Some of these futures seem to be referencing the salsa db and holding it across await points -- that is, they are !Send -- so presumably this is slightly more complex than a tokio::spawn call ?

Hmm. I thought that each test created its own database? Not sure.

I believe that they do, but IIRC some of the Kernel interpretation is async and uses the non-Send db. I"ll check again.

Ah, I see. We should use a "non-workstealing" setup of tokio, or switch to glommio perhaps. We could also make things Send but I"m not sure that"s worth the bother.

That said, I imagine everything is Send...

I spent some time looking at this and got stuck once I started running tests in parallel and in arbitrary order. The shared LSP appears to deadlock within salsa while interning InputFile. It"s pretty unexpected since I"m serializing requests to the LSP and it shouldn"t be experiencing any parallelism.

I"m stumped for now. Here"s my branch: https://github.com/brson/dada/tree/parallel-tests

We just landed a PR fixing a potential deadlock with interning in salsa, so that may be related.

I rebased my parallel testing branch and it still deadlocks. Haven"t investigated further.