The change failure rate formula
One formula, three worked examples, and the windowing rules DORA actually recommends.
The formula
The arithmetic is one line. The interesting work is in counting consistently. Use the same definition of "failed deploy" across teams or the cross-team comparison breaks.
CFR % = (failed_deploys / total_deploys) * 100 with optional exclusions: total_deploys -= fix_only_deploys failed_deploys -= externally_caused_failures
Worked example 1: low-deploy team
A 35-engineer team ships 4 deploys per week over a 30-day window. That is 17 total deploys. Two of them failed (one rollback, one hotfix). No fix-only exclusion applies because both were the original deploys, not their fixes. Both failures were internally caused.
CFR = 2 / 17 = 11.8%. The team is at the 47th percentile vs the 2025 DORA distributionDORA 2024 Report2024. With only 17 events in the window, the 95% confidence interval is wide; treat any week-over-week move as noise until the sample crosses ~30.
Worked example 2: high-frequency team with fix-only exclusion
A 180-engineer team ships 26 deploys per week, 4 of which fail, over 30 days. That is roughly 112 deploys and 17 failures. 13 of the 112 deploys are forward-fixes for the 17 failures. The exclusion drops the denominator from 112 to 99.
Adjusted CFR = 17 / 99 = 17.2%. Without the exclusion, the team would report 17 / 112 = 15.2%. The exclusion is the more honest number: forward-fixes should not pad the success column.
Worked example 3: AI-coding-era variant
A 90-engineer team enables AI pair-coding mid-quarter. Pre-AI: 18 deploys/week, 1.4 failures. Post-AI: 28 deploys/week, 2.6 failures. Pre-AI CFR = 7.8%; post-AI CFR = 9.3%. Throughput rose 56%; CFR rose 1.5 percentage points. The 2024 DORA AI report finds throughput-up, stability-modestly-down is the dominant pattern, with a sub-cohort showing stability-up after a quarter of usage DORA AI Research2024.
Window selection rule
7-day windows fit weekly stand-up discussion but only stabilise above ~30 events. 30-day is the DORA board-reporting default DORA 20242024. 90-day windows are good for quarterly trend lines and low-deploy teams; the lag cost is staleness.