(In reply to Ionuț Goldan [:igoldan], Performance Sheriffing from comment #13)
From the Perf sheriffing's perspective, weaker specs would make build time changes more obvious. We would opt for lower price.
But this isn't that much of a deal. We would rather go with your suggestions, such as picking c5d.4xlarge for Linux.
Builds need to finish before tests can run, and we run many more tests than builds, as I'm sure you're aware as a perf sheriff. I'll always opt for faster builds in CI. The number of people working on build performance is relatively small and we can make individual config arrangements for them if necessary.
- Because these are build machines and we need fewer of them relative to testers, would we be better served by using fewer, higher-spec, reserved instances to better leverage local caches, etc?
Yes, I think using fewer, higher-spec & reserved instances would serve us better.
That seems at odds with your preference for lower price above. Higher spec instances will be more expensive.
(In reply to Ionuț Goldan [:igoldan], Performance Sheriffing from comment #14)
Do you have enough data to proceed with the changes the Perf team requires?
Your preferences are noted, but we're talking about a decision that could potentially cost Mozilla thousands of dollars a month and possibly some developer downtime due to capacity.
I'll float the proposal to align on a single, more performant instance type for builds at the CI mtg on Thursday (Mar 21). I can make the changes quickly thereafter once the other stakeholders have weighed-in. I'll reflect their notes back here, if any.