This PR obsoletes the JSON and Bincode transports and instead introduces a unified transport that
is generic over any tokio-serde serialization format as well as AsyncRead + AsyncWrite medium.
This comes with a slight hit for usability (having to manually specify the underlying transport
and codec), but it can be alleviated by making custom freestanding connect and listen fns.
As part of this, I made an optional tokio feature which, when enabled,
adds utility functions that spawn on the default tokio executor. This
allows for the removal of the runtime crate.
On the one hand, this makes the spawning utils slightly less generic. On
the other hand:
- The fns are just helpers and are easily rewritten by the user.
- Tokio is the clear dominant futures executor, so most people will just
use these versions.
-- Connection Limits
The problem with having ConnectionFilter default-enabled is elaborated on in https://github.com/google/tarpc/issues/217. The gist of it is not all servers want a policy based on `SocketAddr`. This PR allows customizing the behavior of ConnectionFilter, at the cost of not having it enabled by default. However, enabling it is as simple as one line:
incoming.max_channels_per_key(10, ip_addr)
The second argument is a key function that takes the user-chosen transport and returns some hashable, equatable, cloneable key. In the above example, it returns an `IpAddr`.
This also allows the `Transport` trait to have the addr fns removed, which means it has become simply an alias for `Stream + Sink`.
-- Per-Channel Request Throttling
With respect to Channel's throttling behavior, the same argument applies. There isn't a one size fits all solution to throttling requests, and the policy applied by tarpc is just one of potentially many solutions. As such, `Channel` is now a trait that offers a few combinators, one of which is throttling:
channel.max_concurrent_requests(10).respond_with(serve(Server))
This functionality is also available on the existing `Handler` trait, which applies it to all incoming channels and can be used in tandem with connection limits:
incoming
.max_channels_per_key(10, ip_addr)
.max_concurrent_requests_per_channel(10).respond_with(serve(Server))
-- Global Request Throttling
I've entirely removed the overall request limit enforced across all channels. This functionality is easily gotten back via [`StreamExt::buffer_unordered`](https://rust-lang-nursery.github.io/futures-api-docs/0.3.0-alpha.1/futures/stream/trait.StreamExt.html#method.buffer_unordered), with the difference being that the previous behavior allowed you to spawn channels onto different threads, whereas `buffer_unordered ` means the `Channels` are handled on a single thread (the per-request handlers are still spawned). Considering the existing options, I don't believe that the benefit provided by this functionality held its own.
# Changes
## Client is now a trait
And `Channel<Req, Resp>` implements `Client<Req, Resp>`. Previously, `Client<Req, Resp>` was a thin wrapper around `Channel<Req, Resp>`.
This was changed to allow for mapping the request and response types. For example, you can take a `channel: Channel<Req, Resp>` and do:
```rust
channel
.with_request(|req: Req2| -> Req { ... })
.map_response(|resp: Resp| -> Resp2 { ... })
```
...which returns a type that implements `Client<Req2, Resp2>`.
### Why would you want to map request and response types?
The main benefit of this is that it enables creating different client types backed by the same channel. For example, you could run multiple clients multiplexing requests over a single `TcpStream`. I have a demo in `tarpc/examples/service_registry.rs` showing how you might do this with a bincode transport. I am considering factoring out the service registry portion of that to an actual library, because it's doing pretty cool stuff. For this PR, though, it'll just be part of the example.
## Client::new is now client::new
This is pretty minor, but necessary because async fns can't currently exist on traits. I changed `Server::new` to match this as well.
## Macro-generated Clients are generic over the backing Client.
This is a natural consequence of the above change. However, it is transparent to the user by keeping `Channel<Req, Resp>` as the default type for the `<C: Client>` type parameter. `new_stub` returns `Client<Channel<Req, Resp>>`, and other clients can be created via the `From` trait.
## example-service/ now has two binaries, one for client and one for server.
This serves as a "realistic" example of how one might set up a service. The other examples all run the client and server in the same binary, which isn't realistic in distributed systems use cases.
## `service!` trait fns take self by value.
Services are already cloned per request, so this just passes on that flexibility to the trait implementers.
# Open Questions
In the service registry example, multiple services are running on a single port, and thus multiple clients are sending requests over a single `TcpStream`. This has implications for throttling: [`max_in_flight_requests_per_connection`](https://github.com/google/tarpc/blob/master/rpc/src/server/mod.rs#L57-L60) will set a maximum for the sum of requests for all clients sharing a single connection. I think this is reasonable behavior, but users may expect this setting to act like `max_in_flight_requests_per_client`.
Fixes#103#153#205
# New Crates
- crate rpc contains the core client/server request-response framework, as well as a transport trait.
- crate bincode-transport implements a transport that works almost exactly as tarpc works today (not to say it's wire-compatible).
- crate trace has some foundational types for tracing. This isn't really fleshed out yet, but it's useful for in-process log tracing, at least.
All crates are now at the top level. e.g. tarpc-plugins is now tarpc/plugins rather than tarpc/src/plugins. tarpc itself is now a *very* small code surface, as most functionality has been moved into the other more granular crates.
# New Features
- deadlines: all requests specify a deadline, and a server will stop processing a response when past its deadline.
- client cancellation propagation: when a client drops a request, the client sends a message to the server informing it to cancel its response. This means cancellations can propagate across multiple server hops.
- trace context stuff as mentioned above
- more server configuration for total connection limits, per-connection request limits, etc.
# Removals
- no more shutdown handle. I left it out for now because of time and not being sure what the right solution is.
- all async now, no blocking stub or server interface. This helps with maintainability, and async/await makes async code much more usable. The service trait is thusly renamed Service, and the client is renamed Client.
- no built-in transport. Tarpc is now transport agnostic (see bincode-transport for transitioning existing uses).
- going along with the previous bullet, no preferred transport means no TLS support at this time. We could make a tls transport or make bincode-transport compatible with TLS.
- a lot of examples were removed because I couldn't keep up with maintaining all of them. Hopefully the ones I kept are still illustrative.
- no more plugins!
# Open Questions
1. Should client.send() return `Future<Response>` or `Future<Future<Response>>`? The former appears more ergonomic but it doesn’t allow concurrent requests with a single client handle. The latter is less ergonomic but yields back control of the client once it’s successfully sent out the request. Should we offer fns for both?
2. Should rpc service! Fns take &mut self or &self or self? The service needs to impl Clone anyway, technically we only need to clone it once per connection, and then leave it up to the user to decide if they want to clone it per RPC. In practice, everyone doing nontrivial stuff will need to clone it per RPC, I think.
3. Do the request/response structs look ok?
4. Is supporting server shutdown/lameduck important?
Fixes#178#155#124#104#83#38
