DORA Metrics: The 5 Metrics, Benchmarks and a 90-Day Plan

DORA metrics, five gauges measuring speed and stability above a software delivery pipeline

DORA metrics are now five numbers, not four. MTTR became failed deployment recovery time in 2023, and deployment rework rate joined in 2024. You can measure all five from Azure DevOps or GitHub data with a spreadsheet, no paid tool needed. Measure per application, never per person, and pick one bottleneck to fix first. AI coding tools are speeding teams up and making releases less stable, so watch the instability numbers closely.

DORA metrics measure software delivery with five numbers. Change lead time, deployment frequency and failed deployment recovery time cover throughput. Change fail rate and deployment rework rate cover instability. Measure them per application, not per person.

Ask most engineering leaders what the DORA metrics are and you’ll hear four of them, including MTTR. That answer is two years out of date. Ask an AI chatbot and you’ll usually get the same outdated list.

It matters because the newer metrics are the ones that catch what’s going wrong in 2026. This guide uses the definitions DORA publishes today and sizes everything for companies with 5 to 50 engineers. That’s the range our DevOps team works with most, where nobody has budget for a metrics platform and nobody has time for a dashboard that doesn’t change anything.

You’ll get the five definitions, benchmarks with their caveats, a free way to measure in Azure DevOps or GitHub, and a 90-day plan.

What are the five DORA metrics now?

DORA’s current model has five metrics in two groups. Throughput covers change lead time, deployment frequency and failed deployment recovery time. Instability covers change fail rate and deployment rework rate.

Five dials grouped three and two, representing DORA throughput and instability metrics
MetricGroupDORA’s definition, in short
Change lead timeThroughputTime from a change being committed to it running in production
Deployment frequencyThroughputNumber of deployments in a period, or the time between them
Failed deployment recovery timeThroughputTime to recover from a deployment that fails and needs immediate intervention
Change fail rateInstabilityShare of deployments that need immediate intervention afterward
Deployment rework rateInstabilityShare of deployments that are unplanned and happen because of a production incident

Definitions summarized from DORA’s metrics guide.

Two changes trip people up. In 2023, mean time to restore became failed deployment recovery time, which ties recovery to deployments instead of every outage. In 2024, DORA added deployment rework rate and regrouped the metrics into throughput and instability, as its history page explains.

Rework rate is the one we’d watch hardest. It counts the deploys you didn’t plan, the hotfixes and emergency patches. A team can look fast on deployment frequency while half its deploys are firefighting. Rework rate exposes that.

DORA also says something most teams ignore. Speed and stability aren’t tradeoffs. Its research finds them correlated for most teams, so going faster in smaller batches tends to make releases safer, not riskier.

What does good look like? Benchmarks with caveats

DORA’s 2024 report grouped teams into four performance tiers. Here they are, as summarized by Octopus Deploy’s write-up of the 2024 clusters.

TierLead timeDeploy frequencyChange fail rateRecovery time
EliteUnder 1 dayOn demand5%Under 1 hour
High1 day to 1 weekDaily to weekly20%Under 1 day
Medium1 week to 1 monthWeekly to monthly10%Under 1 day
Low1 to 6 monthsMonthly to twice a year40%1 week to 1 month

DORA 2024 performance clusters, via Octopus Deploy.

Notice the oddity. High performers fail 20% of changes and medium performers only 10%. That’s not a typo. The tiers are clusters of real teams, and the high group ships far more frequently and accepts more failures doing it. So don’t chase a single number in isolation.

Most teams aren’t elite, and that’s fine. An analysis of DORA’s 2025 data by Research-Driven Engineering Leadership found only 16.2% of respondents deploying on demand, and 43.5% with lead times over a week. If your team ships weekly with a sane failure rate, you’re already doing better than plenty of larger companies.

And DORA’s own guidance warns against comparing metrics between very different applications. Your ERP integration and your customer portal shouldn’t share a target. Benchmark each against its own last quarter.

How to measure DORA metrics without buying a tool

Pull deployment records and commit timestamps from Azure DevOps or GitHub, tag incidents and hotfixes, and calculate the five metrics monthly in a spreadsheet or Power BI. Paid tools save time later, but they aren’t required to start.

