Jens-G opened a new pull request, #3799:
URL: https://github.com/apache/thrift/pull/3799
Nothing in CI builds or runs `lib/erl`. `build.yml` has jobs for php, go,
java-kotlin, netstd, haxe, rust, python, nodejs, cpp and ruby, and both
`build.yml` and `sca.yml` pass `--without-erlang`; `docker.yml` only builds and
validates the images. `lib/erl` ships enabled by default in the release
tarball, so this is code reaching users with no regression coverage at all —
THRIFT-6163, THRIFT-6164 and THRIFT-6184 each had to be verified by running
`rebar3 eunit` by hand.
Most of what a job needs already exists: `lib/erl/Makefile.am` has `check:
.generated` followed by `$(REBAR) eunit`, and `configure.ac` gates the binding
behind `--with-erlang` and locates `erl`/`erlc`.
### The job
Sets up OTP and rebar3, bootstraps, configures with `--with-erlang` and the
other bindings off, runs `make -C lib/erl check`. The compiler artifact is
needed because `.generated` runs the compiler over the test IDL. `setup-beam`
has to come **before** configure, which looks for `erl` and `erlc` on `PATH`.
### Why a matrix, and not just the 25.3.2.9 the images pin
THRIFT-6184 (#3797) is the argument. The default Erlang TLS client stopped
being able to connect when OTP 26 changed how `ssl:connect/3` reads its option
list — a job pinned to 25.3 alone would have stayed green straight through it.
So 25.3, 26.2, 27.3 and 28.5.
### OTP and rebar3 are one axis, not two
No single rebar3 covers the range:
```
OTP 25.3.2.9 rebar3 3.18.0 ok rebar3 3.27.0 Error loading module
rebar_log
OTP 26.2.5.21 rebar3 3.18.0 ok rebar3 3.27.0 ok
OTP 27.3.4.17 rebar3 3.18.0 undefined function rebar3:main/1
OTP 28.5.0.6 rebar3 3.18.0 undefined function rebar3:main/1
```
3.18.0 is what the docker images install; it does not load on 27 or 28.
3.27.0 does not load on 25. They overlap only on 26, so each matrix entry
carries its own pair.
### OTP 29 is left out
`lib/erl` itself compiles there. The eunit-only `meck 0.9.2` pinned in
`rebar.config` does not — deprecated `catch` is an error as of 29, and
`meck_matcher.erl` still uses it:
```
===> Compiling _build/test/lib/meck/src/meck_matcher.erl failed
76 │ (catch erlang:apply(hamcrest, assert_that, [Value, HamcrestMatcher]))
== true;
│ ╰── 'catch ...' is deprecated; please use 'try ... catch ... end'
instead.
```
Raising that pin is a change to `lib/erl`'s test dependencies rather than to
CI, so it gets its own ticket.
### A guard on configure
configure disables a binding it cannot locate rather than failing, so the
step checks that Erlang was actually found. Without it, a job whose
`setup-beam` step had quietly stopped working would keep reporting success
while running no tests — worse than having no job.
### LANGUAGES.md
It recorded the tested OTP range as 18.3 to 22.0, which matched neither the
docker images nor anything CI ran. #3797 flagged that as stale and said it
should be raised "on the back of real CI coverage" — this is that. It now says
25.3 to 28.5, what the job pins.
### Verification
On a clean `ubuntu:24.04`, per matrix entry, using the same prebuilt OTP
tarballs `setup-beam` pulls from `builds.hex.pm`:
```
OTP 25.3.2.9 349 tests, 0 failures
OTP 26.2.5.21 349 tests, 0 failures
OTP 27.3.4.17 349 tests, 0 failures
OTP 28.5.0.6 349 tests, 0 failures
```
With `erl` off `PATH` the job fails at configure instead of passing.
### Not in this PR
Adding `erl` to the `cross-test` matrix. `test/erl` has a `precross` target
but no `check`, so that is a follow-up on its own merits.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
