Backend & APIs
Scaling an API platform to 40M requests a day at 38ms
Tracklane's analytics API was falling over at 4M requests a day. We rearchitected ingestion, storage, and caching to handle 40M with lower latency and a smaller bill.
0M
requests handled daily
0ms
p99 response time
0%
lower infrastructure cost
01 · The challenge
Tracklane sells an events API to other startups, which means their uptime is everyone's uptime. At 4M requests a day the p99 latency had crossed 800ms, the database was pegged, and every traffic spike meant paged engineers at 3am.
The catch: no downtime allowed and no breaking changes. Hundreds of customer integrations depended on the exact behavior of the existing API.
02 · The solution
We rebuilt the write path as an ingestion pipeline: requests land on a lightweight edge service, stream through Kafka, and batch into a columnar store built for analytics queries. Reads moved behind a two-layer cache with smart invalidation.
The migration ran in shadow mode for three weeks, serving both stacks in parallel and diffing responses on live traffic, then cut over customer by customer with zero downtime and zero contract changes.
03 · How we did it
Measure before touching
Two weeks of profiling to find the real bottlenecks, because guessing at scale problems is how rewrites fail.
Split reads from writes
Kafka-backed ingestion decoupled spiky writes from analytics reads, each scaling independently on its own hardware.
Shadow the migration
Both stacks ran in parallel on live traffic with automated response diffing, so the cutover was proven before it happened.
Load-test past the goal
Synthetic load at 3x the target volume before launch, plus dashboards and alerts wired to the on-call rotation.
“They took us from nightly pages at 4M requests to sleeping through 40M. Same API, same contracts, ten times the headroom.”
Want results like this?
Tell us where your numbers are today and we will show you what they could look like.