An interactive study companion for Martin Kleppmann's classic on reliable, scalable, and maintainable systems.
YouTube must ingest massive files, serve globally with low startup latency, and recommend the next clip — all profitably.
Video is the bulk of bytes; metadata is the bulk of queries. Treating them as one system fails — upload and transcode are write pipelines, playback is a CDN problem, home feed is a ranking problem.
YouTube stores originals in durable object storage (GCS-style), serves via Google's CDN, and keeps metadata in sharded databases. Creators see processing state; viewers see p95 startup time under a few seconds on good networks.
A 2 GB upload cannot block a synchronous API thread. Resumable uploads land in object storage; a job queue fans out transcoding workers that write HLS/DASH segments back to the bucket.
Video files are gigabytes — storing them in Postgres or on API servers is impossible at scale. Object storage offers pennies-per-GB durability with multipart/resumable upload APIs. Metadata (title, status) lives elsewhere.
Accepts chunked uploads, validates auth, and issues pre-signed URLs so bytes stream directly to object storage without proxying through app servers. Keeps API pods memory-light.
Transcoding takes minutes and must not block the upload response. An event ('video.uploaded') fans out to worker pools that scale independently. Also feeds moderation, thumbnails, and search indexing.
CPU-heavy, embarrassingly parallel — one job per resolution. Run on spot instances or batch queues, not on request-serving API pods. Output is immutable HLS/DASH segments back to object storage.
99% of bytes are video segments served repeatedly — edge caching cuts origin egress cost and startup latency. Adaptive bitrate manifests let players switch quality without new API calls.
Titles, channels, comments, and processing state need joins, indexes, and transactions. Sharded or regional when single-node Postgres saturates — not replaced by object storage.
Exact per-view counts are unnecessary; approximate aggregation at millions of events/sec is cheaper than row-level UPDATE on a video table.