openreader/docs-site/docs/deploy/compute-worker.md
Richard R f1aa1c3e3b feat(compute): add external compute worker backend and integration
Introduce support for external compute worker mode (`COMPUTE_MODE=worker`) using a new `WorkerComputeBackend`. This enables offloading heavy ONNX Whisper alignment and PDF layout parsing to a standalone worker service (Redis + BullMQ), improving scalability and compatibility with serverless/limited environments.

- Add `@openreader/compute-core` as a shared package for ONNX inference and PDF parsing logic.
- Implement `WorkerComputeBackend` and worker contract/types for remote job execution.
- Update compute backend selection logic and remove previous worker mode guards.
- Extend `WhisperAlignInput` and `PdfLayoutInput` types to support object keys for remote data access.
- Refactor local compute backend to use `@openreader/compute-core` and support both buffer and object key inputs.
- Update job runner, TTS segment alignment, and PDF layout parsing flows to use new compute backend APIs.
- Add scripts, Docker workflow, and documentation for deploying and running the compute worker.
- Update environment variable docs and examples for worker mode, including storage requirements and configuration.
- Document published images and stack changes to reflect the new compute worker architecture.

BREAKING CHANGE: `COMPUTE_MODE=worker` now requires an external compute worker service and S3-compatible object storage. Embedded SeaweedFS (`weed mini`) is not supported in worker mode. See the new documentation for deployment and configuration details.
2026-05-19 15:21:25 -06:00

64 lines
1.7 KiB
Markdown

---
title: Compute Worker (Redis + BullMQ)
---
Use this guide for `COMPUTE_MODE=worker` deployments where heavy compute runs outside the Next.js app server.
## Overview
The compute worker handles:
- Whisper word alignment (`/align/whisper/jobs`)
- PDF layout parsing (`/layout/pdf/jobs`)
The app server enqueues jobs and polls status. Queue durability and retries are backed by Redis + BullMQ.
## Published image
- App server image: `ghcr.io/richardr1126/openreader`
- Compute worker image: `ghcr.io/richardr1126/openreader-compute-worker`
## Worker environment variables
Required:
- `COMPUTE_WORKER_TOKEN`: bearer token expected by worker routes
- `REDIS_URL`: BullMQ Redis connection string
- `S3_BUCKET`
- `S3_REGION`
- `S3_ACCESS_KEY_ID`
- `S3_SECRET_ACCESS_KEY`
Common optional:
- `S3_ENDPOINT` (for non-AWS S3-compatible storage)
- `S3_FORCE_PATH_STYLE=true` (for many S3-compatible providers)
- `S3_PREFIX=openreader`
- `COMPUTE_WORKER_HOST=0.0.0.0`
- `COMPUTE_WORKER_PORT=8081`
- `COMPUTE_LOG_FORMAT=pretty` (default) or `json`
- `COMPUTE_QUEUE_MAX_DEPTH=64`
- `COMPUTE_PREWARM_MODELS=true`
## App server environment variables (worker mode)
Set on the Next.js app server:
```env
COMPUTE_MODE=worker
COMPUTE_WORKER_URL=http://<worker-host>:8081
COMPUTE_WORKER_TOKEN=<same-token-as-worker>
```
`COMPUTE_MODE=worker` has no local fallback. If worker is unavailable, affected requests fail.
## Production notes
- Worker mode assumes shared object storage is reachable by both app server and worker.
- Non-exposed embedded `weed mini` is not supported with external worker mode.
- Protect `COMPUTE_WORKER_TOKEN` and avoid exposing worker routes publicly without auth.
## Health endpoints
- `GET /health/live`
- `GET /health/ready`