Open neilconway opened 13 years ago
lol, did we seriously replace TCP with UDP without leaving an option for TCP?
Yup.
that sucks, can i check it back in as an option?
Maybe cleanest would be to add support for specifying TCP transport in the channel syntax.
Why do you need TCP, btw?
for flow control and reliable delivery mainly; my RTrace thing sends a lot of trace info at the end of a computation (timings of all sent messages and elapsed timestamps) back to a master and i'm worried that the deluge of data will overflow buffers and cause data to get lost. i haven't actually tested this yet because my RTrace thing isn't yet working.
that brings me to my second point, which is that in debugging my code, i'd like to be able to rule out the possibility of "UDP randomly dropped this message" when considering why my code isn't working like i expect.
We currently provide unreliable channels, implemented via UDP. This is perfectly reasonable, but most applications will want something a bit more sophisticated:
We could implement this stuff in Bloom on top of reliable channels, but another approach would be to provide support for sending messages via TCP. This raises some interesting questions: