API documentation

h11 has a fairly small public API, with all public symbols available directly at the top level:

In [1]: import h11

In [2]: h11.<TAB>
h11.CLIENT                 h11.MUST_CLOSE
h11.CLOSED                 h11.NEED_DATA
h11.Connection             h11.PAUSED
h11.ConnectionClosed       h11.PRODUCT_ID
h11.Data                   h11.ProtocolError
h11.DONE                   h11.RemoteProtocolError
h11.EndOfMessage           h11.Request
h11.ERROR                  h11.Response
h11.IDLE                   h11.SEND_BODY
h11.InformationalResponse  h11.SEND_RESPONSE
h11.LocalProtocolError     h11.SERVER
h11.MIGHT_SWITCH_PROTOCOL  h11.SWITCHED_PROTOCOL

These symbols fall into three main categories: event classes, special constants used to track different connection states, and the Connection class itself. We’ll describe them in that order.

Events

Events are the core of h11: the whole point of h11 is to let you think about HTTP transactions as being a series of events sent back and forth between a client and a server, instead of thinking in terms of bytes.

All events behave in essentially similar ways. Let’s take Request as an example. Like all events, this is a “final” class – you cannot subclass it. And like all events, it has several fields. For Request, there are four of them: method, target, headers, and http_version. http_version defaults to b"1.1"; the rest have no default, so to create a Request you have to specify their values:

In [3]: req = h11.Request(method="GET",
   ...:                   target="/",
   ...:                   headers=[("Host", "example.com")])
   ...: 

Event constructors accept only keyword arguments, not positional arguments.

Events have a useful repr:

In [4]: req
Out[4]: Request(method=b'GET', target=b'/', headers=[(b'host', b'example.com')], http_version=b'1.1')

And their fields are available as regular attributes:

In [5]: req.method
Out[5]: b'GET'

In [6]: req.target
Out[6]: b'/'

In [7]: req.headers
Out[7]: [(b'host', b'example.com')]

In [8]: req.http_version
Out[8]: b'1.1'

Notice that these attributes have been normalized to byte-strings. In general, events normalize and validate their fields when they’re constructed. Some of these normalizations and checks are specific to a particular event – for example, Request enforces RFC 7230’s requirement that HTTP/1.1 requests must always contain a "Host" header:

# HTTP/1.0 requests don't require a Host: header
In [9]: h11.Request(method="GET", target="/", headers=[], http_version="1.0")
Out[9]: Request(method=b'GET', target=b'/', headers=[], http_version=b'1.0')
# But HTTP/1.1 requests do
In [10]: h11.Request(method="GET", target="/", headers=[])
---------------------------------------------------------------------------
LocalProtocolError                        Traceback (most recent call last)
<ipython-input-10-645a3ef89e11> in <module>()
----> 1 h11.Request(method="GET", target="/", headers=[])

/home/docs/checkouts/readthedocs.org/user_builds/h11/envs/latest/lib/python3.5/site-packages/h11-0.7.0+dev-py3.5.egg/h11/_events.py in __init__(self, **kwargs)
     53                 raise LocalProtocolError("status code must be integer")
     54 
---> 55         self._validate()
     56 
     57     def _validate(self):

/home/docs/checkouts/readthedocs.org/user_builds/h11/envs/latest/lib/python3.5/site-packages/h11-0.7.0+dev-py3.5.egg/h11/_events.py in _validate(self)
    125                 host_count += 1
    126         if self.http_version == b"1.1" and host_count == 0:
--> 127             raise LocalProtocolError("Missing mandatory Host: header")
    128         if host_count > 1:
    129             raise LocalProtocolError("Found multiple Host: headers")

LocalProtocolError: Missing mandatory Host: header

This helps protect you from accidentally violating the protocol, and also helps protect you from remote peers who attempt to violate the protocol.

A few of these normalization rules are standard across multiple events, so we document them here:

headers: In h11, headers are represented internally as a list of (name, value) pairs, where name and value are both byte-strings, name is always lowercase, and name and value are both guaranteed not to have any leading or trailing whitespace. When constructing an event, we accept any iterable of pairs like this, and will automatically convert native strings containing ascii or bytes-like objects to byte-strings and convert names to lowercase:

In [11]: original_headers = [("HOST", bytearray(b"Example.Com"))]

In [12]: req = h11.Request(method="GET", target="/", headers=original_headers)

In [13]: original_headers
Out[13]: [('HOST', bytearray(b'Example.Com'))]

In [14]: req.headers
Out[14]: [(b'host', b'Example.Com')]

If any names are detected with leading or trailing whitespace, then this is an error (“in the past, differences in the handling of such whitespace have led to security vulnerabilities” – RFC 7230). We also check for certain other protocol violations, e.g. it’s always illegal to have a newline inside a header value, and Content-Length: hello is an error because Content-Length should always be an integer. We may add additional checks in the future.

