Prompt
Do I need alerting for slow requests?
Latest observation
Nov 19, 2026
OpenAI APIWeb search: off
Usually, yes — if slow requests can hurt users, revenue, or upstream systems, you should have alerting for them.
A good rule:
- Monitor slow requests: always
- Alert on slow requests: when they’re likely to indicate a real incident, not just normal variance
When alerting is useful
Alert if slow requests can mean:
- user-facing degradation
- timeouts or failed checkout/login flows
- queue buildup or cascading latency
- a dependency problem
- resource exhaustion
When alerting can be noisy
Be careful if:
- latency naturally varies a lot by time of day
- occasional slow requests are expected
- you only alert on a few outliers and not the overall impact
- the threshold is arbitrary and not tied to SLOs
Better than “alert if any request is slow”
Use alerts based on:
- percentiles: p95/p99 latency
- error budget / SLOs
- rate of slow requests over a window
- sustained latency regression, not one-off spikes
Example:
- Alert if p95 latency > 500ms for 10 minutes
- Alert if more than 5% of requests exceed 2s for 15 minutes
- Alert if latency is causing SLO burn rate above a threshold
Practical approach
- Set a latency SLO for important endpoints.
- Dashboards for p50/p95/p99, plus slow-request rate.
- Alert only when latency crosses a sustained and meaningful threshold.
- Include context: endpoint, region, dependency, deploy version.
If you’re not sure
Start with:
- dashboards first
- one alert on high sustained p95/p99 latency
- one alert on timeout or slow-request rate above a threshold
If you want, I can help you design a simple alerting strategy for your service (e.g. by endpoint, p95, or SLO-based).