2 min lesson
Scale and stakes at 1M+ DAU
Walk through each part of "Scale and stakes at 1M+ DAU", then explain what each one does.
Step 1 of 2
Over a million daily active users changes which problems are real. At this scale, hot paths, caching and database scaling stop being optimizations you get to later and become the first thing you design around. Answers that ignore scale read as junior here.
- Auth
- Runs on every request, so its latency and availability budgets are extreme - it can't be the slow dependency.
- Webhooks
- Bursty and high-throughput; you design for backpressure and graceful degradation, not the happy path.
- Agent backend
- Long-running LLM jobs that spike under load; queueing and backpressure decide whether it stays up.
- Cost
- Agent/LLM backends have real per-request economics, so reliability trade-offs are measured in dollars, not just nines.
Auth is the clearest example. If it sits on every request, then a 50ms regression there is a 50ms tax on the entire product and an availability dip is a full outage no matter how healthy the rest of the system is. That is why caching strategy and read-path design for auth are not afterthoughts at this scale.
The agent backend introduces a dimension most backend roles never touch: the dollar cost of a single request. LLM calls are expensive, so a reliability decision like "retry three times" is also a cost decision and a capacity decision like "keep warm headroom" burns money while idle.
Latency, complexity and engineering time.
More nines usually means more redundancy and more code.
Each retry, each warm pool, each replica costs LLM spend or compute.
The right answer balances nines and per-request economics together.