What you are seeing
Each blue block is a commit waiting for a free CI runner. Runners (green bars) build and test one commit at a time; a flaky failure (red) sends the commit back to the queue. Add runners, shorten the pipeline or fix flaky tests and lead time collapses; push utilisation near 100 percent and the queue explodes non-linearly.
The science
Model. Commits arrive as a Poisson process with rate lambda. Each pipeline run takes an exponential time with mean 1/mu on one of c runners (FIFO). A run fails spuriously with probability f and the commit re-enters the queue, so the number of runs per commit is geometric with mean 1/(1-f) and the effective arrival rate is lambdaeff = lambda/(1-f).
Theory. M/M/c: a = lambdaeff/mu, rho = a/c, Erlang C probability of waiting C(c, a) = [ac/(c!(1-rho))] / [sumk<c ak/k! + ac/(c!(1-rho))], Wq = C/(c mu - lambdaeff). Lead time = (1/(1-f))(Wq + 1/mu) by Wald's identity. Little's law L = lambda W ties queue content to waiting.
Simplifications: exponential build times (real ones are less variable), constant arrival rate (no working-hours pattern), identical runners, no parallel stages, retry re-queues at the back; retried arrivals are not Poisson, so the lead-time formula is a close approximation (checked within 8 %).