---
title: "MTBF (Mean Time Between Failures): Formula, Calculation, and How to Improve It"
author: "Talha Bektas"
date: 2026-06-22T06:36:12+00:00
updated: 2026-06-22T06:36:12+00:00
canonical: https://www.itoc360.com/mtbf-mean-time-between-failures/
---

# MTBF (Mean Time Between Failures): Formula, Calculation, and How to Improve It

> **Quick answer:** MTBF (mean time between failures) is the average time a repairable system operates between one failure and the next. **MTBF = Total operating time ÷ Number of failures.** A service that runs 2,156 hours in a quarter and fails 6 times has an MTBF of about 359 hours (15 days). A higher MTBF means a more reliable system: it measures how rarely you fail, while MTTR measures how fast you recover.

Your incident response is fast. MTTA under five minutes. MTTR under 30. But your system goes down every two weeks.

That is a low-MTBF problem — and no amount of faster incident response will fix it. When failures are frequent, the issue is the system itself: unstable infrastructure, recurring bugs, fragile dependencies or insufficient capacity planning. MTBF is the metric that makes this visible before your SLA reports do.

This guide covers what MTBF measures, the formula, worked examples, how to calculate MTBF in Excel, how to convert between MTBF, failure rate, FIT and AFR, what MTBF does and does not tell you about reliability, and the engineering work that moves the number in the right direction.

## **What Is MTBF?**

MTBF, or mean time between failures, is the average elapsed operating time between one failure of a repairable system and the next. It is sometimes written as **mean time before failure**; the meaning is the same.

The "mean" matters. MTBF is an average across many failures over a measurement period, not a guarantee about any single interval. A system with an MTBF of 30 days will not fail exactly every 30 days — it might run 60 days without an incident, then fail twice in one week.

MTBF applies to **repairable** systems: things you restore to service after a failure, such as software services, servers, networks and databases. For components you replace rather than repair — a disk, a sensor, a light bulb — the matching metric is MTTF (mean time to failure).

In reliability engineering terms, MTBF is the inverse of failure rate. A system that fails on average once every 100 hours has a failure rate of 0.01 failures per hour and an MTBF of 100 hours.

\[FIGURE 1 — existing diagram: MTBF is the uptime between the end of one repair and the next failure; MTTR is the repair window.\]

## **MTBF Formula**

> **MTBF = Total operating time ÷ Number of failures**

How you measure "total operating time" depends on what you are measuring:

| **Use case** | **Total operating time** | **Example** |
| --- | --- | --- |
| **One repairable system** (a service, a server) | Calendar time in the period **minus downtime** | 90 days of a payment service, minus 4 hours of outages |
| **A fleet of identical units** (hardware, devices) | Units × operating hours per unit | 500 drives × 2,000 hours each = 1,000,000 unit-hours |

Two related formulas complete the picture:

> **Failure rate (λ) = 1 ÷ MTBF**
>
> **Availability = MTBF ÷ (MTBF + MTTR)**

## **How to Calculate MTBF (Step by Step)**

1. **Define what counts as a failure.** For software services this is usually any P1 or P2 incident, or any incident with user-visible impact. The exact definition matters less than using the same one every period.
2. **Choose a measurement window.** Rolling 90-day windows are common: long enough to smooth out a bad week, short enough to act on.
3. **Measure total operating time.** For a single system, subtract downtime from calendar time. For a fleet, multiply units by operating hours.
4. **Count failures** in the same window.
5. **Divide** operating time by the number of failures.

### **Example 1: MTBF for a software service**

A payment processing service runs for 90 days and has 6 failures. Total downtime across all six incidents is 4 hours.

- Calendar time = 90 days × 24 hours = 2,160 hours
- Operating time = 2,160 − 4 = 2,156 hours
- MTBF = 2,156 ÷ 6 = **359.3 hours** (about 15 days between failures)

If the team closes its postmortem action items and the next quarter has only 3 failures with similar downtime, MTBF roughly doubles to about 720 hours — a reliability gain that no amount of faster incident response could produce.

