Be honest — how much of your system design work is gut feel vs actual numbers?
I've been doing backend / system design for a while and there's something about my own process I keep noticing that I'd love to hear other people's perspective on.
Whenever I'm sketching a new service or reviewing someone else's design, I'm constantly making calls based on rough estimates I'm pulling out of my head. "Postgres can probably handle ~5k QPS for this workload." "A queue here will smooth out the spike." "If this service dies, retries should cover us."
Some of those numbers come from benchmarks I actually ran once. Some are from a blog post I half-remember. Some are, honestly… vibes. The unsettling part is I usually can't tell which is which until prod tells me.
So I'm curious how other people deal with this. A few things I'd really like honest answers on:
- Do you keep a personal "cheat sheet" of approximate numbers — typical Redis hit rates, RDS connection limits, gRPC vs REST overhead, Kafka throughput, etc.? Or do you re-google every time? Or just trust your memory and hope?
- When you propose an architecture in a design review, how confident are you in your latency / throughput / capacity numbers? Like 90% confident, 50%, "we'll find out in staging"?
- Has there been a project where you shipped something, watched the metrics, and the actual bottleneck was somewhere completely different from where you expected? What did you assume vs what was actually true?
- When you're choosing between two options — message queue vs direct call, single DB vs read replicas, sync vs async, monolith vs split — how do you actually make the call? Back-of-envelope math? Past scars? "Pick one and refactor later"?
- Do you do anything to validate a design before writing code — load tests on a prototype, capacity spreadsheets, sketches in Excalidraw with annotations — or is the first real signal always staging or prod?
- And for tech leads / architects specifically: when you have to convince a PM, the business, or skeptical seniors that your design will hold up, what do you actually show them? "Trust me" rarely works.
Not pitching anything, genuinely trying to understand what people do. The more embarrassingly honest the answer the better — I have a hunch we all fake the confidence more than we admit, and it'd be kind of nice to normalize that.
What's the dumbest assumption you ever made in a design review that bit you later?