Menu

1 / 4

What our CI fixes measured: shorter queues, one claim of ours that was wrong, and a 16-hour block

DEV Community: cicd·Alo Martin·about 10 hours ago
#G2Kt9RqM
#devtocicd#devto#dev#november#jobs
Long article (3123 words) — no AI summary yet.
Reading 0:00
15s threshold

A reader on dev.to answered our CI cost post with one comment: runner comparisons are most useful when they include queue time, cache hit rate, maintenance labour and failure recovery, not just billed minutes.

We replied that labour and cache hit rate were the two we hadn't put numbers on, and that we would share the results either way. This is that post. We measured the same four things, plus merge to production and the billed-minutes equivalent, on the 30 days of the first post and on 18 to 20 November.

The short version, with the numbers in bold:

Queue time. Node jobs (lint, type-check, tests) waited 11.3 minutes at the 90th percentile before and 1.7 minutes after, at the same load. Image builds and end-to-end tests got worse, 3.0 to 16.9 minutes , because of a mistake in our first post (section 2).

Merge to production. The median did not move: 49.8 to 51.1 minutes for user-facing merges. In quiet hours after the last change, 15 merges took 9 to 18 minutes .

Cache hit rate.…

Continue reading — create a free account

Join HashtagPLUS to read full articles, follow hashtags, vote, and join the conversation.