Skip to content
This repository has been archived by the owner on Sep 9, 2023. It is now read-only.
/ nine43 Public archive

On the path to understanding concurrency...

License

Notifications You must be signed in to change notification settings

math2001/nine43

Repository files navigation

Nine43 Build Status

The successor of Nine42

This project is supposed to teach me concurrency. It's the start of game server with a client, all in Python.

The server

The server is organized in sections that are completely independent. The only way they talk with one an other is through channels. This makes the server much easier to test because all those sections are pure functions (in the sense that they have no side effect, and aren't be affected by any).

Firstly, we have the network section, the interface between the network and the application. It accepts raw trio.SocketStreams, wraps them in net.JSONStream and sends them, through a channel called connch to the initiator.

The initiator has one task: find out the username of the connection (and make sure that there aren't any duplicates in the application). Once it has done that, it can wrap the connection in a Player (which is just a connection and a username). It then sends it off to a channel called playerch.

And the lobby is listening to that channel. The lobby's role is very simplistic: to stack up players in groups of N, and sends the Group on an other channel: groupch.

Notice of up to now, all those section are exclusive on the application: there is only one server running, oneinitiator and one lobby.

But now that we have a group of players, we can ask them what they want to do, and then, depending on, it will start a chain of events: a sub. A sub is just a select section, a game section, and an end section.

The server needs to be able to run multiple subs at the same time, but none of those subs need to know about each other. So we have a sub manager which spawns subs as it receives groups.

In this case, there are 3 subs running:

visualization of process explained above

Made using stackedit.io and mermaid

There are a few key parts I've omitted on this graph because it makes look unnecessarily complex.

  1. Every section is connected to the initiator through a channel called quitch. As soon as a connection is closed, we assumed that the layer below tried it's best to recover the connection, and throw it away. But the username needs to be released. This is the role of the quitch: every section can send a player on it, and the initiator will release the username. initiator is the sole consumer of this channel, every other section receives a clone of the sending end of the channel.
  2. As player leave a sub (leave the fin section), they get sent back to the lobby through a clone of the playerch. This is managed by the sub manager (the server doesn't anything about what's going on inside a sub).

Every section is in fact just a proxy, which reads from a channel, and writes on another.

This means the sections are easy to test independently, and we'll know that they'll behave likewise in the app, because those channels are the only way they have to communicate.

Only the sender should close the connection.

Another important part of the server is cascading: as soon a section detects that its input channel has been closed, it cleans itself up, and closes its output channel, creating a cascading effect, closing everything.

You might notice that since subs don't read from a channel (only the sub manager does, it then spawns one sub per group), it won't know when to stop.

And this is an interesting behavior that could be desired: don't accept anymore connections (close everything up to the sub manager), but finish the game you're running.

If we really want to close everything forcefully, it'd be quite easy to pass a trio.Event to every sub that will be triggered when it needs to be forcefully closed.

Note that this organization only works because when something is written on a channel, it is given up by the sender. For example, this wouldn't be allowed:

async def section1(ch: SendCh[Obj]):
    obj = make_new_obj(arg='value')
    await ch.send(obj)
    obj.alter() # no!! You don't own obj anymore!

It prevents us from thinking of every section as independent, and we have to consider the whole application as a whole.

Client

TODO

About

On the path to understanding concurrency...

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages