Quic-Rpc

A streaming rpc system based on quic

Goals

Interaction patterns

Provide not just request/response RPC, but also streaming in both directions, similar to grpc.

• 1 req -> 1 res
• 1 req, update stream -> 1 res
• 1 req -> res stream
• 1 req, update stream -> res stream

It is still a RPC system in the sense that interactions get initiated by the client.

Transports

• memory transport with very low overhead. In particular, no ser/deser, currently using flume
• quic transport via the quinn crate
• transparent combination of the above

API

• The API should be similar to the quinn api. Basically "quinn with types".

Non-Goals

• Cross language interop. This is for talking from rust to rust
• Any kind of versioning. You have to do this yourself
• Making remote message passing look like local async function calls
• Being runtime agnostic. This is for tokio

Example

computation service

Why?

The purpose of quic-rpc is to serve as an optional rpc framework. One of the main goals is to be able to use it as an in process way to have well specified protocols and boundaries between subsystems, including an async boundary.

It should not have noticeable overhead compared to what you would do anyway to isolate subsystems in a complex single process app, but should have the option to also send messages over a process boundary via one of the non mem transports.

What do you usually do in rust to have isolation between subsystems, e.g. between a database and a networking layer? You have some kind of channel between the systems and define messages flowing back and forth over that channel. For almost all interactions these messages itself will again contain (oneshot or mpsc) channels for independent async communication between the subsystems.

Quic-rpc with the mem channel does exactly the same thing, except that it hides the details and allows you to specify a clean high level interaction protocol in the rust type system.

Instead of having a message that explicitly contains some data and the send side of a oneshot or mpsc channel for the response, it creates a pair of flume channels internally and sends one end of them to the server. This has some slight overhead (2 flume channels vs. 1 oneshot channel) for a RPC interaction. But for streaming interactions the overhead is negligible.

For the case where you have a process boundary, the overhead is very low for transports that already have a concept of cheap substreams (http2, quic, ...). Quic is the poster child of a network transport that has built in cheap substreams including per substream backpressure. However, I found that for raw data transfer http2/tcp has still superior performance. This is why the http2 transport exists.

Currently you would use the quinn transport for cases where you want to have connections to many different peers and can't accept a large per connection overhead, or where you want low latency for small messages.

You would use the hyper transport for cases where you have a small number of connections, so per connection overhead does not matter that much, and where you want maximum throughput at the expense of some latency.

This may change in the future as quic implementations get more optimized.