### **Example 2: MTBF for a fleet of hardware units**

500 servers each run 2,000 hours. During that time, 4 of them fail.

- Total operating time = 500 × 2,000 = 1,000,000 unit-hours
- MTBF = 1,000,000 ÷ 4 = **250,000 hours**

This is how manufacturers and large fleets measure MTBF: across many units, not by waiting for one unit to fail repeatedly.

### **Example 3: MTBF for one month of incidents**

A core API runs for a 30-day month (720 hours) and has 3 failures with 2.5 hours of total downtime.

- Operating time = 720 − 2.5 = 717.5 hours
- MTBF = 717.5 ÷ 3 = **239.2 hours** (about 10 days)

### **What counts as a "failure"?**

Teams get inconsistent MTBF numbers mainly because they count failures differently. Common definitions are any P1 or P2 incident, any incident that pages on-call, or any incident with user-visible degradation. Pick one, write it down, and apply it every period so trends stay comparable.

## **How to Calculate MTBF in Excel or Google Sheets**

*[Interactive calculator: mtbf-calculator]*

Put each failure in a row with its start time in column **B** and its resolved time in column **C**. Put the window start in **F1** and the window end in **F2**.

1. Total downtime in hours: `=SUMPRODUCT(C2:C50-B2:B50)*24`
2. Operating time in hours: `=(F2-F1)*24 - SUMPRODUCT(C2:C50-B2:B50)*24`
3. Number of failures: `=COUNT(B2:B50)`
4. MTBF in hours: `=((F2-F1)*24 - SUMPRODUCT(C2:C50-B2:B50)*24) / COUNT(B2:B50)`
5. Failure rate per hour: `=1/MTBF_cell`

Multiplying by 24 converts Excel's day-based date values into hours. The same formulas work in Google Sheets.

## **MTBF, Failure Rate, FIT and AFR**

Hardware datasheets and reliability reports often express the same idea in different units. Assuming a constant failure rate, they convert like this:

| **Metric** | **What it means** | **Formula from MTBF (hours)** |
| --- | --- | --- |
| **Failure rate (λ)** | Failures per operating hour | λ = 1 ÷ MTBF |
| **FIT** (failures in time) | Failures per 1 billion operating hours | FIT = 1,000,000,000 ÷ MTBF |
| **AFR** (annualized failure rate) | Probability a unit fails within one year of operation | AFR = 1 − e^(−8,760 ÷ MTBF) |
| **Expected failures per year** | Failures per unit per year of continuous operation | 8,760 ÷ MTBF |

**Conversion examples:**

| **MTBF (hours)** | **Failure rate per hour** | **FIT** | **AFR** |
| --- | --- | --- | --- |
| 100,000 | 0.00001 | 10,000 | 8.39% |
| 500,000 | 0.000002 | 2,000 | 1.74% |
| 1,000,000 | 0.000001 | 1,000 | 0.87% |
| 2,000,000 | 0.0000005 | 500 | 0.44% |

To go the other way: **MTBF = 1 ÷ λ**, **MTBF = 1,000,000,000 ÷ FIT**, and **MTBF = −8,760 ÷ ln(1 − AFR)**.

## **What MTBF Does — and Does Not — Tell You About Reliability**

### **MTBF is not a lifespan**

A drive rated at 1,000,000 hours MTBF will not last 114 years. MTBF describes the failure rate during a product's normal service life, across a large population. With 1,000 such drives running continuously, you should expect around 8.8 failures per year (1,000 × 8,760 ÷ 1,000,000).

### **The probability of surviving a given time**

If the failure rate is constant, the probability that a system runs for a time *t* without failing is:

> **Reliability R(t) = e^(−t ÷ MTBF)**

The surprising consequence: the probability of running for a full MTBF period without failing is only **e^(−1) ≈ 36.8%**. For the payment service in Example 1 (MTBF 359 hours), the probability of a 30-day stretch (720 hours) without a failure is e^(−720 ÷ 359.3) ≈ **13.5%**.