It’s not just headers we normalize to being byte-strings: the same type-conversion logic is also applied to the Request.method and Request.target field, and – for consistency – all http_version fields. In particular, we always represent HTTP version numbers as byte-strings like b"1.1". Bytes-like objects and native strings will be automatically converted to byte strings. Note that the HTTP standard specifically guarantees that all HTTP version numbers will consist of exactly two digits separated by a dot, so comparisons like req.http_version < b"1.1" are safe and valid.

When manually constructing an event, you generally shouldn’t specify http_version, because it defaults to b"1.1", and if you attempt to override this to some other value then Connection.send() will reject your event – h11 only speaks HTTP/1.1. But it does understand other versions of HTTP, so you might receive events with other http_version values from remote peers.

Here’s the complete set of events supported by h11:

class h11.Request(**kwargs)[source]

The beginning of an HTTP request.

Fields:

method

An HTTP method, e.g. b"GET" or b"POST". Always a byte string. Bytes-like objects and native strings containing only ascii characters will be automatically converted to byte strings.

target

The target of an HTTP request, e.g. b"/index.html", or one of the more exotic formats described in RFC 7320, section 5.3. Always a byte string. Bytes-like objects and native strings containing only ascii characters will be automatically converted to byte strings.

headers

Request headers, represented as a list of (name, value) pairs. See the header normalization rules for details.

http_version

The HTTP protocol version, represented as a byte string like b"1.1". See the HTTP version normalization rules for details.

class h11.InformationalResponse(**kwargs)[source]

An HTTP informational response.

Fields:

status_code

The status code of this response, as an integer. For an InformationalResponse, this is always in the range [100, 200).

headers

Request headers, represented as a list of (name, value) pairs. See the header normalization rules for details.

http_version

The HTTP protocol version, represented as a byte string like b"1.1". See the HTTP version normalization rules for details.

reason

The reason phrase of this response, as a byte string. For example: b"OK", or b"Not Found".

class h11.Response(**kwargs)[source]

The beginning of an HTTP response.

Fields:

status_code

The status code of this response, as an integer. For an Response, this is always in the range [200, 600).

headers

Request headers, represented as a list of (name, value) pairs. See the header normalization rules for details.

http_version

The HTTP protocol version, represented as a byte string like b"1.1". See the HTTP version normalization rules for details.

reason

The reason phrase of this response, as a byte string. For example: b"OK", or b"Not Found".

class h11.Data(**kwargs)[source]

Part of an HTTP message body.

Fields:

data

A bytes-like object containing part of a message body. Or, if using the combine=False argument to Connection.send(), then any object that your socket writing code knows what to do with, and for which calling len() returns the number of bytes that will be written – see Support for sendfile() for details.

chunk_start

A marker that indicates whether this data object is from the start of a chunked transfer encoding chunk. This field is ignored when when a Data event is provided to Connection.send(): it is only valid on events emitted from Connection.next_event(). You probably shouldn’t use this attribute at all; see Chunked Transfer Encoding Delimiters for details.

chunk_end

A marker that indicates whether this data object is the last for a given chunked transfer encoding chunk. This field is ignored when when a Data event is provided to Connection.send(): it is only valid on events emitted from Connection.next_event(). You probably shouldn’t use this attribute at all; see Chunked Transfer Encoding Delimiters for details.

class h11.EndOfMessage(**kwargs)[source]

The end of an HTTP message.

Fields:

headers

Default value: []

Any trailing headers attached to this message, represented as a list of (name, value) pairs. See the header normalization rules for details.

Must be empty unless Transfer-Encoding: chunked is in use.

class h11.ConnectionClosed(**kwargs)[source]

This event indicates that the sender has closed their outgoing connection.

Note that this does not necessarily mean that they can’t receive further data, because TCP connections are composed to two one-way channels which can be closed independently. See Closing connections for details.

No fields.

The state machine

Now that you know what the different events are, the next question is: what can you do with them?

A basic HTTP request/response cycle looks like this:

And once that’s finished, both sides either close the connection, or they go back to the top and re-use it for another request/response cycle.

To coordinate this interaction, the h11 Connection object maintains several state machines: one that tracks what the client is doing, one that tracks what the server is doing, and a few more tiny ones to track whether keep-alive is enabled and whether the client has proposed to switch protocols. h11 always keeps track of all of these state machines, regardless of whether it’s currently playing the client or server role.

The state machines look like this (click on each to expand):

client-image
server-image
special-image

If you squint at the first two diagrams, you can see the client’s IDLE -> SEND_BODY -> DONE path and the server’s IDLE -> SEND_RESPONSE -> SEND_BODY -> DONE path, which encode the basic sequence of events we described above. But there’s a fair amount of other stuff going on here as well.

The first thing you should notice is the different colors. These correspond to the different ways that our state machines can change state.

  • Dark blue arcs are event-triggered transitions: if we’re in state A, and this event happens, when we switch to state B. For the client machine, these transitions always happen when the client sends an event. For the server machine, most of them involve the server sending an event, except that the server also goes from IDLE -> SEND_RESPONSE when the client sends a Request.
  • Green arcs are state-triggered transitions: these are somewhat unusual, and are used to couple together the different state machines – if, at any moment, one machine is in state A and another machine is in state B, then the first machine immediately transitions to state C. For example, if the CLIENT machine is in state DONE, and the SERVER machine is in the CLOSED state, then the CLIENT machine transitions to MUST_CLOSE. And the same thing happens if the CLIENT machine is in the state DONE and the keep-alive machine is in the state disabled.
  • There are also two purple arcs labeled start_next_cycle(): these correspond to an explicit method call documented below.

Here’s why we have all the stuff in those diagrams above, beyond what’s needed to handle the basic request/response cycle:

  • Server sending a Response directly from IDLE: This is used for error responses, when the client’s request never arrived (e.g. 408 Request Timed Out) or was unparseable gibberish (400 Bad Request) and thus didn’t register with our state machine as a real Request.
  • The transitions involving MUST_CLOSE and CLOSE: keep-alive and shutdown handling; see Re-using a connection: keep-alive and pipelining and Closing connections.
  • The transitions involving MIGHT_SWITCH_PROTOCOL and SWITCHED_PROTOCOL: See Switching protocols.
  • That weird ERROR state hanging out all lonely on the bottom: to avoid cluttering the diagram, we don’t draw any arcs coming into this node, but that doesn’t mean it can’t be entered. In fact, it can be entered from any state: if any exception occurs while trying to send/receive data, then the corresponding machine will transition directly to this state. Once there, though, it can never leave – that part of the diagram is accurate. See Error handling.

And finally, note that in these diagrams, all the labels that are in italics are informal English descriptions of things that happen in the code, while the labels in upright text correspond to actual objects in the public API. You’ve already seen the event objects like Request and Response; there are also a set of opaque sentinel values that you can use to track and query the client and server’s states.

Special constants

h11 exposes some special constants corresponding to the different states in the client and server state machines described above. The complete list is:

h11.IDLE
h11.SEND_RESPONSE
h11.SEND_BODY
h11.DONE
h11.MUST_CLOSE
h11.CLOSED
h11.MIGHT_SWITCH_PROTOCOL
h11.SWITCHED_PROTOCOL
h11.ERROR

For example, we can see that initially the client and server start in state IDLE / IDLE:

In [15]: conn = h11.Connection(our_role=h11.CLIENT)

In [16]: conn.states
Out[16]: {CLIENT: IDLE, SERVER: IDLE}

And then if the client sends a Request, then the client switches to state SEND_BODY, while the server switches to state SEND_RESPONSE:

In [17]: conn.send(h11.Request(method="GET", target="/", headers=[("Host", "example.com")]));

In [18]: conn.states
Out[18]: {CLIENT: SEND_BODY, SERVER: SEND_RESPONSE}

And we can test these values directly using constants like SEND_BODY:

In [19]: conn.states[h11.CLIENT] is h11.SEND_BODY
Out[19]: True

This shows how the Connection type tracks these state machines and lets you query their current state.

The above also showed the special constants that can be used to indicate the two different roles that a peer can play in an HTTP connection:

h11.CLIENT
h11.SERVER

And finally, there are also two special constants that can be returned from Connection.next_event():

h11.NEED_DATA
h11.PAUSED

All of these behave the same, and their behavior is modeled after None: they’re opaque singletons, their __repr__() is their name, and you compare them with is.

Finally, h11’s constants have a quirky feature that can sometimes be useful: they are instances of themselves.

