Block · ops
Latency Percentiles
p50 / p95 / p99 request latency over time, auto-flagging the worst p99 deviation by z-score in a plain-language narrative.
State
Size
Mode
API Gateway · request latency
297 ms
p99 · latest
- p50
- p95
- p99
p99 latency spiked to 536 on May 21, 2026, 4.1σ above a 285 baseline — driven by a deploy regression.
APM telemetry · updated every 60s
Demo data is illustrative. Replace with your own typed data prop.
How it computes
p50 / p95 / p99 latency lines over time, with z-score anomaly flagging on the p99 series.
zᵢ = (p99ᵢ − μ) / σ; flag when |zᵢ| ≥ threshold (default 2σ)Assumptions
- Percentiles are precomputed upstream, not recomputed from raw latencies.
- Anomaly detection runs on p99 only; p50/p95 are context.
- The p99 series is roughly Gaussian for the z-score to be meaningful.
Honest about
- When all p99 readings are equal, σ = 0 ⇒ no anomalies (no false positives).
- Only the single highest-|z| point drives the narrative even if several breach the threshold.
- The narrative's suggested cause ("a deploy regression") is a fixed hint, not inferred from the data.