Engineering a React Native Live-Streaming App for a Creator Platform

The server issued each viewer a token for one channel, so the vendor's secret key never reached the phone.

The vendor's signed "stream started" and "stream ended" webhooks were verified with a body hash and an HMAC signature before they entered Kafka.

Creators went live from the phone camera or from OBS, the open-source broadcasting software, over RTMP. Streams were capped at 4 hours, with warnings 30, 10 and 2 minutes before the end.

On iPhone, the app ran its own camera session, so creators saw a preview before going live. At most 15 frames waited in the queue, and video went out at 720x1280 and 30 fps.

Leaving the app paused a broadcast, and Resume restarted video and chat. On Android, the back button shrank a live broadcast into picture-in-picture, and a creator who reopened the app during a studio stream was offered to jump back in.

A nightly end-to-end suite published test video through the vendor and checked the real-time events a creator and a viewer should receive, across 12 tests and 114 assertions.

Neighboring streams played muted, so the next swipe opened on a stream that was already connected.

On iPhone, neighbors were also capped at about 830 kbps of video bandwidth.

Streams two or more positions away stayed disconnected and used no data.

A feature went live by changing a setting in Firebase Remote Config, with no new app build or store review.

JavaScript fixes shipped over the air through CodePush, and a remote switch could require users to update to a newer app version.

Backend releases promoted the exact container image already deployed to stage. The pipeline stopped with an error when stage had never received that image.

The gateway called 14 services over gRPC, an open-source framework for typed service-to-service calls. The other services exchanged events through Kafka with Avro schemas, and the data lived in MongoDB and Redis.

Internal services shared typed contracts: 20 protobuf files defined 18 gRPC services and 171 calls, and the TypeScript clients were generated from them at build time.

A message that crashed a consumer moved to a dead-letter topic, a holding queue for failed messages.

Image uploads were capped at 15 MB, and the file type was read from the file's first bytes, so only GIF, PNG and JPEG files were accepted.

Scheduled streams could be booked 1 hour to 1 month ahead, on :00 or :30 slots. Reminders reached the creator 60 minutes before and followers 15 minutes before, in push batches of 500 devices.

A circuit breaker cut off a failing dependency after 5 errors and tried it again after 15 seconds. The video vendor's API client used a stricter limit of 3 errors.

All 23 services loaded the Datadog tracer before anything else, and the gateway passed each request's trace ID to the services it called, so one user action could be followed across the system. Sentry reported errors, and all services shared one set of 314 error codes, numbered in blocks per area.