In [20]: type(h11.NEED_DATA) is h11.NEED_DATA
Out[20]: True

In [21]: type(h11.PAUSED) is h11.PAUSED
Out[21]: True

The main application of this is that when handling the return value from Connection.next_event(), which is sometimes an instance of an event class and sometimes NEED_DATA or PAUSED, you can always call type(event) to get something useful to dispatch one, using e.g. a handler table, functools.singledispatch(), or calling getattr(some_object, "handle_" + type(event).__name__). Not that this kind of dispatch-based strategy is always the best approach – but the option is there if you want it.

The Connection object

class h11.Connection(our_role, max_incomplete_event_size=16384)[source]

An object encapsulating the state of an HTTP connection.

Parameters:
  • our_role – If you’re implementing a client, pass h11.CLIENT. If you’re implementing a server, pass h11.SERVER.
  • max_incomplete_event_size (int) – The maximum number of bytes we’re willing to buffer of an incomplete event. In practice this mostly sets a limit on the maximum size of the request/response line + headers. If this is exceeded, then next_event() will raise RemoteProtocolError.
receive_data(data)[source]

Add data to our internal recieve buffer.

This does not actually do any processing on the data, just stores it. To trigger processing, you have to call next_event().

Parameters:

data (bytes-like object) –

The new data that was just received.

Special case: If data is an empty byte-string like b"", then this indicates that the remote side has closed the connection (end of file). Normally this is convenient, because standard Python APIs like file.read() or socket.recv() use b"" to indicate end-of-file, while other failures to read are indicated using other mechanisms like raising TimeoutError. When using such an API you can just blindly pass through whatever you get from read to receive_data(), and everything will work.

But, if you have an API where reading an empty string is a valid non-EOF condition, then you need to be aware of this and make sure to check for such strings and avoid passing them to receive_data().

Returns:

Nothing, but after calling this you should call next_event() to parse the newly received data.

Raises:

RuntimeError – Raised if you pass an empty data, indicating EOF, and then pass a non-empty data, indicating more data that somehow arrived after the EOF.

(Calling receive_data(b"") multiple times is fine, and equivalent to calling it once.)

next_event()[source]

Parse the next event out of our receive buffer, update our internal state, and return it.

This is a mutating operation – think of it like calling next() on an iterator.

Returns:One of three things:
  1. An event object – see Events.
  2. The special constant NEED_DATA, which indicates that you need to read more data from your socket and pass it to receive_data() before this method will be able to return any more events.
  3. The special constant PAUSED, which indicates that we are not in a state where we can process incoming data (usually because the peer has finished their part of the current request/response cycle, and you have not yet called start_next_cycle()). See Flow control for details.
Raises:RemoteProtocolError – The peer has misbehaved. You should close the connection (possibly after sending some kind of 4xx response).

Once this method returns ConnectionClosed once, then all subsequent calls will also return ConnectionClosed.

If this method raises any exception besides RemoteProtocolError then that’s a bug – if it happens please file a bug report!

If this method raises any exception then it also sets Connection.their_state to ERROR – see Error handling for discussion.

send(event)[source]

Convert a high-level event into bytes that can be sent to the peer, while updating our internal state machine.

Parameters:event – The event to send.
Returns:If type(event) is ConnectionClosed, then returns None. Otherwise, returns a bytes-like object.
Raises:LocalProtocolError – Sending this event at this time would violate our understanding of the HTTP/1.1 protocol.

If this method raises any exception then it also sets Connection.our_state to ERROR – see Error handling for discussion.

send_with_data_passthrough(event)[source]

Identical to send(), except that in situations where send() returns a single bytes-like object, this instead returns a list of them – and when sending a Data event, this list is guaranteed to contain the exact object you passed in as Data.data. See Support for sendfile() for discussion.

start_next_cycle()[source]

Attempt to reset our connection state for a new request/response cycle.

If both client and server are in DONE state, then resets them both to IDLE state in preparation for a new request/response cycle on this same connection. Otherwise, raises a LocalProtocolError.

See Re-using a connection: keep-alive and pipelining.

our_role

CLIENT if this is a client; SERVER if this is a server.

their_role

SERVER if this is a client; CLIENT if this is a server.

states

A dictionary like:

{CLIENT: <client state>, SERVER: <server state>}

See The state machine for details.

our_state

The current state of whichever role we are playing. See The state machine for details.

their_state

The current state of whichever role we are NOT playing. See The state machine for details.

their_http_version

The version of HTTP that our peer claims to support. None if we haven’t yet received a request/response.

