Duration percentiles

P50 and P95 pipeline duration are percentiles of how long pipeline runs take: P50 (the median) is the duration half of all runs finish under, and P95 is the duration 95% of runs finish under — only one run in twenty is slower.

How it is measured

Take a window of completed run durations, sort them, and read the value at the 50th and 95th positions of the distribution. Duration here means execution time, start to finish — the wait before the run starts is queue time, a separate metric with separate fixes. Providers draw that line in slightly different places: a GitHub Actions workflow run carries no completion timestamp — only created_at, run_started_at and updated_at (workflow runs API, as of 2026-07-28) — so a run-level duration there is an upper bound that still includes the queue wait, while GitLab reports a pipeline duration with queue time excluded (jobs API, as of 2026-07-28). Percentiles are therefore compared within a provider, not across two.

Percentiles exist because the average lies. A mean is dragged by outliers and hides bimodal distributions: a pipeline that runs in 4 minutes on a warm cache and 20 minutes on a cold one averages out to a duration nobody ever experiences. P50 describes the typical run; P95 describes the tail. The gap between them is itself diagnostic — a narrow gap means a consistent pipeline, while a wide one means high variance: cache misses, runner contention, retried jobs, or flaky behavior somewhere in the suite.

Why it matters

Developer experience is set by the tail, not the median. Each run has a 5% chance of landing beyond P95, so a developer who triggers ten runs a day has roughly a 40% chance of hitting the tail at least once that day — the slow runs are not rare events in anyone's week. And people plan around the worst case they remember: if P95 is 25 minutes, developers context-switch to something else after pushing, even when the median run would have finished in 8. The cost of that switching is a large share of what failed and slow builds actually cost.

Tail runs also point at causes averages smear away. A P95 dominated by one job's cold starts, or by retries piling extra attempts onto already-slow runs (the retry tax again), is actionable in a way "average duration went up 9%" never is — which makes percentiles the right starting point for diagnosing a slow pipeline.

TrimCI reports p50 and p95 run duration per provider in its PDF reports — overall and for successful runs only, plus a daily median trend for single-provider scopes — alongside a slowest-pipelines ranking per repository and pipeline, queue time and failure cost — so the optimization list starts with the runs that actually hurt, not the ones the average says are fine.