Replit

Replit is the fastest way to build and deploy software. Millions of developers run code on Replit's infrastructure every day, which means the platform team is operating one of the busiest multi-tenant runtimes on the internet.

The challenge

At Replit's scale, "a few users are affected" is still thousands of people. The platform team needed to catch a degrading region or a noisy-neighbor container before it became a trend on social media — but the signal was buried in a volume of logs no human could read in real time.

The old model was reactive: an alert fired, someone woke up, and the investigation started from zero at 3am.

The solution

Replit fed its runtime and scheduler logs into Sazabi and set it loose on the firehose continuously, surfacing anomalies as questions answer themselves.

Sazabi turned our pager rotations into conversations. Instead of waking up to a cryptic alert, we wake up to a diagnosis — and half the time it's already paged itself before a single user noticed.

The shift was from reactive to ahead-of-it: most regressions are now caught while they're still measured in single-digit affected sessions.

Results

• 68% of incidents detected before the first user report
• 3am pages down 45% — most issues surface and resolve during business hours
• 2 minutes average from anomaly to named root cause