### **When the constant-failure-rate assumption breaks**

These formulas assume failures happen at a steady rate — the flat middle of the "bathtub curve". They do not hold during early-life failures (infant mortality) or at end of life (wear-out), when failure rates rise. For software, frequent deploys and changing traffic also make the rate less stable, which is why tracking the MTBF trend matters more than any single value.

### **Rated, field and demonstrated MTBF**

- **Rated (predicted) MTBF** is calculated by manufacturers from component models and standards such as MIL-HDBK-217F or Telcordia SR-332, before field data exists.
- **Field MTBF** is measured from real failures in deployed units. It often differs from the rated figure — public drive-failure reports such as Backblaze's Drive Stats show real annualized failure rates that vary widely by model.
- **Demonstrated MTBF** is proven through testing, stated as a lower bound at a confidence level (for example, "MTBF ≥ 50,000 hours at 90% confidence"). For a time-terminated test with total test time T and r failures, the lower bound is 2T ÷ χ²(confidence; 2r + 2).

For software services, your own incident data — field MTBF — is always more meaningful than any external figure.

## **MTBF vs MTTF vs MTTR**

| **Metric** | **Full name** | **Applies to** | **What it measures** | **Formula** |
| --- | --- | --- | --- | --- |
| **MTBF** | Mean time between failures | Repairable systems | How long a system runs between failures | Total operating time ÷ Number of failures |
| **MTTF** | Mean time to failure | Non-repairable components | How long a component runs before it fails for good | Total operating time ÷ Number of units |
| **MTTR** | Mean time to resolve / repair | Repairable systems | How long it takes to recover from a failure | Total resolution time ÷ Number of incidents |

**MTBF vs MTTF:** use MTBF for anything you repair and return to service (software, servers, networks); use MTTF for parts you replace. Some vendors publish "MTBF" for parts that are really measured by MTTF, which is a common source of confusion.

**MTBF vs MTTR:** MTBF measures how rarely you fail — a property of the system, improved by engineering work such as better architecture, testing, capacity planning and dependency hardening. [MTTR](https://www.itoc360.com/mttr-mean-time-to-resolve/) measures how fast you recover — a property of your process and tooling, improved by alert routing, runbooks and automation. The two strategies do not overlap: you cannot raise MTBF by improving incident response, and you cannot cut MTTR by making the system fail less often.

For all four core reliability metrics side by side, see the [MTTR, MTTA, MTBF and MTTD guide](https://www.itoc360.com/mttr-mtta-mtbf-mttd/).

\[FIGURE 3 — existing diagram: MTBF is the only pre-failure metric; MTTD, MTTA and MTTR measure what happens after a failure.\]

## **MTBF and Availability**

> **Availability = MTBF ÷ (MTBF + MTTR)**

Availability depends on both how rarely you fail and how fast you recover. The payment service in Example 1 has an MTBF of 359.3 hours and an average MTTR of 40 minutes (4 hours ÷ 6 failures), so its availability is 359.3 ÷ (359.3 + 0.67) ≈ **99.81%**.

\[FIGURE 2 — existing diagram: MTBF (reliability lever) + MTTR (response lever) = Availability.\]

## **What Is a Good MTBF?**

There is no universal "good" MTBF. The right target depends on your availability commitment and how fast you recover. Rearranging the availability formula gives the MTBF you need:

> **Required MTBF = MTTR × Availability ÷ (1 − Availability)**

**Required MTBF for common availability targets:**

| **Availability target** | **MTTR 15 min** | **MTTR 1 hour** | **MTTR 4 hours** |
| --- | --- | --- | --- |
| 99% | 25 hours (1 day) | 99 hours (4 days) | 396 hours (16.5 days) |
| 99.5% | 50 hours (2 days) | 199 hours (8.3 days) | 796 hours (33 days) |
| 99.9% | 250 hours (10.4 days) | 999 hours (41.6 days) | 3,996 hours (166.5 days) |
| 99.95% | 500 hours (20.8 days) | 1,999 hours (83.3 days) | 7,996 hours (333 days) |
| 99.99% | 2,500 hours (104 days) | 9,999 hours (417 days) | 39,996 hours (4.6 years) |

Read across a row and the trade-off is clear: to hold 99.9% with a 4-hour MTTR, a service can fail only about once every 166 days. Cut MTTR to 15 minutes and it can fail every 10 days and still meet the same target. This is why reliability teams work on both levers at once.

