git client
Wsec-->>C: Return piped messages from Git