Depending on the workflows of individual developers, it can either be
convenient or annoying that our GitHub Actions CI jobs are run on every
branch. As an example of annoying: if you carry many half-finished
work-in-progress branches and rebase them frequently against master,
you'd get tons of failure reports that aren't interesting (not to
mention the wasted CPU).
This commit adds a new job which checks a special branch within the
repository for CI config, and then runs a shell script it finds there to
decide whether to skip the rest of the tests. The default will continue
to run tests for all refs if that branch or script is missing.
There have been a few alternatives discussed:
One option is to carry information in the commit itself about whether it
should be tested, either in the tree itself (changing the workflow YAML
file) or in the commit message (a "[skip ci]" flag or similar). But
these are frustrating and error-prone to use:
- you have to manually apply them to each branch that you want to mark
- it's easy for them to leak into other workflows, like emailing patches
We could likewise try to get some information from the branch name. But
that leads to debates about whether the default should be "off" or "on",
and overriding still ends up somewhat awkward. If we default to "on",
you have to remember to name your branches appropriately to skip CI. And
if "off", you end up having to contort your branch names or duplicate
your pushes with an extra refspec.
By comparison, this commit's solution lets you specify your config once
and forget about it, and all of the data is off in its own ref, where it
can be changed by individual forks without touching the main tree.
There were a few design decisions that came out of on-list discussion.
I'll summarize here:
- we could use GitHub's API to retrieve the config ref, rather than a
real checkout (and then just operate on it via some javascript). We
still have to spin up a VM and contact GitHub over the network from
it either way, so it ends up not being much faster. I opted to go
with shell to keep things similar to our other tools (and really
could implement allow-refs in any language you want). This also makes
it easy to test your script locally, and to modify it within the
context of a normal git.git tree.
- we could keep the well-known refname out of refs/heads/ to avoid
cluttering the branch namespace. But that makes it awkward to
manipulate. By contrast, you can just "git checkout ci-config" to
make changes.
- we could assume the ci-config ref has nothing in it except config
(i.e., a branch unrelated to the rest of git.git). But dealing with
orphan branches is awkward. Instead, we'll do our best to efficiently
check out only the ci/config directory using a shallow partial clone,
which allows your ci-config branch to be just a normal branch, with
your config changes on top.
- we could provide a simpler interface, like a static list of ref
patterns. But we can't get out of spinning up a whole VM anyway, so
we might as well use that feature to make the config as flexible as
possible. If we add more config, we should be able to reuse our
partial-clone to set more outputs.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
e76eec3554