ASGI (Asynchronous Server Gateway Interface) Draft Spec

Note

This is still in-progress, but is now mostly complete.

Abstract

This document proposes a standard interface between network protocol servers (particularly web servers) and Python applications, intended to allow handling of multiple common protocol styles (including HTTP, HTTP2, and WebSocket).

This base specification is intended to fix in place the set of APIs by which these servers interact and the guarantees and style of message delivery; each supported protocol (such as HTTP) has a sub-specification that outlines how to encode and decode that protocol into messages.

The set of sub-specifications is available in the Message Formats section.

Rationale

The WSGI specification has worked well since it was introduced, and allowed for great flexibility in Python framework and web server choice. However, its design is irrevocably tied to the HTTP-style request/response cycle, and more and more protocols are becoming a standard part of web programming that do not follow this pattern (most notably, WebSocket).

ASGI attempts to preserve a simple application interface, but provide an abstraction that allows for data to be sent and received at any time, and from different application threads or processes.

It also take the principle of turning protocols into Python-compatible, asynchronous-friendly sets of messages and generalises it into two parts; a standardised interface for communication and to build servers around (this document), and a set of standard message formats for each protocol.

Its primary goal is to provide a way to write HTTP/2 and WebSocket code, alongside normal HTTP handling code, however, and part of this design is ensuring there is an easy path to use both existing WSGI servers and applications, as a large majority of Python web usage relies on WSGI and providing an easy path forwards is critical to adoption. Details on that interoperability are covered in HTTP & WebSocket ASGI Message Format (Draft Spec).

The end result of this process has been a specification for generalised inter-process communication between Python processes, with a certain set of guarantees and delivery styles that make it suited to low-latency protocol processing and response. It is not intended to replace things like traditional task queues, but it is intended that it could be used for things like distributed systems communication, or as the backbone of a service-oriented architecure for inter-service communication.

Overview

ASGI consists of three different components - protocol servers, a channel layer, and application code. Channel layers are the core part of the implementation, and provide an interface to both protocol servers and applications.

A channel layer provides a protocol server or an application server with a send callable, which takes a channel name and message dict, and a receive callable, which takes a list of channel names and returns the next message available on any named channel.

Thus, rather than under WSGI, where you point the protocol server to the application, under ASGI you point both the protocol server and the application to a channel layer instance. It is intended that applications and protocol servers always run in separate processes or threads, and always communicate via the channel layer.

ASGI tries to be as compatible as possible by default, and so the only implementation of receive that must be provided is a fully-synchronous, nonblocking one. Implementations can then choose to implement a blocking mode in this method, and if they wish to go further, versions compatible with the asyncio or Twisted frameworks (or other frameworks that may become popular, thanks to the extension declaration mechanism).

The distinction between protocol servers and applications in this document is mostly to distinguish their roles and to make illustrating concepts easier. There is no code-level distinction between the two, and it’s entirely possible to have a process that does both, or middleware-like code that transforms messages between two different channel layers or channel names. It is expected, however, that most deployments will fall into this pattern.

There is even room for a WSGI-like application abstraction on the application server side, with a callable which takes (channel, message, send_func), but this would be slightly too restrictive for many use cases and does not cover how to specify channel names to listen on. It is expected that frameworks will cover this use case.

Channels and Messages

All communication in an ASGI stack uses messages sent over channels. All messages must be a dict at the top level of the object, and contain only the following types to ensure serializability:

  • Byte strings
  • Unicode strings
  • Integers (within the signed 64 bit range)
  • Floating point numbers (within the IEEE 754 double precision range)
  • Lists (tuples should be treated as lists)
  • Dicts (keys must be unicode strings)
  • Booleans
  • None

Channels are identified by a unicode string name consisting only of ASCII letters, ASCII numerical digits, periods (.), dashes (-) and underscores (_), plus an optional type character (see below).

Channels are a first-in, first out queue with at-most-once delivery semantics. They can have multiple writers and multiple readers; only a single reader should get each written message. Implementations must never deliver a message more than once or to more than one reader, and must drop messages if this is necessary to achieve this restriction.

