When removing `concurrency`, I overlooked this case where we want to
serialize deployments.
Queuing deployments also doesn't give us much. I switched from using an
action for this to using built-in support.
This drops us down to just a handlful of jobs, allowing us full
parallelism (github caps max parallel jobs). This is dependent on us
using bors to run the "ci" before merging into master.
There is a balance in what to run. We should consider what is most
likely to break for the widest variety of PRs. Contributors that expect
an uncovered case to fail can always specify `@bors try`
Motivation
- Mac is similar enough to Linux, we only need to run one of them and
Linux has more parallel runners on Github.
- Since we deal with `OsStr`, test Windows because its different than
the others.
- People are most likely to make changes on `stable` and break support
for MSRV, so we should verify that
- Still test on `stable` to not block feedback if we run into problems
with dependencies and our MSRV run.
- On the other hand, beta and nightly are less likely to break on an
individual PR
- Remove benchmarks because most changes are not performance sensitive
and we aren't looking at the results enough to justify a 30 minute run.
Fixes#2801
Using `head_ref`, we are making it so PRs are all in the same group.
When a new PR comes in (not just an update), it then cancels all other
PRs. Switching to `ref` makes it so each PR is in its own concurrency
group.
push(master) is not run because they would be run on staging anyway.
pull_request(staging, trying) are not run because pull requests shouldn't be made to them
