“At Shopify I watched traffic climb from 1K to 1M RPS, and the thing that saved us was not adding new tech but building one dashboard that showed queuing time, worker utilization, and per-tenant latency sliced together. If you can't see where the spike hurts, you're fixing blind.”

“I’d stop guessing and instrument first: measure drain time, consumer headroom, and where retries are amplifying the spike. At Lambda scale, the backend usually falls over because backpressure and autoscaling are out of tune, not because raw throughput is impossible.”

“I’d stop guessing and instrument first: at JPM I saw a notification system that looked fine until spikes hit, and the dashboards we added exposed that autoscaling was lagging behind queue depth, not CPU. Get latency, queue length, and per-pod resource metrics in front of you before you redesign anything.”

“We used to get obliterated every time a seller did a flash drop, request queuing through the roof on 2x Heroku dynos. We hired a random Rails guy named Sam who fixed it in an hour by killing some atrocious N+1s and caching our JS vendor folder, so I'd start there before you rewrite anything.”

“I would stop guessing and look at the spike first: at TuLotero, our push service fanned out ~1M notifications per broadcast, and the real bottlenecks were always the database connection pool or a synchronous downstream call, not CPU. Profile one failing path, fix that, then move to caching and autoscaling.”

“At Metro Markets, adding Redis caching on our homepage paths cut traffic costs by about 60% and took the spike pressure off the backend. I'd instrument first with Grafana or whatever you have, find the hot path, and cache that before rewriting anything.”

“I would start by instrumenting the spike path so you actually know where the time goes, because at Airbnb for Work the fixes that looked like scaling problems were usually N+1 queries, missing indexes, or a single hot shard that nobody had named yet.”

“I spent too many nights paging on database CPU when the real problem was synchronous batch jobs hammering the same API our users needed. If your backend collapses on traffic spikes, decouple your heavy or batch work into queues with async consumers so the real-time path stops absorbing bursts it was never built for.”

“I had a product hit 500K views in a few hours and got DDoS'd into the ground, which forced me to reverse-engineer Cloudflare WAF rules to stay up. For you, I'd put rate limits and a WAF in front of everything first, because spikes will find the unprotected endpoint every time.”