In order to aid with scaling and network architecture, a distinction is made between channels that have multiple readers (such as the http.request channel that web applications would listen on from every application worker process), single-reader channels that are read from a single unknown location (such as http.request.body?ABCDEF), and process-specific channels (such as a http.response.A1B2C3!D4E5F6 channel tied to a client socket).

Normal channel names contain no type characters, and can be routed however the backend wishes; in particular, they do not have to appear globally consistent, and backends may shard their contents out to different servers so that a querying client only sees some portion of the messages. Calling receive on these channels does not guarantee that you will get the messages in order or that you will get anything if the channel is non-empty.

Single-reader channel names contain a question mark (?) character in order to indicate to the channel layer that it must make these channels appear globally consistent. The ? is always preceded by the main channel name (e.g. http.response.body) and followed by a random portion. Channel layers may use the random portion to help pin the channel to a server, but reads from this channel by a single process must always be in-order and return messages if the channel is non-empty. These names must be generated by the new_channel call.

Process-specific channel names contain an exclamation mark (!) that separates a remote and local part. These channels are received differently; only the name up to and including the ! character is passed to the receive() call, and it will receive any message on any channel with that prefix. This allows a process, such as a HTTP terminator, to listen on a single process-specific channel, and then distribute incoming requests to the appropriate client sockets using the local part (the part after the !). The local parts must be generated and managed by the process that consumes them. These channels, like single-reader channels, are guaranteed to give any extant messages in order if received from a single process.

Messages should expire after a set time sitting unread in a channel; the recommendation is one minute, though the best value depends on the channel layer and the way it is deployed.

The maximum message size is 1MB if the message were encoded as JSON; if more data than this needs to be transmitted it must be chunked or placed onto its own single-reader or process-specific channel (see how HTTP request bodies are done, for example). All channel layers must support messages up to this size, but protocol specifications are encouraged to keep well below it.

Handling Protocols

ASGI messages represent two main things - internal application events (for example, a channel might be used to queue thumbnails of previously uploaded videos), and protocol events to/from connected clients.

As such, there are sub-specifications that outline encodings to and from ASGI messages for common protocols like HTTP and WebSocket; in particular, the HTTP one covers the WSGI/ASGI interoperability. It is recommended that if a protocol becomes commonplace, it should gain standardized formats in a sub-specification of its own.

The message formats are a key part of the specification; without them, the protocol server and web application might be able to talk to each other, but may not understand some of what the other is saying. It’s equivalent to the standard keys in the environ dict for WSGI.

The design pattern is that most protocols will share a few channels for incoming data (for example, http.request, websocket.connect and websocket.receive), but will have individual channels for sending to each client (such as http.response!kj2daj23). This allows incoming data to be dispatched into a cluster of application servers that can all handle it, while responses are routed to the individual protocol server that has the other end of the client’s socket.

Some protocols, however, do not have the concept of a unique socket connection; for example, an SMS gateway protocol server might just have sms.receive and sms.send, and the protocol server cluster would take messages from sms.send and route them into the normal phone network based on attributes in the message (in this case, a telephone number).

Extensions

Extensions are functionality that is not required for basic application code and nearly all protocol server code, and so has been made optional in order to enable lightweight channel layers for applications that don’t need the full feature set defined here.

The extensions defined here are:

  • groups: Allows grouping of channels to allow broadcast; see below for more.
  • flush: Allows easier testing and development with channel layers.
  • statistics: Allows channel layers to provide global and per-channel statistics.
  • twisted: Async compatibility with the Twisted framework.
  • asyncio: Async compatibility with Python 3’s asyncio.

There is potential to add further extensions; these may be defined by a separate specification, or a new version of this specification.

If application code requires an extension, it should check for it as soon as possible, and hard error if it is not provided. Frameworks should encourage optional use of extensions, while attempting to move any extension-not-found errors to process startup rather than message handling.

Groups

While the basic channel model is sufficient to handle basic application needs, many more advanced uses of asynchronous messaging require notifying many users at once when an event occurs - imagine a live blog, for example, where every viewer should get a long poll response or WebSocket packet when a new entry is posted.