Pipeline log data flowing into a simple spreadsheet and chart, representing measuring DORA metrics without a paid tool

The free options shrank recently. Google’s open-source Four Keys project was archived in January 2024, and GitLab’s built-in DORA dashboards need its Ultimate tier. So here’s the do-it-yourself version.

Azure DevOps

  • Use Environments for production and route every production release through a pipeline that targets it. The deployment history becomes your deployment frequency.
  • Each deployment links to the commits it shipped. Commit time to deploy time is change lead time.
  • Tag rollbacks, hotfix pipelines and incident work items so you can count failed and unplanned deployments.

GitHub

  • GitHub Actions deployments to a named production environment give you the deployment log.
  • Merge or commit timestamps on the deployed SHA give you lead time.
  • Label hotfix pull requests and incident issues consistently. Labels are free. Unlabeled hotfixes make rework rate invisible.

Then export monthly into a spreadsheet or Power BI. Ten engineers don’t need real-time dashboards. They need five numbers, once a month, discussed for twenty minutes.

The hardest part isn’t tooling. It’s agreeing what a failed deployment is. Our rule is any deploy that needs a rollback, a hotfix or someone paged within 24 hours. Write your definition down, because without one the change fail rate is just an opinion.

Which applications should you measure first?

Start with one or two applications that change regularly and matter to the business, such as a customer portal or an integration your sales team depends on. Skip systems that rarely change, because their numbers won’t tell you much.

Measuring everything at once is the fastest way to stall. Mid-market companies usually run a mix of very different software, and each kind behaves differently.

Application typeMeasure it?What to expect
Customer-facing web app or portalYes, start hereFrequent changes, clear business impact, the most to gain
Integrations and APIs between systemsYesFailures show up fast in sales or finance, so change fail rate means something
ERP or CRM customizationsLaterRelease cycles tied to vendor updates. Track lead time and fail rate, not frequency
Internal tools that rarely changeNoToo few deployments for the numbers to mean anything

Where to start measuring DORA metrics.

That third row trips up teams running Dynamics 365 or Salesforce. Release timing there partly depends on the vendor’s update schedule, so deployment frequency is the wrong yardstick. Lead time and fail rate still tell you plenty.

A 90-day plan to improve

  1. Weeks 1 to 2. Baseline. Agree definitions, pull the last three months of data for one or two key applications, and calculate all five metrics. No targets yet.
  2. Weeks 3 to 6. Fix the biggest bottleneck. Usually it’s one of three. Pull requests waiting days for review, a manual regression test cycle, or a release approval that only one person can give.
  3. Weeks 7 to 9. Shrink batch size. Smaller pull requests, feature flags for unfinished work, and more frequent deploys of less change each time.
  4. Weeks 10 to 12. Make failure cheap. Automated rollback, a deployment checklist, and a short review after every failed deployment that ends in one concrete change.

A 90-day DORA improvement plan.

Four ascending steps with a bottleneck being widened, representing a 90-day DORA improvement plan

When the bottleneck is testing, and for mid-market teams it very regularly is, test automation is the lever. Our QA automation team builds the regression suites that let a team deploy without a three-day manual test cycle. We covered why that shift matters in QA vs quality engineering.

If the bottleneck is infrastructure, like environments that take a week to provision, our cloud engineering group builds them as code so a test environment takes minutes.

AI coding tools: faster, and less stable

DORA’s research shows AI coding tools raising throughput but hurting delivery stability. Its 2025 report found 90% of respondents using AI at work, with AI still negatively related to software delivery stability.

Robotic arm stacking code blocks quickly over a safety net, representing AI coding speed versus delivery stability

The trend is worth knowing in detail. DORA’s 2024 report estimated that a 25% increase in AI adoption came with a 1.5% drop in delivery throughput and a 7.2% drop in delivery stability, even as documentation and code quality improved. By 2025, the picture shifted. AI now helped throughput, but stability still suffered, and 30% of respondents reported little or no trust in AI-generated code.

DORA’s own summary of the 2025 research calls AI an amplifier. It magnifies whatever strengths and weaknesses a team already has. Good tests and small batches get better. Big batches and thin tests get worse, faster.

Practical guardrails for a small team.

  • Keep AI-assisted pull requests small. Big AI-generated diffs get skimmed, not reviewed.
  • Require tests to pass before merge, no exceptions for AI-written code
  • Watch rework rate monthly. It’s the first number to rise when AI speed outruns review quality.
  • Don’t celebrate deployment frequency alone. Speed without stability is just faster firefighting.

