How to track API response times and get alerted when they slow down
Set up an API monitor, read p50, p95 and p99 response times from several regions, and get a warning when your API gets slow, before it starts timing out.
· 6 min read · By the Spot Downtime team
APIs rarely go from fast to down in one step. They get slower first: a missing index, a growing table, a dependency that's struggling. If you can see that slope, you can fix it before it becomes an outage. This guide shows how to track your API's response times in Spot Downtime and get warned when they climb.
Step 1: add an API monitor
- Open Monitors, click Add monitor and choose API.
- Pick the Method and paste the URL, for example
GET https://api.example.com/v1/health. - If the endpoint needs a key, add it under Headers, one per line:
Authorization: Bearer …. Headers are visible to everyone in the workspace, so use a read-only key. - Set the Expected status if it's something specific like
200. Leave it empty to accept any 2xx or 3xx. - Choose Check every: 5 minutes on Free, down to 1 minute on Starter and 30 seconds on Pro. More checks give you a smoother, more precise picture.
Which endpoint should you time?
A health endpoint that returns {"status":"ok"} without touching anything tells you little about speed. Pick a request that does real work, like a read from the database, a search, or listing recent orders, so its timing moves when your users' does. Add JSON assertions such as data.items.length greater than 0 so a fast but empty answer still counts as down. Our API monitoring guide has more on what to assert.
Step 2: check from where your users are
Response time depends on distance. An API in Virginia might answer in 80 ms from California and 250 ms from Singapore. On Starter and Pro, use Check from to pick the regions that match your users: California, Amsterdam, Singapore, or several at once. With several, each region checks every interval, so you see how the API performs for everyone.
Step 3: read the response time chart
Open the monitor. The Response time card charts every successful check over 24 hours, 7, 30 or 90 days, and lists the last 24 hours as:
Response time
- Average · 24h
- 196 ms
- p50 · 24h
- 184 ms
- p75 · 24h
- 203 ms
- p95 · 24h
- 318 ms
- p99 · 24h
- 412 ms
| Figure | What it tells you |
|---|---|
| Average | A rough summary. Easily pulled up or down by a few outliers. |
| p50 | The median: half of checks were faster. Your typical response time. |
| p75 | Three in four checks were faster. |
| p95 | The time 95% of checks answered within. The slowest 1 in 20 took longer. |
| p99 | The slowest 1 in 100. Where timeouts start to show up first. |
Click p95 in the chart's legend to draw the p95 line over the individual checks. Failed checks aren't counted in these figures, since a timeout has no response time; they show up as downtime instead.
Why p95 and not the average?
Patterns to look for
- A slow upward slope over weeks: data growth. Look for queries that scan whole tables.
- A step up on a particular day: a deploy or config change. Compare with your release log.
- Spikes at the same time every day: a cron job, backup or batch import competing for the database.
- p50 steady but p95 rising: some requests are hitting a slow path, like a cold cache, lock contention or a slow dependency.
- One region much slower than the others: a CDN or routing issue, or a sign you need a server closer to those users.
Step 4: get alerted when it gets slow
On Starter and Pro, edit the monitor and tick Alert when it's slow. Then set:
- p95 response time above (ms): your threshold.
- Measured over: 5, 15 or 30 minutes, or 1 hour. The window has to hold at least 3 checks, so a monitor that checks every 5 minutes needs at least 15.
You get a warning when the p95 over that window crosses the threshold, and another when it's fast again. A slowdown isn't counted as downtime and isn't shown on status pages, and slowness warnings aren't sent by text, call or WhatsApp, so they never spend SMS credits. Every slowdown is listed on the monitor with its worst p95.
Choosing a threshold
- Let the monitor run for a few days, ideally including a busy day.
- Note the 24-hour p95 on a normal day.
- Set the threshold to about 1.5 to 2 times that. If your normal p95 is 400 ms, start at 700 ms over 15 minutes.
- If it fires on days you wouldn't act on, raise it or lengthen the window; if a real slowdown slipped past, lower it.
Requests time out after 10 seconds and count as down, so keep the threshold well below that to get the warning first.
It works for more than APIs
Website, TCP, UDP, ping, DNS and LLM API monitors all record response times and support slowness alerts in the same way. A database port that takes longer and longer to accept connections, or an AI model whose replies are getting slower, is worth knowing about too.
If you use Claude, the Spot Downtime connector can pull these numbers for you: ask “how have the checkout API's response times changed this week?”
More in this series
Get new guides by email
Practical posts on monitoring, outages, error fixes and on-call, when they're published. No spam.
Keep reading
- How-to · Monitoring · APIsHow to configure gRPC health checkingLearn how to implement and configure gRPC health checking for reliable service discovery and load balancing: server status, Kubernetes probes, client-side health checks, Envoy, cloud load balancers and Consul.November 7, 2026 · 7 min read
- How-to · APIs · MonitoringHow to monitor an LLM API (OpenAI, Anthropic, Gemini) in Spot DowntimeA step-by-step guide to LLM API monitoring: use a separate key, pick the model, set the prompt and time limit, and keep the cost of checks to a fraction of a cent.November 2, 2026 · 5 min read