24 KiB
Extension pipelining
websocket-extensions models the extension negotiation and processing pipeline
of the WebSocket protocol. Between the driver parsing messages from the TCP
stream and handing those messages off to the application, there may exist a
stack of extensions that transform the message somehow.
In the parlance of this framework, a session refers to a single instance of an
extension, acting on a particular socket on either the server or the client
side. A session may transform messages both incoming to the application and
outgoing from the application, for example the permessage-deflate extension
compresses outgoing messages and decompresses incoming messages. Message streams
in either direction are independent; that is, incoming and outgoing messages
cannot be assumed to 'pair up' as in a request-response protocol.
Asynchronous processing of messages poses a number of problems that this pipeline construction is intended to solve.
Overview
Logically, we have the following:
+-------------+ out +---+ +---+ +---+ +--------+
| |------>| |---->| |---->| |------>| |
| Application | | A | | B | | C | | Driver |
| |<------| |<----| |<----| |<------| |
+-------------+ in +---+ +---+ +---+ +--------+
\ /
+----------o----------+
|
sessions
For outgoing messages, the driver receives the result of
C.outgoing(B.outgoing(A.outgoing(message)))
or, [A, B, C].reduce(((m, ext) => ext.outgoing(m)), message)
For incoming messages, the application receives the result of
A.incoming(B.incoming(C.incoming(message)))
or, [C, B, A].reduce(((m, ext) => ext.incoming(m)), message)
A session is of the following type, to borrow notation from pseudo-Haskell:
type Session = {
incoming :: Message -> Message
outgoing :: Message -> Message
close :: () -> ()
}
(That () -> () syntax is intended to mean that close() is a nullary void
method; I apologise to any Haskell readers for not using the right monad.)
The incoming() and outgoing() methods perform message transformation in the
respective directions; close() is called when a socket closes so the session
can release any resources it's holding, for example a DEFLATE de/compression
context.
However because this is JavaScript, the incoming() and outgoing() methods
may be asynchronous (indeed, permessage-deflate is based on zlib, whose API
is stream-based). So their interface is strictly:
type Session = {
incoming :: Message -> Callback -> ()
outgoing :: Message -> Callback -> ()
close :: () -> ()
}
type Callback = Either Error Message -> ()
This means a message m2 can be pushed into a session while it's still processing the preceding message m1. The messages can be processed concurrently but they must be given to the next session in line (or to the application) in the same order they came in. Applications will expect to receive messages in the order they arrived over the wire, and sessions require this too. So ordering of messages must be preserved throughout the pipeline.
Consider the following highly simplified extension that deflates messages on the
wire. message is a value conforming the type:
type Message = {
rsv1 :: Boolean
rsv2 :: Boolean
rsv3 :: Boolean
opcode :: Number
data :: Buffer
}
Here's the extension:
var zlib = require('zlib');
var deflate = {
outgoing: function(message, callback) {
zlib.deflateRaw(message.data, function(error, result) {
message.rsv1 = true;
message.data = result;
callback(error, message);
});
},
incoming: function(message, callback) {
// decompress inbound messages (elided)
},
close: function() {
// no state to clean up
}
};
We can call it with a large message followed by a small one, and the small one will be returned first:
var crypto = require('crypto'),
large = crypto.randomBytes(1 << 14),
small = new Buffer('hi');
deflate.outgoing({data: large}, function() {
console.log(1, 'large');
});
deflate.outgoing({data: small}, function() {
console.log(2, 'small');
});
/* prints: 2 'small'
1 'large' */
So a session that processes messages asynchronously may fail to preserve message ordering.
Now, this extension is stateless, so it can process messages in any order and still produce the same output. But some extensions are stateful and require message order to be preserved.
For example, when using permessage-deflate without no_context_takeover set,
the session retains a DEFLATE de/compression context between messages, which
accumulates state as it consumes data (later messages can refer to sections of
previous ones to improve compression). Reordering parts of the DEFLATE stream
will result in a failed decompression. Messages must be decompressed in the same
order they were compressed by the peer in order for the DEFLATE protocol to
work.
Finally, there is the problem of closing a socket. When a WebSocket is closed by the application, or receives a closing request from the other peer, there may be messages outgoing from the application and incoming from the peer in the pipeline. If we close the socket and pipeline immediately, two problems arise:
- We may send our own closing frame to the peer before all prior messages we sent have been written to the socket, and before we have finished processing all prior messages from the peer
- The session may be instructed to close its resources (e.g. its de/compression context) while it's in the middle of processing a message, or before it has received messages that are upstream of it in the pipeline
Essentially, we must defer closing the sessions and sending a closing frame until after all prior messages have exited the pipeline.
