"Levente Uzonyi"<leves at elte.hu> wrote:
> Okay, I really tracked down the cause of the problem. The tests perform 
> simple producer-consumer scenarios, but the consumers won't wait at all. 
> If there's nothing to consume, the test fails. The randomness come from 
> the scheduler. The server process is started first, the client is the 
> second. If the server process can produce enough input to the client 
> process the test will pass.
> If you decrease the priority of the client process by one in 
> #timeout:server:client:, the randomness will be gone and the tests will 
> reliably pass, because the client won't be able to starve the server. To 
> avoid false timeouts I had to increase the timeout value to 2000 
> milliseconds using SqueakVM.

An alternative solution I was thinking about would be to dispense with the processes in default case and simply run the server and client blocks sequentially. The cases where we really need the processes are channels with limited capacity (e.g. Socket or Pipe) where the writer needs to be able to block (when the OS buffers fill up) and let the reader consume some of the output before it continues writing. In these cases there aren't any shared resources so process contention should not be an issue. The other advantage would be that cases with the default behavior would be easier to debug and would not be subject to timeouts. As it is I still get timeouts sporadically even when it's set to 1 second.
 
> I also found an issue: the process in XTTransformWriteStream doesn't 
> terminate. If you run the tests, you'll get 12 lingering processes.

I don't see that happening in Pharo 1.1. Normally the process should end normally when the stream is closed (the close synchronizes with the process via the closeReady semaphore). Maybe there's a missing close somewhere in the test suite ?