This reflects what we actually do when we apply the commit, and also
means we can easily iterate the commits.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Changelog-EXPERIMENTAL: Build: all experimental features are now runtime-enabled; no more ./configure --enable-experimental-features
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Without this, a stuck test (such as before the previous commit, where a plugin crashed
when running a command) simply gets timed out by the full CI timeout.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
There are several benefits of doing this:
- prevent fuzz target bit rot
- more test coverage in CI
- greater visibility of fuzz tests, encouraging contributions to the
seed corpus and tests themselves
We can require a status to be successful in PRs, but they are
referenced by name, and so we'd have to add each matrix job as
required. That's rather cumbersome, so have this artificial gather
step at the end which signals success on the entire run.
We often have a number of changes in flight, and we amend PRs before the previous run completes. This allows us to cancel that prior run, preserving our precious runners.
They are run in parallel with the integration tests, not on the
critical path, but can be a bit lengthy since they build a log of C
code, and run in valgrind.
The tester CI job uses absolute paths to ensure it is testing the
correct binaries. That clashes with the transfer between builder and
tester job using the `install` target because that switches things
around. This commit introduces a new target that just collects
artifacts in place, and tars them. Then we can use `tar` to unpack
them on the tester jobs again.
We want to compile with one set of dependencies, and run the tests
with another. This also helps us cut down on the times we compile CLN
itself, by sharing them among stages, and simplifies the logic of each
stage to have one specific goal.
The lnprototests are often blocking PRs, due to them failing, but we
can't restart them when valgrind runs are still ongoing, since they'd
also restart. Splitting allows us to rerun them selectively and waste
less time.
Ideally we'd just fix them, but I am by no means knowledgeable enough
to fix them now.
The runner version was recently bumped causing our tests to fail
because it couldn't find the `poetry` version it just installed. We
instead install python3.10, and force its use, since otherwise we end
up compiling grpcio and fail to do so.
Changelog-None
I was wondering why the workflow never published, turns out we were triggering on
branch pushes only, not tags. So the branch would get pushed, but the tag is pushed
afterwards, and thus not triggering.
Requires us to update to latest lnproto which is now using the most up
to date python-bitcoinlib, as well as updating our python lock files
(which pin the grpcio deps, because of locking problems h/t @cdecker)
Over time, it has cost us more developer cycles than it has gained.
It has hidden intermittant bugs, and allowed cruft to accumulate:
when we eventually tried to figure out what was going wrong, the
actual change which caused it was now stale and forgotten.
This was a particular bane during the connectd rewrite, and I
worked through some issues which had occurred before, but were not
more likely.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
I have no idea why someone else owns the directory suddenly, but all git
commands fail. Workaround as suggested by the error message.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Mostly comments and docs: some places are actually paths, which
I have avoided changing. We may migrate them slowly, particularly
when they're user-visible.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