Design goals
- Message order must be preserved between the protocol driver, the extension sessions, and the application
- Messages should be handed off to sessions and endpoints as soon as possible, to maximise throughput of stateless sessions
- The closing procedure should block any further messages from entering the pipeline, and should allow all existing messages to drain
- Sessions should be closed as soon as possible to prevent them holding memory and other resources when they have no more messages to handle
- The closing API should allow the caller to detect when the pipeline is empty and it is safe to continue the WebSocket closing procedure
- Individual extensions should remain as simple as possible to facilitate modularity and independent authorship
The final point about modularity is an important one: this framework is designed to facilitate extensions existing as plugins, by decoupling the protocol driver, extensions, and application. In an ideal world, plugins should only need to contain code for their specific functionality, and not solve these problems that apply to all sessions. Also, solving some of these problems requires consideration of all active sessions collectively, which an individual session is incapable of doing.
For example, it is entirely possible to take the simple deflate extension
above and wrap its incoming() and outgoing() methods in two Transform
streams, producing this type:
type Session = {
incoming :: TransformStream
outtoing :: TransformStream
close :: () -> ()
}
The Transform class makes it easy to wrap an async function such that message
order is preserved:
var stream = require('stream'),
session = new stream.Transform({objectMode: true});
session._transform = function(message, _, callback) {
var self = this;
deflate.outgoing(message, function(error, result) {
self.push(result);
callback();
});
};
However, this has a negative impact on throughput: it works by deferring
callback() until the async function has 'returned', which blocks Transform
from passing further input into the _transform() method until the current
message is dealt with completely. This would prevent sessions from processing
messages concurrently, and would unnecessarily reduce the throughput of
stateless extensions.
So, input should be handed off to sessions as soon as possible, and all we need is a mechanism to reorder the output so that message order is preserved for the next session in line.
Solution
We now describe the model implemented here and how it meets the above design goals. The above diagram where a stack of extensions sit between the driver and application describes the data flow, but not the object graph. That looks like this:
+--------+
| Driver |
+---o----+
|
V
+------------+ +----------+
| Extensions o----->| Pipeline |
+------------+ +-----o----+
|
+---------------+---------------+
| | |
+-----o----+ +-----o----+ +-----o----+
| Cell [A] | | Cell [B] | | Cell [C] |
+----------+ +----------+ +----------+
A driver using this framework holds an instance of the Extensions class, which
it uses to register extension plugins, negotiate headers and transform messages.
The Extensions instance itself holds a Pipeline, which contains an array of
Cell objects, each of which wraps one of the sessions.
Message processing
Both the Pipeline and Cell classes have incoming() and outgoing()
methods; the Pipeline interface pushes messages into the pipe, delegates the
message to each Cell in turn, then returns it back to the driver. Outgoing
messages pass through A then B then C, and incoming messages in the
reverse order.
Internally, a Cell contains two Functor objects. A Functor wraps an async
function and makes sure its output messages maintain the order of its input
messages. This name is due to @fronx, on the basis
that, by preserving message order, the abstraction preserves the mapping
between input and output messages. To use our simple deflate extension from
above:
var functor = new Functor(deflate, 'outgoing');
functor.call({data: large}, function() {
console.log(1, 'large');
});
functor.call({data: small}, function() {
console.log(2, 'small');
});
/* -> 1 'large'
2 'small' */
A Cell contains two of these, one for each direction:
+-----------------------+
+---->| Functor [A, incoming] |
+----------+ | +-----------------------+
| Cell [A] o------+
+----------+ | +-----------------------+
+---->| Functor [A, outgoing] |
+-----------------------+
This satisfies the message transformation requirements: the Pipeline simply
loops over the cells in the appropriate direction to transform each message.
Because each Cell will preserve message order, we can pass a message to the
next Cell in line as soon as the current Cell returns it. This gives each
Cell all the messages in order while maximising throughput.
Session closing
We want to close each session as soon as possible, after all existing messages
have drained. To do this, each Cell begins with a pending message counter in
each direction, labelled in and out below.
