Break the Backend
Turn on a fault and read what moved. Several completely different faults raise p99, so p99 diagnoses nothing on its own — the skill is finding the signal that separates them, and this page is built to make that the only way through.
Every reading below is produced by a model inside Engineer Atlas, and every value is a multiple of that signal's healthy baseline — 3.0x means three times normal, not three milliseconds. Nothing here was measured on a real system, and a real system would not give you these numbers anyway. What transfers is the shape: which signal moves, and which one tells two faults apart.
Signal readings
Relative to the healthy baseline. Read down the column, not across the headline.
| Signal | Now | Healthy | Where you read it |
|---|---|---|---|
| p99 latency | 1.0x | 1.0x | Request metrics at the edge |
| 5xx rate | 0% | 0% | Request metrics at the edge |
| Completed requests | 1.0x | 1.0x | Request metrics at the edge |
| DB query time | 1.0x | 1.0x | Database client instrumentation |
| Cache hit rate | 90% | 90% | Cache client metrics |
| Requests waiting for a connection | 1.0x | 1.0x | Connection pool gauge |
| Event-loop lag | 1.0x | 1.0x | Runtime metrics |
| External call time | 1.0x | 1.0x | Outbound HTTP client spans |
| External call errors | 0% | 0% | Outbound HTTP client metrics |
| Queue depth | 1.0x | 1.0x | Broker metrics |
| Healthy instances | 3 | 3 | Load balancer / orchestrator |
| Duplicated side effects | 0 | 0 | Business data — not an infrastructure metric |
What an engineer would notice
In roughly the order they would notice it.
Nothing. Every signal is at its healthy baseline — turn on a fault to the left.
Diagnose it
Name the cause from the signals before you open the answer. Getting the right answer from the wrong signal does not transfer to the next incident.
Nothing is wrong. Every signal is at its healthy baseline.