Tarpaulin: Coverage for executables

[JP-Ellis]

In the case where the crate is for a binary, cargo tarpaulin will run all the tests within the crate, but is it possible to test the executable itself (with arguments)? Or similarly, handling library with associated binaries (in src/bin/main.rs)?

Effectively, this would be in analogy to cargo run -- <args>, perhaps as cargo tarpaulin --run -- <args> or maybe a different command cargo tarpaulin-run -- <args>?

[JP-Ellis]

One comment:

It might be possible to have tests which execute the binary under crate/tests. For example

/// Get name of the binary
fn bin() -> path::PathBuf {
    let root = env::current_exe()
        .unwrap()
        .parent()
        .unwrap()
        .parent()
        .unwrap()
        .to_path_buf();
    if cfg!(target_os = "windows") {
        root.join("my-binary.exe")
    } else {
        root.join("my-binary")
    }
}

#[test]
fn test_my_binary() {
    let mut cmd = process::Command::new(bin());
    cmd.arg(...);

    match cmd.status() {
        Err(e) => {
            println!("Error: {}.", e);
            panic!();
        }
        Ok(status) => {
            println!("Exited with code {:?}.", status.code());
            assert!(!status.success());
        }
    }
}

but cargo tarpaulin doesn't follow the spawned program (unlike kcov).

[xd009642]

I now have support running tarpaulin on benchmarks and examples in the develop branch, running on arbitrary binaries is still unsupported but something I'm actively working towards

[dbrgn]

I'm working on a application server project where we have a few unit tests but a lot of integration tests. The tests are written in Python and simply run against the Rust server running on localhost.

It would be great if we could track coverage of the server binary (tarpaulin run) when running integration tests, and ideally merge the results together with the unit tests.

[pickfire]

How does one run tests for project with multiple binaries?

[xd009642]

There's a WIP PR for this if people want to test it, or comment on the UI I put in the comment https://github.com/xd009642/tarpaulin/pull/604

[xd009642]

Also as per my last comment on this repo the benchmark and example stuff changed... I realised the semantics of cargo test --examples actually looks for test annotated functions in examples, it doesn't run the example binaries directly. Same with benchmarks, so they were changed to match cargo test behaviour

[xd009642]

You can now do this via --command build which will call cargo build and take the binary you build and run it with the args. Still very much an "alpha" feature but I've tried it on a few projects and it works. Also while making it fixed another issue so pushed out a release.

So feel free to try in 0.16.0 and if there are any issues if you comment on this more recent issue https://github.com/xd009642/tarpaulin/issues/507 😄

[dbrgn]

Thanks @xd009642! Excited to test this!

My integration test suite is written in Python. It simply does calls against the main binary being run (a WebSocket server).

If I understand things correctly, I would run tarpaulin with --command build and then run the test suite against the started server, right? The problem is that the server does not stop by itself, and if I send SIGINT with Ctrl+C then both the binary and tarpaulin are stopped. I guess I'd have to modify the server so that it can be stopped from the outside (e.g. with code only built when the "coverage-testing" feature is enabled)?

When running tarpaulin both with --command test and --command build, is there a way to merge the resulting coverage information?

[xd009642]

@dbrgn Yeah I guess you'd have to modify it for a close message. There is a --forward-signals option for tarpaulin that forwards some signals to the test executable but I don't think it could help here :thinking:. And you can merge the two with a config file, I think this should work in a tarpaulin.toml

[py_tests]
command = "build"

[rust_tests]
command = "test"

[xd009642]

@dbrgn oh another thought is that you can use the more verbose logging to get the test PID and send the SIGINT to the test itself bypassing tarpaulin. Then with --forward-signals it might work, if not tarpaulin will possible just complain about an unexpected signal and not give you any results :thinking:

I need to figure out some smarter semantics for the signalling stuff... I've opened #622 to track any work on it

[dbrgn]

I now just shut down the server with std::process::exit(0) when I receive a certain message from a client (but of course only behind an optional feature flag). That seems to work! :tada: (Depending on the size of the integration test suite, the timeout needs to be raised as well.) And finally I can confirm that while unit tests in my project only cover 30%, the integration tests cover 70% :slightly_smiling_face:

[dbrgn]

@xd009642 after some more thought, maybe tarpaulin could allow running a custom shell command (hooks?) before the binary has been started, after it has been started and after it has been stopped? Maybe even with the PID being passed to the hooks. That way, the integration tests could automatically be started after the binary was started, and if the testsuite has access to the PID it could send a signal after it's done.

[xd009642]

Maybe I'd have to think about it, feel free to open an issue for the feature where we can always discuss it further :smile:

[dbrgn]

Done! https://github.com/xd009642/tarpaulin/issues/623