Skip to content

Latest commit

 

History

History
141 lines (104 loc) · 6.29 KB

CHANGELOG.md

File metadata and controls

141 lines (104 loc) · 6.29 KB
Raw

0.10.0

  • Renamed Endpoint to SessionEndpoint
  • Added ServerEndpoint to mark opening servers
  • Made Server only be added to opened servers, and have this component store the opened_at: Instant
  • Removed Opened
  • All relevant aeronet_io types are now registered in the type registry

0.9.0

  • Redesigned to be Bevy native

0.8.0

  • Split the associated type Transport::Error into ::SendError and ::PollError
  • Transport::send now returns Result<.., Self::SendError>
  • {Client,Server}Event's error type is now PollError
  • flush, disconnect, close are now infallible
  • Fixed how the WebTransportServer handles local disconnections

0.7.0

  • Updated to Bevy 0.14.0
  • Added disconnection reasons to API
    • ClientTransport::disconnect, ServerTransport::{disconnect, close} now take a reason: String parameter
    • {Client, Server}Event::Disconnected now hold a reason: DisconnectReason, specifying why the client was disconnected
    • ServerEvent::Closed now holds a reason: CloseReason, specifying why the server closed
  • Fixes to WASM transport (there may still be bugs or instability)
  • Added client/server features to separate the two sides

0.6.0

  • Removed the concept of protocols
    • u8s and lanes are baked into the core -Transport traits, simplifying working with networked transports
    • Originally, the concept of bytes and lanes was abstracted out of the core API. However, this made working with transports very tedious due to the extra protocol type parameter, and didn't make much sense since only MPSC channels needed this abstraction
    • Messages are just bytes::Bytes now
  • Removed the core aeronet client/server plugins
    • Since we're now using bytes directly, it makes no sense to receive messages as events - the API user is encouraged to own their own Bytes messages, which events don't let you do
    • Users should use transport.poll() and transport.flush() manually in their own systems
  • Removed derive macros since we don't have a use for them anymore
  • Added egui stats visualizer

0.5.0

  • Complete overhaul of the crate (again)

0.4.1

  • TryIntoBytes renamed to TryAsBytes since it doesn't consume self
  • Doc updates

0.4.0

Overhaul of basically the entire crate. Treat this as a completely new crate.

What worked:

  • Using a finite state machine internally for the transport impls
    • Makes it much easier to isolate logic and determine what happens when
  • Using a single TransportProtocol type parameter instead of <C2S, S2C>
    • I originally had this in 0.1.0, but I wasn't experienced enough with the type system to implement it properly
    • Switching back to a single protocol type makes the API simpler for consumers
  • ChannelKind moved into aeronet core, as opposed to being WT-specific
    • Reliability is a general-purpose feature which should be transport-agnostic, and channels are perfect for this

What didn't work:

  • Representing WT native clients as Closed <-> (Open <-> Connected) FSM
    • Overly complicated state machine, made the logic confusing
    • Not much benefit to the consumer
  • Exposing the internal FSM types to the user
    • Complicated the API a lot, and how many users are going to be using the FSM directly?

0.3.0

No major features in this version, since I'm still working on a large rework of API for 0.4.0.

Bevy 0.12

Dependencies of this crate have been updated to use Bevy 0.12. This is the main feature of this release.

src -> aeronet/src

The core aeronet crate has been put into its own directory. Just for code organisation.

0.2.0

WebTransport channels API

The old streams API for aeronet_wt_native kind of sucked, since it required making a TransportStreams and accumulating references to streams on setup. Then you as an app user were responsible for storing these stream references in e.g. a Bevy resource. Although this -worked-, the stream API surface kind of sucked with TransportStream/ClientStream/ServerStream. Internally the code also didn't benefit from these distinctions.

Consequently, streams have been renamed to channels, and there are now two types of channels:

  • Datagram - remains the same
  • Stream - previously Bi, is now the general-purpose ordered + reliable channel type

Note that support for unidirectional streams is --dropped--, however this shouldn't be too much of a problem considering bidirectional streams cover many of the same use cases. In addition, dropping support for uni streams -greatly- simplifies the API surface, and your app can use the same channel logic for both client and server.

Transport update loop v2

Transport::recv(..) -> Result<..> has been replaced with two different functions:

  • Transport::recv(..) -> ()
  • Transport::take_events(..) -> impl Iterator<Item = ..>

Now, an app using networking must call recv first, then take_events to consume any events raised by the transport while receiving updates. This frees implementations' internal logic, as users are no longer required to keep calling recv in a loop. Instead, recv is called once, and take_events is also called once, returning an iterator with ownership over the events.

Although this is slightly less efficient as now transport implementations buffer their events in a Vec in recv and then drain them in take_events, the internal logic is -much- nicer to reason about. Still, I'm not entirely happy with this approach. Maybe the two can be merged, and a single iterator can be returned immediately. I'll experiment more.

Misc

  • Plugins will no longer remove their transport resource on failure

    The resources are designed to be reusable as much as possible (unless it's like an in-memory channel transport, in which case you would need to put a new one in). Therefore, the plugins will no longer directly remove the resources whenever a failure occurs.

    I'm also still not entirely sure on this decision, but I can come back to it later.

  • Connecting for both clients and servers is removed

    This was a bit of implementation-specific logic which I was never happy with, and with in-memory channels this event makes little sense. So I've removed it from all transports. Maybe it would still be useful to log incoming connections though? I would want to implement this in a more transport-agnostic way, not requiring incompatible transports to support it as well.