Runner right-sizing: stop paying for capacity your jobs don't use
An 8 vCPU Linux runner costs four times what a 2 vCPU one does, $0.016 a minute against $0.004. If a job on the 8 vCPU runner averages a fifth of its CPU, you're paying four times the 2 vCPU rate on every run for capacity that sits idle. Avrea now spots those jobs and tells you which label to move them to.
It works from the CPU, memory and disk a job's recent successful runs used, compared with the capacity of the runner it ran on. A job that stays well under capacity on every resource gets a Size down suggestion with the smallest label it still fits, which can be several sizes down, 8 vCPU to 2 for example. A job whose memory or disk peaks near capacity, or whose CPU stays high on average, gets a Size up suggestion with the next size that has more of what ran out. A job needs three successful runs on a size before it gets either, and a job near a threshold isn't bounced back and forth.
You'll see the suggestion on every finished job: job and run pages flag jobs that should size up or down and name the label to switch to. For a whole workflow, choose Right-size in your repository's Workflows tab and you get every job's suggestion over the last day, 7 days or 30 days. The same suggestions are in the API, per workflow or for up to 50 workflows of a repository in one call.
The Right-size view also has a ready-made prompt for your coding assistant. It lists each job's current and suggested label and says to change runs-on and nothing else. Paste it in and review the change.
One thing to keep an eye on. The CPU check uses the average over the run, so a job that idles through a long download and then compiles on every core for a minute can come out a little slower on the smaller runner. If its duration grows after the switch, the runs on the new size get their own suggestion.
The thresholds, the API and the response format are in the runner right-sizing docs.
