How VectorGrid absorbed a 50× launch spike without a single page
An analytics platform scaled from a steady baseline to fifty times its normal traffic on launch day, then back down, with no manual steps.
Results
Launch day, by the numbers
- Peak traffic
- ×
- Over the weekly baseline, within two hours.
- Pages
- 0
- No on-call alerts during the launch.
- Lower compute cost
- %
- Scaling down overnight instead of running at peak.
- p95 latency
- ms
- Held steady through the spike.
The challenge
VectorGrid’s dashboards are used by thousands of teams, and a public launch was expected to bring a large but unknown spike. The previous setup ran at a fixed size, so the team either paid for peak capacity all month or risked falling over on the day that mattered most.
The solution
They set scaling rules on CPU and request latency for each web service, with a minimum instance count for the morning traffic and a ceiling to cap spend. Query workers moved behind private networking so they could scale on queue depth without exposing an endpoint.
Before launch, the team replayed production traffic at 60 times the normal rate against a staging copy to confirm the limits held.
The results
On launch day traffic climbed to 50 times the baseline within two hours. New instances came up in seconds, latency stayed flat, and no one was paged. By the next morning the services had scaled back down on their own, and the month’s compute bill came in below the old fixed setup.
“Autoscaling carried us through a launch that brought fifty times our normal traffic. Nobody got paged, and the bill stayed predictable.”