We went deeper on the production side of AI in our post on getting AI pilots into production.

How to keep the metrics honest

Every metric gets gamed once people are judged by it. DORA metrics are no exception. Split one change into ten deploys and frequency soars. Stop labeling hotfixes and the fail rate looks great.

  • Measure teams and applications. Never individuals, and never in performance reviews.
  • Look at all five together. Gaming one usually shows up in another.
  • Pair the numbers with a three-question survey each quarter. Are reviews slow? Is testing painful? What slowed you down most?
  • Use the metrics to start conversations, not to end them

DORA shows what is happening. It can’t tell you why. The survey and the twenty-minute monthly discussion fill that gap.

What a monthly DORA review looks like

Twenty minutes. One slide or one spreadsheet tab. The engineers who did the work in the room, plus whoever owns delivery. That’s it.

  1. Show the five numbers for each measured application, this month next to the last three.
  2. Ask what moved and why. The team usually already knows. A big release, a flaky test suite, a week of reviewers on vacation.
  3. Pick one thing to change before the next review, with one owner.
  4. Check last month’s change. Did it help? Keep it, adjust it or drop it.

A monthly DORA review agenda.

What kills these reviews is turning them into status reporting for leadership. The moment the team feels judged, the hotfix labels start disappearing. Leadership can see a quarterly summary. The monthly review belongs to the people who can change the numbers.

DORA vs SPACE vs DX Core 4

Three frameworks compete for the same attention. SPACE, from Microsoft and GitHub researchers, covers satisfaction, performance, activity, communication and efficiency. DX Core 4 rolls DORA-style delivery measures together with developer experience and business impact.

For a mid-market team, our advice is simple. Start with DORA, because the data already exists in your tools and the definitions are public. Add a short developer survey for the human side. Move to a bigger framework only when you have more than one team and a reason to compare them.

Who we are

Upen Group is a digital transformation and engineering company headquartered in Irving, Texas. Our DevOps, cloud engineering and QA automation teams help mid-market companies measure delivery with DORA metrics and then improve it, through faster pipelines, automated testing and infrastructure as code.

The short answer

Use the current five DORA metrics, measured per application. Pull them free from Azure DevOps or GitHub into a monthly spreadsheet, define a failed deployment in writing, and fix one bottleneck at a time. Watch rework rate closely as AI coding tools spread, and never turn the numbers into individual scorecards.

Want a baseline without buying a tool? Tell us whether you run Azure DevOps or GitHub and how much your teams deploy, and we’ll set up the five metrics for one application and walk your team through the first month of numbers. Short on engineers? Our engagement models range from one embedded DevOps engineer to a full team.

What engineering leaders ask us

Are there four DORA metrics or five, and what happened to MTTR?

Five. MTTR was renamed failed deployment recovery time in 2023, and deployment rework rate was added in 2024. Most articles and chatbots still list the old four.

What’s a realistic benchmark for a 10 to 30 person team?

Weekly deploys, lead time under a week and recovery within a day put you in DORA’s high or medium groups, ahead of plenty of larger companies. Compare against your own last quarter rather than Google-scale teams.

So what counts as a good change fail rate?

DORA’s 2024 elite cluster sat at 5%, but its high cluster sat at 20% while shipping far more frequently. Define a failed deployment clearly, then aim to bring your own number down quarter over quarter.

Can we measure DORA metrics in Azure DevOps or GitHub for free?

Yes. Deployment records give frequency, commit and deploy timestamps give lead time, and tagged hotfixes and incidents give fail rate, recovery time and rework rate. A monthly spreadsheet or Power BI report is enough for most mid-market teams.

Should DORA metrics go into performance reviews?

No. The moment individuals are judged by them, the numbers get gamed and stop telling you anything. Keep them at the team and application level.

Is AI coding making us faster or less stable?

Probably both. DORA’s 2025 research found AI helping throughput while still hurting stability. Watch rework rate and change fail rate, keep AI-assisted pull requests small, and make tests mandatory before merge.

How soon will we see improvement?

Usually within one or two months of fixing the first bottleneck, and lead time usually moves first. A clean 90-day cycle of baseline, one fix, smaller batches and cheaper rollback is a sensible first target.