Incident Metrics Examples: What to Track and How to Read Them

Concrete incident metrics examples with definitions, sample numbers and a ready dashboard layout, plus how to avoid the traps that make MTTR and counts misleading.

The incident metrics worth tracking fall into four groups: timing (detect, acknowledge, resolve), frequency and severity, alert quality, and on-call load. Below are concrete examples of each, with definitions, a sample calculation, a dashboard layout you can copy, and the statistical traps that make a single number lie.

Uptime Beacon
uptimebeacon.eu
Reduce False Alarms on Call
Uptime Beacon confirms repeated failures from a second location on paid plans, helping teams distinguish incidents from network blips.
Explore Uptime Beacon

TL;DR

  • Start with incident count by severity and service. Add timing metrics (MTTD, MTTA, MTTR) only once your timestamps are consistent.
  • Report timing at a percentile (median and p90), not only as a mean. Google’s SRE team warns that MTTx averages can be dangerously misleading.
  • Track alert noise and pages per person alongside speed. Fast response means little if the pager is crying wolf.
  • Pair every metric with an owner and a written definition, or the trend line is meaningless.
  • Fewer false alarms upstream improves almost every metric below.

Table of Contents

The timing metrics: detect, acknowledge, resolve

Timing metrics measure how fast your team moves through the life of an incident. Atlassian’s guidance on incident management KPIs lists mean time to acknowledge (MTTA), mean time to detect (MTTD), average time to resolve and average time between incidents among the useful measures, along with incident timestamps as the raw data behind them.

Here is how each reads in practice, using one imaginary outage on an API:

Metric What it measures Example timeline
Time to detect (MTTD) Failure starts to monitoring notices Failure begins 02:00, alert opens 02:04: 4 minutes
Time to acknowledge (MTTA) Alert opens to a human takes ownership Alert 02:04, engineer acknowledges 02:09: 5 minutes
Time to mitigate (MTTM) First impact to customer impact stops Rollback done 02:25: 25 minutes
Time to resolve or recover (MTTR) First impact to full recovery Root fix deployed 03:10: 70 minutes
Time between incidents (MTBF) Gap from one incident to the next Previous incident 11 days earlier

The definition matters more than the name. Google’s Incident Metrics in SRE report defines MTTR as the mean duration from first product impact to recovery, and MTTM as the mean duration from first product impact to mitigation. Note where the clock starts: at first impact, not when someone noticed. If your clock starts at acknowledgment, you are hiding your detection time.

Teams use “R” to mean repair, recovery, resolve, respond and restore. Our own write-up on one outage with three different MTTRs shows how the same incident produces different numbers depending on the definition. Pick one, write it down, and keep it.

The timing metrics: detect, acknowledge, resolve

Benchmarks, with a caveat

One vendor, Automation Anywhere, publishes target benchmarks of MTTD under 5 minutes, MTTA under 15 minutes and MTTR under 60 minutes. Treat these as a starting conversation, not a grade. A team running a payment API and a team running an internal reporting tool should not share targets. Set yours from your own baseline and your service level objectives (we cover that in how to set service level objectives that work).

Detection deserves its own line

Of all timing metrics, detection is the one a monitoring tool controls directly. Material from SREcon Europe 2016 reports an operational observation that incidents not detected by monitoring took 10 to 15 times longer to mitigate. That is one organization’s experience, not a law, but it explains why the same slides recommend tracking the percentage of outages that are auto-detected. If customers tell you about outages before your monitors do, that percentage is your first improvement target.

Frequency and severity metrics

Speed metrics say nothing about how often you are in trouble. Google’s report makes this point sharply: doubling the number of incidents while keeping a similar duration distribution can worsen reliability without materially changing MTTR. So frequency must be reported separately.

Examples to put on a dashboard:

A caution from Metabase: trend count by severity and service rather than as one blended number, and remember that fewer incidents may reflect under-reporting rather than better reliability. If your count drops the week after you made declaring an incident feel heavy, that is not a win.

A worked example

Say a small SaaS team logs these over a four-week month:

Week Sev 1 Sev 2 Sev 3 Total
1 0 2 5 7
2 1 1 4 6
3 0 3 6 9
4 0 1 3 4

Total: 26 incidents, one Sev 1. The headline “incidents fell from 9 to 4” is tempting, but with samples this small the useful question is which service produced the Sev 2s in week 3, and whether any were repeats. The table is illustrative, not data from a real team.

Alert quality metrics

This group is the one most teams skip, and it is where on-call pain lives.

High noise inflates every timing metric. Responders learn to wait before acknowledging, so MTTA rises, and real incidents hide among false ones. This is why we built Uptime Beacon around fewer false alarms: an incident opens only after several failures in a row, and on paid plans it is confirmed from a second location before anyone is paged. For the tactics behind that, see how to reduce alert noise without missing real incidents.

On-call health metrics

Response speed comes from people, and people burn out. The SREcon 2025 Americas slides list pages per person, especially overnight, length and frequency of on-call shifts, and perceived fairness of on-call schedules. Metabase’s dashboard likewise includes pages per on-call shift and escalation rate. Atlassian lists on-call time among its suggested incident metrics.

A simple example: if one engineer received 14 overnight pages in a month and the other three received 2, 1 and 0, your average of 4.25 looks fine and your rotation is not. Always look at the distribution per person.

Uneven overnight pages across four engineers

Follow-through metrics

An incident is not finished at recovery. Useful governance examples:

  • Postmortem completion. Metabase suggests tracking postmortem completion for high-severity incidents.
  • Action item completion, one of the measures the SREcon 2025 material recommends alongside postmortem completeness.
  • Time to customer messaging, from the same slides. It measures how long customers waited to hear anything.
  • Percentage mitigated within 30 minutes and percentage internally detected, which the GitLab handbook uses alongside MTTR to track incident performance against targets.

If you need a format for the review itself, our incident postmortem template is a copy-paste starting point, and the root cause analysis guide covers verifying what actually went wrong.

Why a single number misleads

This is the part most listicles skip. Google’s SRE report cautions that MTTx metrics can be dangerously misleading in typical SRE settings and says conclusions should rest on a whole body of incidents, not a single event. Incident durations are skewed: most are short, a few are very long, and the long ones dominate the mean.

The SREcon 2025 slides go further. They note that MTTR as a mean is not a robust summary statistic, and that, per Štěpán Davidovič’s analysis, with typical incident counts and spread, measured changes in MTTR are almost deterministically noise. A drop from 62 to 55 minutes quarter over quarter probably tells you nothing.

Practical responses:

  1. Report medians and a high percentile. Metabase’s dashboard pairs median MTTA with p90 and median MTTR with p90. The 2016 SREcon slides likewise prefer a stated percentile over a bare mean.
  2. Show the distribution. A simple histogram of durations beats a single figure.
  3. Segment. Split by severity and service before comparing periods.
  4. Read the biggest incidents individually. The numbers tell you where to look, the postmortems tell you why.
  5. Match the statistic to the question. Google’s perspective is that no universal MTTx number reliably summarizes reliability, so choose the measure for the decision in front of you.

A starter dashboard you can copy

Do not build all of the above on day one. FireHydrant advises starting with incident count per day, week or month, then adding days and times of occurrence, affected services and who is most often involved, before moving to MTTR and its components.

A sensible progression:

  1. Month 1: incidents per week by severity and service; currently open incidents.
  2. Month 2: add detection time and acknowledgment time, with median and p90.
  3. Month 3: add alert-noise rate and pages per person per shift.
  4. Month 4 onward: add postmortem and action completion, SLO compliance, and repeat incidents.

On SLO compliance, Metabase suggests reporting periods meeting the published SLO target divided by all periods. If you promise uptime to clients, this is the number they will ask about.

Before you start, gather a baseline. Gremlin’s guide, written for chaos engineering but useful anywhere, lists alert counts by service per week, time to resolution per service, incidents per week by SEV level and MTTD, MTTR and MTBF by service. Daydream adds a governance point: standardize definitions and timestamps and name an owner for incident response metrics. Without an owner, definitions drift and the trend breaks.

For more on what to report and to whom, see our guide to incident reporting metrics.

Perspective

Most metric programs fail at the first timestamp. If detection is slow or noisy, everything downstream is skewed: acknowledgment times stretch because responders have learned to hesitate, and the incident log fills with events that were never real.

That is why we put false-alarm reduction first. Uptime Beacon opens an incident only after repeated failures, and on paid plans confirms from a second region, so the incidents you count are incidents. Your noise rate falls, your pages per person fall, and your timing metrics start to mean something.

The second change is who does the analysis. With MCP on every plan, including Free, you can connect Claude, ChatGPT, Cursor, VS Code or Codex and ask it to read your monitors, incidents and check history, then diagnose a failing monitor region by region. Pulling the raw data for a weekly review becomes a question instead of a spreadsheet exercise. Changes such as pausing monitors or acknowledging incidents are off until a team turns on “Allow changes”, and nothing can be deleted over MCP.

Our advice: count fewer, truer incidents, report percentiles, and read the worst ones in full.

Aaditya Parashar

Cleaner data from the first check

If you want these metrics built on incidents you can trust, Uptime Beacon’s Free plan gives you 5 monitors checked every 5 minutes from EU Central, one alert channel (email or Slack) and one status page, with no card. Paid plans start at €14.99 per month for Starter, and add second-location confirmation of failures, so a network blip does not become a Sev 2 in your log. Pro is €39.99 per month and Business is €89.99 per month. Every result is stored in the EU.

Uptime Beacon

Running APIs and cron jobs? See API and cron monitoring for SaaS teams, and for how AI assistants fit into a response, read our piece on AI incident response for on-call teams.

FAQ

What are the most common incident management metrics?

The most common are number of incidents, MTTD, MTTA, MTTR, MTBF and uptime. Atlassian’s KPI guide lists these along with alerts created, on-call time and SLA or SLO measures. Start with incident count by severity, then add timing once your timestamps are consistent.

Is MTTR a good metric?

It is useful as one input, but poor as a lone target. Google’s SRE research says MTTx metrics can mislead, and that a mean is easily dominated by a few long incidents. Report median and p90, and read the longest incidents individually.

How do I measure alert noise?

Divide alerts that never became incidents by all alerts over the same period, as Metabase defines it. Break it down by service to find the worst sources. A falling rate means fewer wasted pages.

What should a small team track first?

Incidents per week by severity and service, plus time to detect and acknowledge. Add noise rate and pages per person next. Leave the long tail of governance metrics until the basics are reliable.

Sources

Put your site on it. Sleep through the blips.

Add your first monitor now. If your site is fine, you will not hear from us.

Monitor it free
Sign up with GoogleOpening Google…
Free plan, no cardSign up with Google or email