Skip to content

PROXY_ARCHITECTURE.md

Latest commit

 

History

History
27 lines (21 loc) · 1.85 KB

PROXY_ARCHITECTURE.md

File metadata and controls

27 lines (21 loc) · 1.85 KB
Raw

Daeva Proxy Architecture

Goal

Transform Daeva into a Dual-Mode Engine to seamlessly support both custom automation (via the internal job queue) and generic, off-the-shelf native integrations (like OpenClaw's comfy provider) that expect to talk directly to the underlying service API.

Core Concepts

1. Native Mode (Existing)

The original POST /jobs architecture remains. This is for custom integrations that want to submit background jobs, track state asynchronously, and handle explicit pod management.

2. Proxy Mode (New)

A transparent HTTP/WebSocket proxy layer exposed at /proxy/:podId/*. This allows generic clients to point their baseUrl to Daeva (e.g., http://razerblade.local:8787/proxy/comfyapi).

To the generic client, it looks exactly like the target service. Under the hood, Daeva acts as an API-aware Queueing Proxy:

  1. Intercept: Catch incoming HTTP/WS requests on the proxy route.
  2. Queue / Lock: Acquire the GPU exclusivity lock (waiting if the GPU is currently used by another pod, like Whisper).
  3. Wake: Ensure the requested pod (:podId) is running (starting it via pod-controller if it's currently stopped).
  4. Proxy: Forward the raw traffic to the pod's internal baseUrl.
  5. Release: Once the connection closes (or the request completes), release the GPU lock so other queued pods can spin up.

Implementation Details

  • Hook into the Fastify server (src/server.ts) to add a catch-all proxy route.
  • Use a proxy library (like @fastify/http-proxy or standard Node HTTP proxying) to pipe the requests.
  • Integrate with job-manager.ts or pod-controller.ts to request and hold the execution lock for the duration of the proxy request.
  • Ensure WebSocket upgrades are supported (critical for ComfyUI's /ws endpoint).
  • Keep it transparent: headers, query params, and body must pass through unchanged.