Service Topology

Who talks to whom, over what, and with which credential. This is the page to read before adding a service or debugging a cross-service failure.

Communication matrix

From	To	Transport	Auth	Purpose
Mobile	api	HTTPS REST	Authorization: Bearer <JWT>	Everything
Mobile	esu	Socket.IO (WSS)	JWT in the handshake	Realtime
Mobile	LiveKit	WebRTC	LiveKit access token minted by the API	Voice rooms
Mobile	Mux	HLS / RTMP	Public playback ID; stream key for broadcast	Livestreaming
Web	api	HTTPS REST	—	Public data
api	esu	HTTP POST /admin/emit/*	X-Realtime-Admin-Token	Fan out realtime events
esu	api	HTTP GET /api/v1/{dm,room}/:id/authz	Forwards the user's JWT	Authorize a join
ogun	esu	HTTP POST /admin/emit/*	X-Realtime-Admin-Token	Media-ready notification
api	Redis	BullMQ	password	Enqueue media jobs
ogun	Redis	BullMQ	password	Consume media jobs
esu	Redis	Socket.IO Redis adapter	password	Cross-pod pub/sub
api	Redis	ioredis	password	Feed / trending cache
all services	PostgreSQL	Prisma	connection string	Data
RevenueCat	api	HTTPS webhook	shared secret	Purchase events
Mux	api	HTTPS webhook	signing secret	Asset and livestream state
anansi	Stripe	HTTPS API	secret key	Top-ups and creator transfers

The interservice channel

esu exposes an admin router at /admin, guarded by a shared token (apps/esu/src/interservice/adminRoutes.ts):

router.use(requireAdminToken);        // checks X-Realtime-Admin-Token

router.post("/emit/dm/message/new", …);
router.post("/emit/dm/read", …);
router.post("/emit/room/message/new", …);
router.post("/emit/channel/room/created", …);
router.post("/emit/notification/new", …);
router.post("/emit/media/processing/update", …);

Each handler does one thing: emit a Socket.IO event into a room. No database access, no authorization beyond the token.

The API's helper adds the /admin prefix itself (apps/api/src/lib/interservice.ts):

fetch(`${process.env.REALTIME_SERVICE_URL}/admin${path}`, …)

Ogun's helper is identical (apps/ogun/src/utils/interservice.ts) — keep the two copies in sync.

Both normalise the base URL before use, so REALTIME_SERVICE_URL works whether or not it already carries a trailing /admin:

const base = (process.env.REALTIME_SERVICE_URL ?? "").replace(/\/+$/, "").replace(/\/admin$/, "");
fetch(`${base}/admin${path}`, …)

Prefer the bare service root (http://playpalz-esu:4010) in new configuration.

Previously: REALTIME_SERVICE_URL meant two different things

The helpers used to disagree about whether the env value included /admin, while the manifests set one value for both. That was correct for the API and wrong for ogun, so the "your photo finished processing" event silently 404'd in production; locally the polarity reversed and the API POSTed to /admin/admin/…. .catch(() => {}) at every call site hid both.

Two things stop this recurring: the normalisation above, and the helpers now logging and throwing on a non-2xx response instead of returning it unexamined.

Realtime room naming

esu groups sockets into rooms (apps/esu/src/rooms/roomNames.ts). Every emit targets one:

Room	Contains	Used for
user:<userId>	Every socket for one user	Notifications, media-ready, presence-targeted events
dm:<conversationId>	Participants of a direct conversation	Messages, read receipts, typing
room:<roomId>	Members currently in a channel room	Messages, typing
channel:<channelId>	Channel members	Room created / updated, member joined / left

A socket joins user:<id> automatically on connect. All other joins are explicit and authorized via the callback to the API.

Why esu calls back into the API

It would be faster for esu to query Postgres directly for "is this user allowed in this conversation". It does not, on purpose:

• Access rules for channels involve subscription state, blocks, and membership. Duplicating them would mean two implementations that drift.
• esu stays free of business logic, so it can be restarted, scaled, or replaced without coordinating a data-model change.

The cost is one HTTP round trip per join. That is acceptable — joins are rare compared to messages.

Scaling characteristics

Service	Scales by	Constraint
api	Horizontal, stateless	Postgres connection count
esu	Horizontal via the Redis adapter	Ingress needs sticky sessions for long-polling — hence the ESUSESSION cookie on the socket ingress
ogun	Horizontal, more workers = more throughput	CPU-bound (Sharp); Mux is rate-limited
anansi	Single effective runner	Guarded by a Redis distributed lock; a second replica would double-pay creators
igdb-heartbeat	Single	Cron
web	Horizontal, stateless	—

anansi and igdb-heartbeat must not be scaled up casually

Both are cron-driven. anansi uses a Redis lock to make duplicate runs safe, but that lock is the only thing standing between you and duplicate Stripe transfers. Treat replica counts for those two deployments as a deliberate decision, not a knob.

Failure modes

If this dies	What breaks	What keeps working
api	Everything	Nothing meaningful
esu	Live message delivery, typing, presence	Sending and reading messages (the API persists them; clients see them on refresh)
ogun	Thumbnails and video transcoding	Uploads still succeed; media stays pending and processes when the worker returns
anansi	Payout runs	The product; earnings still accrue in the database
igdb-heartbeat	Game catalog freshness	Everything
Redis	Queue, realtime fan-out, caching	Degraded but mostly functional — cache misses fall through to Postgres
PostgreSQL	Everything	Nothing

The useful property here is that esu and ogun are both deferrable: losing them degrades the experience but does not lose data.