+----------+
| Pipeline |
+-----o----+
|
+---------------+---------------+
| | |
+-----o----+ +-----o----+ +-----o----+
| Cell [A] | | Cell [B] | | Cell [C] |
+----------+ +----------+ +----------+
in: 0 in: 0 in: 0
out: 0 out: 0 out: 0
When a message m1 enters the pipeline, say in the outgoing direction, we
increment the pending.out counter on all cells immediately.
+----------+
m1 => | Pipeline |
+-----o----+
|
+---------------+---------------+
| | |
+-----o----+ +-----o----+ +-----o----+
| Cell [A] | | Cell [B] | | Cell [C] |
+----------+ +----------+ +----------+
in: 0 in: 0 in: 0
out: 1 out: 1 out: 1
m1 is handed off to A, meanwhile a second message m2 arrives in the same
direction. All pending.out counters are again incremented.
+----------+
m2 => | Pipeline |
+-----o----+
|
+---------------+---------------+
m1 | | |
+-----o----+ +-----o----+ +-----o----+
| Cell [A] | | Cell [B] | | Cell [C] |
+----------+ +----------+ +----------+
in: 0 in: 0 in: 0
out: 2 out: 2 out: 2
When the first cell's A.outgoing functor finishes processing m1, the first
pending.out counter is decremented and m1 is handed off to cell B.
+----------+
| Pipeline |
+-----o----+
|
+---------------+---------------+
m2 | m1 | |
+-----o----+ +-----o----+ +-----o----+
| Cell [A] | | Cell [B] | | Cell [C] |
+----------+ +----------+ +----------+
in: 0 in: 0 in: 0
out: 1 out: 2 out: 2
As B finishes with m1, and as A finishes with m2, the pending.out
counters continue to decrement.
+----------+
| Pipeline |
+-----o----+
|
+---------------+---------------+
| m2 | m1 |
+-----o----+ +-----o----+ +-----o----+
| Cell [A] | | Cell [B] | | Cell [C] |
+----------+ +----------+ +----------+
in: 0 in: 0 in: 0
out: 0 out: 1 out: 2
Say C is a little slow, and begins processing m2 while still processing
m1. That's fine, the Functor mechanism will keep m1 ahead of m2 in the
output.
+----------+
| Pipeline |
+-----o----+
|
+---------------+---------------+
| | m2 | m1
+-----o----+ +-----o----+ +-----o----+
| Cell [A] | | Cell [B] | | Cell [C] |
+----------+ +----------+ +----------+
in: 0 in: 0 in: 0
out: 0 out: 0 out: 2
Once all messages are dealt with, the counters return to 0.
+----------+
| Pipeline |
+-----o----+
|
+---------------+---------------+
| | |
+-----o----+ +-----o----+ +-----o----+
| Cell [A] | | Cell [B] | | Cell [C] |
+----------+ +----------+ +----------+
in: 0 in: 0 in: 0
out: 0 out: 0 out: 0
The same process applies in the incoming direction, the only difference being
that messages are passed to C first.
This makes closing the sessions quite simple. When the driver wants to close the
socket, it calls Pipeline.close(). This immediately calls close() on all
the cells. If a cell has in == out == 0, then it immediately calls
session.close(). Otherwise, it stores the closing call and defers it until
in and out have both ticked down to zero. The pipeline will not accept new
messages after close() has been called, so we know the pending counts will not
increase after this point.
This means each session is closed as soon as possible: A can close while the
slow C session is still working, because it knows there are no more messages
on the way. Similarly, C will defer closing if close() is called while m1
is still in B, and m2 in A, because its pending count means it knows it
has work yet to do, even if it's not received those messages yet. This concern
cannot be addressed by extensions acting only on their own local state, unless
we pollute individual extensions by making them all implement this same
mechanism.
The actual closing API at each level is slightly different:
type Session = {
close :: () -> ()
}
type Cell = {
close :: () -> Promise ()
}
type Pipeline = {
close :: Callback -> ()
}
This might appear inconsistent so it's worth explaining. Remember that a
Pipeline holds a list of Cell objects, each wrapping a Session. The driver
talks (via the Extensions API) to the Pipeline interface, and it wants
Pipeline.close() to do two things: close all the sessions, and tell me when
it's safe to start the closing procedure (i.e. when all messages have drained
from the pipe and been handed off to the application or socket). A callback API
works well for that.
