Protohackers learnings

Protohackers ?

Protohackers is a very cool website presenting several network programming problems. So far, I've done 11 out of 12 problems and it was a fun experience with a bunch of things I learn. Each problem describe a protocol to implement, and then you give an IP address with a port, and an automated checker will test your implementation.

The neat thing about protohackers is that the challenges are fairly realistic, many of them are simplified version of real world protocols and services. For example there's a chat server, a man in the middle, a simplified tcp over udp and a simplified encryption protocol. I quite like this aspect since one can draw inspiration from the real things.

Because you only provide an IP address and the port, you can do these challenges with any programming language of your choice. If it can handle tcp and udp socket, binary and text data (with a dash of json here and there), you're good to go. I did these in rust and they're on github.

Following is a list of things I learned, some useful, some surprising, in no particular order.

Details matter, a lot

When doing network programming, it is imperative to get the details right. A classic example is read_line may or may not return a newline character at the end, if EOF has been reached. This is often important. Sending one byte wrong, especially if it's a delimiter between frame or some other control byte can be the difference a sucess and a client hanging until time out.

Another important detail that trip me up was read_exact that reads until the given buffer is full. This is fairly common to use that for length prefixed protocols. The following however, will always immediately returns with an empty buffer:

// we expect to read 32 bytes
let mut buf = Vec::with_capacity(32);
input.read_exact(&mut buf).await?;

However, with_capacity may allocates, but doesn't change the length of the buffer, so effectively, the buffer is empty and cannot hold anything. The correct thing to do is

// Correct way to initialize the buffer
let mut buf = vec![0; 32];
input.read_exact(&mut buf).await?;

Put timeouts everywhere

These are annoying, but it is quite important to think about timeout from the very beginning. What happens if a client suddenly goes away? What happens if a client is criminally slow to send data? For many problems such paranoia was vastly overkill, but I found it a good exercise. It's especially important to put timeout in tests, it provides a nicer experience when debugging a hanging server.

Channels are very good !

This was perhaps the most surprising to me. A typical server is one task owning a TcpListener and spawns tasks for clients when new connections arrive.

I started using Mutex and references between clients and server for communications and synchronisation. However, that turned out to be awkward and fiddly. Communication between central state and tasks handling client was particularly thorny, because often one needs some kind of async operation like broadcasting a message to other tasks. And doing so while holding a lock is a recipe for deadlock and/or grumpy compiler.

So I turned to another architecture, where the central state isn't shared anywhere and not held within a Mutex. When a new connection arrives, the server spawns a client, giving it a clone of a mpsc::Sender to communicate with the server.

Most of my servers ended up looking roughly like:

struct Server {
    // to give a clone to every new Client
    server_tx: tokio::sync::mpsc::Sender<Message>,

    // recv on this channel to process commands from the
    // tasks managing connections.
    server_rx: tokio::sync::mpsc::Receiver<Message>,
}

tokio::select!(
    loop {
        tokio::select! {
            // handle new connection
            x = self.listener.accept() => {
                let (stream, addr) = x?;
                tracing::info!("new client connecting from {addr}");
                // spawns a new client with a way to communicate with the server
                Client::new(stream, addr, self.server_tx.clone()).spawn_run();
            }

            // handle commands coming from the clients
            Some(cmd) = self.server_rx.recv() => {
                self.handle_command(cmd)?;
            }
        }
    }
)

And when the clients needs a response from the central server, for example, the ID of the job that was enqueued like in the job centre problem the message sent holds a mpsc::oneshot::Sender and the client can then awaits on the response:

enum Message {
    // notify the server that there's a new job to enqueue
    // the details of the job are omitted here for brevity
    NewJob{assigned_id: oneshot::Sender<u64>}
}

let (tx, rx) = oneshot::channel();
// notify the server
self.server_tx.send(Message{assigned_id: tx}).await?;
let new_id = rx.await;
// and now, we can answer the remote client with the ID for the job
// they have just submitted.

This pattern turns out to be very nice to use. It's a tad verbose because it requires a new enum and variants for every type of messages to be exchanged between server and tasks. However, this scales very well, and once the base is in place, adding messages is straightforward.

I may expand on this idea in a later post. The gist of it is that message passing + share nothing architecture works stupidly well for network code.

Compartimentalize mutable state

Even more than usual, as soon as async is involved, holding onto many mutable references becomes very tricky across await points. For example, the following is not possible:


impl Server {
    async fn run(&mut self) {
        tokio::select!(
            conn = self.accept() => { todo!() },
            cmd = self.client_command() => { todo!() },
        )
    }

    async fn accept(&mut self) { todo!() }
    async fn client_command(&mut self) { todo!() }
}

This doesn't compile because accept and client_command both take &mut self. Thankfully, you can use a &mut self method inside the branches. So to compile the previous example should be:

impl Server {
    async fn run(&mut self) {
        tokio::select!(
            conn = self.listener.accept() => { self.handle_accept(conn).await },
            cmd = self.chan.recv() => { self.handle_client_command(cmd).await },
        )
    }

    async fn handle_accept(&mut self, conn: …) { todo!() }
    async fn handle_client_command(&mut self, cmd: …) { todo!() }
}

This works because, once a branch is chosen, the many multiple borrows in select are "released".

This is where the TcpStream::split and TcpStream::into_split methods come into play, allowing one to share the read and write parts of a tcp stream without having to manually share a mutable borrow.

This forces you to think hard about what is mutable and what isn't, which is a very rust thing, and not necessarilly bad, but can be annoying sometimes.

Nom nom nom parsing

Parsing is always an important part in these challenges, if only because that's the first thing you need to do when receiving data.

I relied heavily on nom for the parsing. Sometimes it was clearly overkill, where regexp could have worked equally well. However, contrary to regexps, nom allow to build composable and extensible parsers which is very nice. Also, nom can handle incomplete input, which came handy for one challenge, and works equally well for text or binary data.

The downsides of nom are that it's a bit verbose. I like to put the parsing inside a submodule which only exports one or two function and keep the nom imports and details hidden. Also, the error handling can be complicated if you want something a bit custom. (I may write something more elaborate on this topic later.)

Think about clients as well

Even when the task is only about developping a server, it's quite handy to have a simple client ready to use as well. Especially for testing, once you're reasonably sure the low level details are fine and you want to test more complex interactions, it cuts down the verbosity of tests by a lot.

Some special mentions

line reversal

I found this problem to be the toughest. It's a (much) simpler version of quic/tcp over udp. It's quite interesting but also complex, because you need to implement yourself what the kernel usually do for you. This lead to having multiple tasks per connections, interacting together. Very satisfying to get right though.

speed daemon

This is the first problem where there are a lot of functional requirements with some subtle interactions. I found the most annoying was to have clients that could request heartbeat before or after identifying themselves as camera or dispatcher. That lead to some code duplication. I'm not sure what's the best way to deal with these kind of state machine (yet?)

Conclusion

Overall it was quite fun and educational. I liked the flexibility where you can make it as simple or as over-engineered as you'd like. The automatic checker also does test for some modicum of performance as well as many edge cases. It is also cool that all edge cases are mentionned, but sometimes somewhat hidden in the text.