For hardware, compare manufacturers' rated MTBF with field data from your own fleet. For software, the most useful benchmark is your own trend: a service whose MTBF improves from 14 to 21 days over two quarters is getting more reliable, whatever an industry average says.

## **How to Improve MTBF**

Improving MTBF means reducing how often failures happen. It is an engineering problem, not an operational one.

### **Complete postmortem action items**

The single highest-leverage investment is closing [postmortem](https://www.itoc360.com/post-mortem-meeting/) action items before the next incident. The database connection pool that exhausts every six weeks and the memory leak that needs a weekly restart are failures that were probably identified once and left unresolved. Track action-item completion as a leading indicator of future MTBF. Eliminating recurring causes is the job of [problem management](https://www.itoc360.com/itil-5-problem-management/).

### **Reduce deployment-related failures**

In high-deployment environments, many incidents are caused by code changes. Canary releases, feature flags and blue-green deployments limit the blast radius of each change, and automated rollback on error-rate thresholds stops many bad deploys before they become full incidents.

### **Harden dependencies**

Many failures come from dependencies: slow upstream APIs, exhausted database connections, third-party outages. Circuit breakers, retries with exponential backoff, fallbacks and dependency health checks stop upstream problems from cascading into your service.

### **Plan capacity before it fails**

Disk, memory, connection pools and queues run out predictably. Alerting at 70–80% of capacity instead of at 100% prevents whole categories of incidents.

### **Use chaos testing to find weaknesses first**

[Chaos testing](https://www.itoc360.com/what-is-chaos-testing/) surfaces hidden assumptions — that failover is clean, that retries handle dropped connections — before production does.

### **Alert on leading indicators**

Rising error rates below incident thresholds, gradual memory growth and slowing queries often precede failures. Investigating them early catches failures in progress and extends MTBF.

## **How to Track MTBF in Practice**

- **Define failure events precisely** — usually incidents at or above a severity threshold — and log them automatically from your incident management platform so every failure has a start and end timestamp.
- **Use a rolling 90-day window.** Single months swing too much.
- **Segment by service.** A platform average of 20 days can hide a payment service at 60 days and an authentication service at 7.
- **Track the trend, not the absolute value.** Treat a significant MTBF regression as an incident in itself: something changed, and it needs investigating.
- **Report MTBF with MTTR and availability** so reviewers see both levers. See [SLA vs SLO vs SLI](https://www.itoc360.com/sla-vs-slo-vs-sli/) for how these feed your reliability targets.

## **How ITOC360 Supports MTBF Tracking**

Accurate MTBF depends on accurate incident data: every qualifying failure, with a start and resolution timestamp. ITOC360 records this automatically as part of normal incident operations, so you can calculate MTBF per service, team or platform without manual data collection — from the same data that powers your MTTR reporting. When incidents start occurring more often for a service, the pattern shows up in ITOC360's incident history before it reaches a quarterly review. [Request a demo](https://www.itoc360.com/demo/) to see it with your own incident data.

## **Sources**

- Google, *Site Reliability Engineering* — "Embracing Risk" chapter: [sre.google/sre-book/embracing-risk](http://sre.google/sre-book/embracing-risk)
- Backblaze, *Drive Stats* (hard drive failure rates): [backblaze.com/cloud-storage/resources/hard-drive-test-data](http://backblaze.com/cloud-storage/resources/hard-drive-test-data)

- [Telcordia, ](http://www.baoshunkj.cn/file/SR_332_TOC.i04.pdf)[*SR-332*](http://www.baoshunkj.cn/file/SR_332_TOC.i04.pdf)*, Reliability Prediction Procedure for Electronic Equipment*
