2a53d675f6
* fix(jobs): survive Redis Cloud maintenance reconnects BullMQ duplicates ioredis connections without inheriting error listeners, so Redis Cloud maintenance (ECONNABORTED) crashed the API process. Attach error handlers on duplicates, harden reconnect options for failover, and add API process-level safety nets. Co-authored-by: Rotimi Best <rotimi-best@users.noreply.github.com> * refactor(jobs): address review feedback on Redis resilience fix - reconnectOnError returns 2 for READONLY/LOADING so the failed command is resent against the newly promoted primary after failover - export prepareRedisConnection and test the real implementation instead of a reimplemented copy; add a no-double-listener regression case - API uncaughtException now reports to Sentry and fails fast (flush + exit 1) so the orchestrator restarts cleanly instead of running in an undefined state Co-authored-by: Rotimi Best <rotimi-best@users.noreply.github.com> * fix(api): don't fail-fast on transient Redis errors in process guard End-to-end testing (restarting Redis to simulate Cloud maintenance) revealed node-redis surfaces reconnect failures as an uncaught AggregateError [ECONNREFUSED], which the fail-fast uncaughtException guard turned into a process exit. Classify transient Redis/socket errors (incl. AggregateError sub-errors) and log-and-continue for those; still fail fast on unknown errors. Co-authored-by: Rotimi Best <rotimi-best@users.noreply.github.com> * fix(api): address PR review on Redis process guards and logging - Fail fast on non-transient unhandledRejection (Greptile) - Filter Sentry OnUncaughtException/OnUnhandledRejection integrations to avoid double-capture; keep custom guards with transient Redis filtering - Make wrapDuplicate idempotent; throttle reconnect/error logs during maintenance - Drop READONLY/LOADING from node-redis transient classifier (command-level only) Co-authored-by: Rotimi Best <rotimi-best@users.noreply.github.com> * fix(api): rate-limit transient Redis warnings in process guards Reuse logTransientRedisEvent from core so maintenance bursts do not flood logs from uncaughtException/unhandledRejection handlers. Co-authored-by: Rotimi Best <rotimi-best@users.noreply.github.com> --------- Co-authored-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rotimi Best <rotimi-best@users.noreply.github.com>
ClassroomIO Backend
Supported Features
- Certificate download for students.
- Uploading videos to cloudflare (for storage) and Muse.ai (for transcription)
Running locally with video upload + playback
Lesson videos run through an HLS pipeline on Cloud. The browser encodes a
360/720/1080 ABR ladder with Mediabunny, uploads segments to R2 via presigned
PUTs, and the player loads /hls/{assetId}/master.m3u8 from the API.
How /hls/* resolves depends on the environment:
| URL the browser hits | Who serves the bytes | |
|---|---|---|
| Cloud production | https://<tenant-host>/hls/... |
Cloudflare Worker (apps/tenant-router) — intercepts the path and streams directly from an R2 binding. Render is never in the bytestream. |
| Local dev / self-hosted | ${PUBLIC_SERVER_URL}/hls/... (direct to API) |
Hono streaming route on the API — session-auth, then getS3Client().send(GetObjectCommand). |
Self-hosted (PUBLIC_IS_SELFHOSTED=true) skips HLS entirely and falls
back to raw-MP4 upload — no R2-specific code path. Everything below is for
Cloud-style local development.
Required env
apps/api/.env:
# Object storage — Cloudflare R2 OR an S3-compatible local backend (MinIO, etc.)
CLOUDFLARE_ACCOUNT_ID=…
CLOUDFLARE_ACCESS_KEY=…
CLOUDFLARE_SECRET_ACCESS_KEY=…
# (or OBJECT_STORAGE_ENDPOINT / OBJECT_STORAGE_ACCESS_KEY_ID / OBJECT_STORAGE_SECRET_ACCESS_KEY)
# Optional — only needed if you want extract-audio → Whisper transcripts
OPENAI_API_KEY=…
# Optional — Redis enables the BullMQ media job pipeline. Without it the
# asset row is still created, but no transcription/thumbnail jobs run.
REDIS_URL=redis://localhost:6379
# Only required when you want to exercise the production cookie-mint path
# (`POST /organization/assets/:assetId/hls/cookie`). Locally the Hono /hls/*
# route uses session auth, so most of the time you can leave this unset.
# Same value must also be set as a Cloudflare secret on tenant-router for prod.
HLS_SIGNING_SECRET=…
apps/dashboard/.env:
# Already required by the existing API client — used as the HLS base URL too.
PUBLIC_SERVER_URL=http://localhost:3000
Run
pnpm --filter @cio/api dev
pnpm --filter @cio/dashboard dev
# (optional) pnpm --filter @cio/jobs dev # for transcription + thumbnail jobs
Drop a video into a lesson; the dashboard encodes locally, you'll see
segments arrive under videos/{assetId}/hls/… in R2 and an assets row with
hls_manifest_key populated. Playback flows directly browser → API → R2 —
no wrangler dev needed locally.
Production setup (one-time)
For Cloud deployment, generate the HMAC signing key the Worker uses to verify playback cookies and set it on both sides:
node -e "console.log(require('crypto').randomBytes(48).toString('base64url'))"
# On Render (apps/api service env):
HLS_SIGNING_SECRET=<paste>
# On Cloudflare (apps/tenant-router):
pnpm --filter @cio/tenant-router exec wrangler secret put HLS_SIGNING_SECRET
# paste the same value when prompted
Debugging Fly.io
View logs from fly.io in real time