Integration tests for redaction failing

[gnunicorn]

since 1a57170 the redaction integration tests fail - name_static almost every time on CI, the other has been observed outside of CI to fail, too:

test tests::redaction::test_redacting_name_static ... FAILED
test tests::redaction::test_redacting_name ... ok

failures:

---- tests::redaction::test_redacting_name stdout ----
thread 'tests::redaction::test_redacting_name' panicked at 'dispatch dropped without returning error', /home/ben/.cargo/registry/src/jackfan.us.kg-1ecc6299db9ec823/hyper-0.14.20/src/client/conn.rs:329:35
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace


failures:
    tests::redaction::test_redacting_name

test result: FAILED. 2 passed; 1 failed; 0 ignored; 0 measured; 0 filtered out; finished in 5.61s

[jplatte]

This seems like an issue other people are facing too with hyper. Are we sharing the hyper client between multiple async runtimes perhaps? That's one way to trigger the error according to issues. These two seem to contain the most info:

• Consider removing "dispatch dropped without returning error" hyperium/hyper#2649
• panicked at dispatch dropped without returning error hyperium/hyper#2892

[gnunicorn]

multiple async runtimes perhaps?

does #[tokio::test] create a fresh runtime for every test? It would make sense, that it does. And then, yes. Our integration tests reuse existing clients stored in a dashmap static USERS: Lazy<Mutex<HashMap<String, (Client, TempDir)>>>.

[jplatte]

Yes, that is what #[tokio::test] does. Do we really need to reuse clients across tests, are they that expensive to create?

[gnunicorn]

I think the main problem here is e2ee, so they are created for the entire test suite with the proper sled store in tmp (which he hold the tmpdir for so that it doesn't get scraped on drop), the other that we register with the server if they don't exist before (locally) - which saves a few roundtrips, too. I suppose we could change that. Maybe it does make sense to provide some TestClient that we can just hammer in some info (like the homeserver) to prevent the unnecessary roundtrips?

[jplatte]

The other option would be to not use #[tokio::test], and share the runtime across tests as well.