How to implement backpressure on imap-flow?

[superboum]

TCP/IP is great as if you don't read the socket, once the TCP OS buffer will be full, the information will be back-propagated through the network to the client, that will slow-down its sending. Unfortunately, we break the "backpressure-chain" as, to send data on imap-flow, you also have to read them, and thus either store it or drop it.

My current workaround is closing the socket if the client sends 10 commands in a row. But it's not ideal, and still can lead to DoS: if we consider 25MB literals, with 10 commands, it's 250MB of memory that can be allocated. Doing that in 10 or 100 of connections, and you can easily trigger an OOM on the server.

[jakoschiko]

We investing heavily in separating the IMAP state machine from the I/O bits. The I/O bits are now located inside of imap_next::stream::Stream and are optional. We tried to keep Stream simple because complex I/O features should not be part of imap-next, at least for now.

If something like backpressure is necessary, I would suggest to omit imap_next::stream::Stream (you can remove it with default-features = false and implement the I/O bits yourself. The state machine should give you enough control to implement anything you want and that fits your use case.