This concept could be kept external to the ASGI spec, and would be, if it were not for the significant performance gains a channel layer implementation could make on the send-group operation by having it included - the alternative being a send_many callable that might have to take tens of thousands of destination channel names in a single call. However, the group feature is still optional; its presence is indicated by the supports_groups attribute on the channel layer object.

Thus, there is a simple Group concept in ASGI, which acts as the broadcast/multicast mechanism across channels. Channels are added to a group, and then messages sent to that group are sent to all members of the group. Channels can be removed from a group manually (e.g. based on a disconnect event), and the channel layer will garbage collect “old” channels in groups on a periodic basis.

How this garbage collection happens is not specified here, as it depends on the internal implementation of the channel layer. The recommended approach, however, is when a message on a process-specific channel expires, the channel layer should remove that channel from all groups it’s currently a member of; this is deemed an acceptable indication that the channel’s listener is gone.

Implementation of the group functionality is optional. If it is not provided and an application or protocol server requires it, they should hard error and exit with an appropriate error message. It is expected that protocol servers will not need to use groups.

Linearization

The design of ASGI is meant to enable a shared-nothing architecture, where messages can be handled by any one of a set of threads, processes or machines running application code.

This, of course, means that several different copies of the application could be handling messages simultaneously, and those messages could even be from the same client; in the worst case, two packets from a client could even be processed out-of-order if one server is slower than another.

This is an existing issue with things like WSGI as well - a user could open two different tabs to the same site at once and launch simultaneous requests to different servers - but the nature of the new protocols specified here mean that collisions are more likely to occur.

Solving this issue is left to frameworks and application code; there are already solutions such as database transactions that help solve this, and the vast majority of application code will not need to deal with this problem. If ordering of incoming packets matters for a protocol, they should be annotated with a packet number (as WebSocket is in its specification).

Single-reader and process-specific channels, such as those used for response channels back to clients, are not subject to this problem; a single reader on these must always receive messages in channel order.

Capacity

To provide backpressure, each channel in a channel layer may have a capacity, defined however the layer wishes (it is recommended that it is configurable by the user using keyword arguments to the channel layer constructor, and furthermore configurable per channel name or name prefix).

When a channel is at or over capacity, trying to send() to that channel may raise ChannelFull, which indicates to the sender the channel is over capacity. How the sender wishes to deal with this will depend on context; for example, a web application trying to send a response body will likely wait until it empties out again, while a HTTP interface server trying to send in a request would drop the request and return a 503 error.

Process-local channels must apply their capacity on the non-local part (that is, up to and including the ! character), and so capacity is shared among all of the “virtual” channels inside it.

Sending to a group never raises ChannelFull; instead, it must silently drop the message if it is over capacity, as per ASGI’s at-most-once delivery policy.

Specification Details

A channel layer must provide an object with these attributes (all function arguments are positional):

  • send(channel, message), a callable that takes two arguments: the channel to send on, as a unicode string, and the message to send, as a serializable dict.
  • receive(channels, block=False), a callable that takes a list of channel names as unicode strings, and returns with either (None, None) or (channel, message) if a message is available. If block is True, then it will not return a message arrives (or optionally, a built-in timeout, but it is valid to block forever if there are no messages); if block is false, it will always return immediately. It is perfectly valid to ignore block and always return immediately, or after a delay; block means that the call can take as long as it likes before returning a message or nothing, not that it must block until it gets one.
  • new_channel(pattern), a callable that takes a unicode string pattern, and returns a new valid channel name that does not already exist, by adding a unicode string after the ! or ? character in pattern, and checking for existence of that name in the channel layer. The pattern must end with ! or ? or this function must error. If the character is !, making it a process-specific channel, new_channel must be called on the same channel layer that intends to read the channel with receive; any other channel layer instance may not receive messages on this channel due to client-routing portions of the appended string.
  • MessageTooLarge, the exception raised when a send operation fails because the encoded message is over the layer’s size limit.
  • ChannelFull, the exception raised when a send operation fails because the destination channel is over capacity.
  • extensions, a list of unicode string names indicating which extensions this layer provides, or an empty list if it supports none. The possible extensions can be seen in Extensions.