At the other end of the stack, Session.close() is a nullary void method with
no callback or promise API because we don't care what it does, and whatever it
does do will not block the WebSocket protocol; we're not going to hold off
processing messages while a session closes its de/compression context. We just
tell it to close itself, and don't want to wait while it does that.
In the middle, Cell.close() returns a promise rather than using a callback.
This is for two reasons. First, Cell.close() might not do anything
immediately, it might have to defer its effect while messages drain. So, if
given a callback, it would have to store it in a queue for later execution.
Callbacks work fine if your method does something and can then invoke the
callback itself, but if you need to store callbacks somewhere so another method
can execute them, a promise is a better fit. Second, it better serves the
purposes of Pipeline.close(): it wants to call close() on each of a list of
cells, and wait for all of them to finish. This is simple and idiomatic using
promises:
var closed = cells.map((cell) => cell.close());
Promise.all(closed).then(callback);
(We don't actually use a full Promises/A+ compatible promise here, we use a much simplified construction that acts as a callback aggregater and resolves synchronously and does not support chaining, but the principle is the same.)
Error handling
We've not mentioned error handling so far but it bears some explanation. The above counter system still applies, but behaves slightly differently in the presence of errors.
Say we push three messages into the pipe in the outgoing direction:
+----------+
m3, m2, m1 => | Pipeline |
+-----o----+
|
+---------------+---------------+
| | |
+-----o----+ +-----o----+ +-----o----+
| Cell [A] | | Cell [B] | | Cell [C] |
+----------+ +----------+ +----------+
in: 0 in: 0 in: 0
out: 3 out: 3 out: 3
They pass through the cells successfully up to this point:
+----------+
| Pipeline |
+-----o----+
|
+---------------+---------------+
m3 | m2 | m1 |
+-----o----+ +-----o----+ +-----o----+
| Cell [A] | | Cell [B] | | Cell [C] |
+----------+ +----------+ +----------+
in: 0 in: 0 in: 0
out: 1 out: 2 out: 3
At this point, session B produces an error while processing m2, that is m2
becomes e2. m1 is still in the pipeline, and m3 is queued behind m2.
What ought to happen is that m1 is handed off to the socket, then m2 is
released to the driver, which will detect the error and begin closing the
socket. No further processing should be done on m3 and it should not be
released to the driver after the error is emitted.
To handle this, we allow errors to pass down the pipeline just like messages do,
to maintain ordering. But, once a cell sees its session produce an error, or it
receives an error from upstream, it should refuse to accept any further
messages. Session B might have begun processing m3 by the time it produces
the error e2, but C will have been given e2 before it receives m3, and
can simply drop m3.
Now, say e2 reaches the slow session C while m1 is still present,
meanwhile m3 has been dropped. C will never receive m3 since it will have
been dropped upstream. Under the present model, its out counter will be 3
but it is only going to emit two more values: m1 and e2. In order for
closing to work, we need to decrement out to reflect this. The situation
should look like this:
+----------+
| Pipeline |
+-----o----+
|
+---------------+---------------+
| | e2 | m1
+-----o----+ +-----o----+ +-----o----+
| Cell [A] | | Cell [B] | | Cell [C] |
+----------+ +----------+ +----------+
in: 0 in: 0 in: 0
out: 0 out: 0 out: 2
When a cell sees its session emit an error, or when it receives an error from upstream, it sets its pending count in the appropriate direction to equal the number of messages it is currently processing. It will not accept any messages after it sees the error, so this will allow the counter to reach zero.
Note that while e2 is in the pipeline, Pipeline should drop any further
messages in the outgoing direction, but should continue to accept incoming
messages. Until e2 makes it out of the pipe to the driver, behind previous
successful messages, the driver does not know an error has happened, and a
message may arrive over the socket and make it all the way through the incoming
pipe in the meantime. We only halt processing in the affected direction to avoid
doing unnecessary work since messages arriving after an error should not be
processed.
Some unnecessary work may happen, for example any messages already in the
pipeline following m2 will be processed by A, since it's upstream of the
error. Those messages will be dropped by B.
Alternative ideas
I am considering implementing Functor as an object-mode transform stream
rather than what is essentially an async function. Being object-mode, a stream
would preserve message boundaries and would also possibly help address
back-pressure. I'm not sure whether this would require external API changes so
that such streams could be connected to the downstream driver's streams.
Acknowledgements
Credit is due to @mnowster for helping with the design and to @fronx for helping name things.