At a glance
- Did you know
- Document databases store nested JSON-like records, so one read can fetch a whole aggregate.
- Rule of thumb
- Model around your queries; unbounded arrays inside one document become slow reads.
- Watch out
- Index write amplification; large documents slow every read.
Its settings in ScaleMySystems
What you can change on it in the lab, and what each number does.
- Concurrency (slots)
- Calls it can work on at once (per instance when autoscaling). Needed slots ≈ calls per second × seconds each takes. Extra calls wait in the queue, then are rejected.
- Queue capacity
- Calls allowed to wait for a free slot. A bigger queue absorbs bursts but adds waiting time; 0 rejects straight away.
- vCPU
- CPU cores available. Needed ≈ calls per second × CPU per request ÷ 1000. Above about 90% busy it slows down and may crash.
- CPU per request (ms)
- CPU time each call burns. Part of what the component is; make the code cheaper to lower it.
- Latency (p50) (ms)
- Typical time one call takes here, before any calls it makes itself. Slots are held for this long.
- Latency (p99) (ms)
- The slowest 1% of calls. Callers must set their timeout above this or they cut off healthy calls.
- Timeout to dependencies (ms)
- How long it waits for each call it makes before giving up. Set it above the dependency’s p99 but low enough that a dead dependency does not hold every slot.
- Error rate
- Share of calls that fail on their own (bugs, declines, flaky tools). An agent that calls it on every step multiplies this.
- Quota (req/min)
- Calls allowed per minute (an API key or provider limit). Calls over it get 429s. Needed = calls per second × 60, with headroom.
Problems that start with it
- YouTube end to end 20,000 requests/s routed by path: watch, search, comments and uploads. Metadata cache in front of Vitess MySQL, recommendations, view events into Kafka, uploads to storage and a transcoding queue.
- Netflix: from launch build to real traffic Netflix without Open Connect or EVCache: every request reaches the origin, title metadata is read from the catalog database each time, and the playback, browse, license and recommendation services are an eighth of their real size. 6,000 requests a second: half are playback starts, 45% browsing, 5% DRM licenses. Playback starts fail.
- ChatGPT: from launch build to real traffic A chat product on launch-week hardware: the GPU prefill and decode pools, the KV cache store, moderation and the chat API are an eighth of the size real traffic needs. 300 chats a second arrive with long prompts. Answers queue for 20 seconds and most fail.
- Twitch: from launch build to real traffic Live streaming as first built: no manifest cache in front of the live streams catalog, and the playback and chat APIs, chat fan-out, live ingest and transcoders at an eighth of their real size. Real viewers arrive, watching and chatting, and three in four requests fail.
- Amazon storefront: from launch build to real traffic A storefront as first built: product pages read DynamoDB with no product cache, and the product, cart and checkout APIs, search, recommendations, the orders database, payments and fulfilment workers at an eighth of their real size. Shoppers arrive at real traffic and half of all requests fail, checkout included.
- YouTube: from launch build to real traffic YouTube as first built: no media CDN, no metadata cache in front of Vitess MySQL, and the watch, search, comment and upload services, recommendations, view counting and transcoding at an eighth of their real size. 20,000 requests a second arrive, routed by path. Almost everything fails.
Full lesson coming soon
Lessons are being added one component at a time: how it works, its settings, a live exercise to break and fix it, interview questions and a short check. See the Circuit breaker lesson for what one looks like.