A channel layer implementing the groups extension must also provide:

  • group_add(group, channel), a callable that takes a channel and adds it to the group given by group. Both are unicode strings. If the channel is already in the group, the function should return normally.
  • group_discard(group, channel), a callable that removes the channel from the group if it is in it, and does nothing otherwise.
  • group_channels(group), a callable that returns an iterable which yields all of the group’s member channel names. The return value should be serializable with regards to local adds and discards, but best-effort with regards to adds and discards on other nodes.
  • send_group(group, message), a callable that takes two positional arguments; the group to send to, as a unicode string, and the message to send, as a serializable dict. It may raise MessageTooLarge but cannot raise ChannelFull.
  • group_expiry, an integer number of seconds that specifies how long group membership is valid for after the most recent group_add call (see Persistence below)

A channel layer implementing the statistics extension must also provide:

  • global_statistics(), a callable that returns statistics across all channels
  • channel_statistics(channel), a callable that returns statistics for specified channel
  • in both cases statistics are a dict with zero or more of (unicode string keys):
    • messages_count, the number of messages processed since server start
    • messages_count_per_second, the number of messages processed in the last second
    • messages_pending, the current number of messages waiting
    • messages_max_age, how long the oldest message has been waiting, in seconds
    • channel_full_count, the number of times ChannelFull exception has been risen since server start
    • channel_full_count_per_second, the number of times ChannelFull exception has been risen in the last second
  • Implementation may provide total counts, counts per seconds or both.

A channel layer implementing the flush extension must also provide:

  • flush(), a callable that resets the channel layer to a blank state, containing no messages and no groups (if the groups extension is implemented). This call must block until the system is cleared and will consistently look empty to any client, if the channel layer is distributed.

A channel layer implementing the twisted extension must also provide:

  • receive_twisted(channels), a function that behaves like receive but that returns a Twisted Deferred that eventually returns either (channel, message) or (None, None). It is not possible to run it in nonblocking mode; use the normal receive for that.

A channel layer implementing the async extension must also provide:

  • receive_async(channels), a function that behaves like receive but that fulfills the asyncio coroutine contract to block until either a result is available or an internal timeout is reached and (None, None) is returned. It is not possible to run it in nonblocking mode; use the normal receive for that.

Channel Semantics

Channels must:

  • Preserve ordering of messages perfectly with only a single reader and writer if the channel is a single-reader or process-specific channel.
  • Never deliver a message more than once.
  • Never block on message send (though they may raise ChannelFull or MessageTooLarge)
  • Be able to handle messages of at least 1MB in size when encoded as JSON (the implementation may use better encoding or compression, as long as it meets the equivalent size)
  • Have a maximum name length of at least 100 bytes.

They should attempt to preserve ordering in all cases as much as possible, but perfect global ordering is obviously not possible in the distributed case.

They are not expected to deliver all messages, but a success rate of at least 99.99% is expected under normal circumstances. Implementations may want to have a “resilience testing” mode where they deliberately drop more messages than usual so developers can test their code’s handling of these scenarios.

Persistence

Channel layers do not need to persist data long-term; group memberships only need to live as long as a connection does, and messages only as long as the message expiry time, which is usually a couple of minutes.

That said, if a channel server goes down momentarily and loses all data, persistent socket connections will continue to transfer incoming data and send out new generated data, but will have lost all of their group memberships and in-flight messages.

In order to avoid a nasty set of bugs caused by these half-deleted sockets, protocol servers should quit and hard restart if they detect that the channel layer has gone down or lost data; shedding all existing connections and letting clients reconnect will immediately resolve the problem.

If a channel layer implements the groups extension, it must persist group membership until at least the time when the member channel has a message expire due to non-consumption, after which it may drop membership at any time. If a channel subsequently has a successful delivery, the channel layer must then not drop group membership until another message expires on that channel.

Channel layers must also drop group membership after a configurable long timeout after the most recent group_add call for that membership, the default being 86,400 seconds (one day). The value of this timeout is exposed as the group_expiry property on the channel layer.

Protocol servers must have a configurable timeout value for every connection-based protocol they serve that closes the connection after the timeout, and should default this value to the value of group_expiry, if the channel layer provides it. This allows old group memberships to be cleaned up safely, knowing that after the group expiry the original connection must have closed, or is about to be in the next few seconds.

It’s recommended that end developers put the timeout setting much lower - on the order of hours or minutes - to enable better protocol design and testing. Even with ASGI’s separation of protocol server restart from business logic restart, you will likely need to move and reprovision protocol servers, and making sure your code can cope with this is important.

Message Formats

These describe the standardized message formats for the protocols this specification supports. All messages are dicts at the top level, and all keys are required unless explicitly marked as optional. If a key is marked optional, a default value is specified, which is to be assumed if the key is missing. Keys are unicode strings.

The one common key across all protocols is reply_channel, a way to indicate the client-specific channel to send responses to. Protocols are generally encouraged to have one message type and one reply channel type to ensure ordering.

A reply_channel should be unique per connection. If the protocol in question can have any server service a response - e.g. a theoretical SMS protocol - it should not have reply_channel attributes on messages, but instead a separate top-level outgoing channel.

Messages are specified here along with the channel names they are expected on; if a channel name can vary, such as with reply channels, the varying portion will be represented by !, such as http.response!, which matches the format the new_channel callable takes.

There is no label on message types to say what they are; their type is implicit in the channel name they are received on. Two types that are sent on the same channel, such as HTTP responses and response chunks, are distinguished apart by their required fields.

Message formats can be found in the sub-specifications:

Protocol Format Guidelines

Message formats for protocols should follow these rules, unless a very good performance or implementation reason is present:

  • reply_channel should be unique per logical connection, and not per logical client.
  • If the protocol has server-side state, entirely encapsulate that state in the protocol server; do not require the message consumers to use an external state store.
  • If the protocol has low-level negotiation, keepalive or other features, handle these within the protocol server and don’t expose them in ASGI messages.
  • If the protocol has guaranteed ordering and does not use a specific channel for a given connection (as HTTP does for body data), ASGI messages should include an order field (0-indexed) that preserves the ordering as received by the protocol server (or as sent by the client, if available). This ordering should span all message types emitted by the client - for example, a connect message might have order 0, and the first two frames order 1 and 2.
  • If the protocol is datagram-based, one datagram should equal one ASGI message (unless size is an issue)

Approximate Global Ordering

While maintaining true global (across-channels) ordering of messages is entirely unreasonable to expect of many implementations, they should strive to prevent busy channels from overpowering quiet channels.

For example, imagine two channels, busy, which spikes to 1000 messages a second, and quiet, which gets one message a second. There’s a single consumer running receive(['busy', 'quiet']) which can handle around 200 messages a second.

In a simplistic for-loop implementation, the channel layer might always check busy first; it always has messages available, and so the consumer never even gets to see a message from quiet, even if it was sent with the first batch of busy messages.

A simple way to solve this is to randomize the order of the channel list when looking for messages inside the channel layer; other, better methods are also available, but whatever is chosen, it should try to avoid a scenario where a message doesn’t get received purely because another channel is busy.

Strings and Unicode

In this document, and all sub-specifications, byte string refers to str on Python 2 and bytes on Python 3. If this type still supports Unicode codepoints due to the underlying implementation, then any values should be kept within the 0 - 255 range.

Unicode string refers to unicode on Python 2 and str on Python 3. This document will never specify just string - all strings are one of the two exact types.

Some serializers, such as json, cannot differentiate between byte strings and unicode strings; these should include logic to box one type as the other (for example, encoding byte strings as base64 unicode strings with a preceding special character, e.g. U+FFFF).

Channel and group names are always unicode strings, with the additional limitation that they only use the following characters:

  • ASCII letters
  • The digits 0 through 9
  • Hyphen -
  • Underscore _
  • Period .
  • Question mark ? (only to delineiate single-reader channel names, and only one per name)
  • Exclamation mark ! (only to delineate process-specific channel names, and only one per name)

Common Questions

  1. Why are messages dicts, rather than a more advanced type?

    We want messages to be very portable, especially across process and machine boundaries, and so a simple encodable type seemed the best way. We expect frameworks to wrap each protocol-specific set of messages in custom classes (e.g. http.request messages become Request objects)