This is preserved by start_next_cycle(), so it can be handy for a client making multiple requests on the same connection: normally you don’t know what version of HTTP the server supports until after you do a request and get a response – so on an initial request you might have to assume the worst. But on later requests on the same connection, the information will be available here.

client_is_waiting_for_100_continue

True if the client sent a request with the Expect: 100-continue header, and is still waiting for a response (i.e., the server has not sent a 100 Continue or any other kind of response, and the client has not gone ahead and started sending the body anyway).

See RFC 7231 section 5.1.1 for details.

they_are_waiting_for_100_continue

True if their_role is CLIENT and client_is_waiting_for_100_continue.

trailing_data

Data that has been received, but not yet processed, represented as a tuple with two elements, where the first is a byte-string containing the unprocessed data itself, and the second is a bool that is True if the receive connection was closed.

See Switching protocols for discussion of why you’d want this.

Error handling

Given the vagaries of networks and the folks on the other side of them, it’s extremely important to be prepared for errors.

Most errors in h11 are signaled by raising one of ProtocolError‘s two concrete base classes, LocalProtocolError and RemoteProtocolError:

exception h11.ProtocolError(msg, error_status_hint=400)[source]

Exception indicating a violation of the HTTP/1.1 protocol.

This as an abstract base class, with two concrete base classes: LocalProtocolError, which indicates that you tried to do something that HTTP/1.1 says is illegal, and RemoteProtocolError, which indicates that the remote peer tried to do something that HTTP/1.1 says is illegal. See Error handling for details.

In addition to the normal Exception features, it has one attribute:

error_status_hint

This gives a suggestion as to what status code a server might use if this error occurred as part of a request.

For a RemoteProtocolError, this is useful as a suggestion for how you might want to respond to a misbehaving peer, if you’re implementing a server.

For a LocalProtocolError, this can be taken as a suggestion for how your peer might have responded to you if h11 had allowed you to continue.

The default is 400 Bad Request, a generic catch-all for protocol violations.

exception h11.LocalProtocolError(msg, error_status_hint=400)[source]
exception h11.RemoteProtocolError(msg, error_status_hint=400)[source]

There are four cases where these exceptions might be raised:

  • When trying to instantiate an event object (LocalProtocolError): This indicates that something about your event is invalid. Your event wasn’t constructed, but there are no other consequences – feel free to try again.
  • When calling Connection.start_next_cycle() (LocalProtocolError): This indicates that the connection is not ready to be re-used, because one or both of the peers are not in the DONE state. The Connection object remains usable, and you can try again later.
  • When calling Connection.next_event() (RemoteProtocolError): This indicates that the remote peer has violated our protocol assumptions. This is unrecoverable – we don’t know what they’re doing and we cannot safely proceed. Connection.their_state immediately becomes ERROR, and all further calls to next_event() will also raise RemoteProtocolError. Connection.send() still works as normal, so if you’re implementing a server and this happens then you have an opportunity to send back a 400 Bad Request response. But aside from that, your only real option is to close your socket and make a new connection.
  • When calling Connection.send() or Connection.send_with_data_passthrough() (LocalProtocolError): This indicates that you violated our protocol assumptions. This is also unrecoverable – h11 doesn’t know what you’re doing, its internal state may be inconsistent, and we cannot safely proceed. Connection.our_state immediately becomes ERROR, and all further calls to send() will also raise LocalProtocolError. The only thing you can reasonably due at this point is to close your socket and make a new connection.

Message body framing: Content-Length and all that

There are two different headers that HTTP/1.1 uses to indicate a framing mechanism for request/response bodies: Content-Length and Transfer-Encoding. Our general philosophy is that the way you tell h11 what configuration you want to use is by setting the appropriate headers in your request / response, and then h11 will both pass those headers on to the peer and encode the body appropriately.

Currently, the only supported Transfer-Encoding is chunked.

On requests, this means:

  • No Content-Length or Transfer-Encoding: no body, equivalent to Content-Length: 0.

  • Content-Length: ...: You’re going to send exactly the specified number of bytes. h11 will keep track and signal an error if your EndOfMessage doesn’t happen at the right place.

  • Transfer-Encoding: chunked: You’re going to send a variable / not yet known number of bytes.

    Note 1: only HTTP/1.1 servers are required to support Transfer-Encoding: chunked, and as a client you have to decide whether to send this header before you get to see what protocol version the server is using.

    Note 2: even though HTTP/1.1 servers are required to support Transfer-Encoding: chunked, this doesn’t necessarily mean that they actually do – e.g., applications using Python’s standard WSGI API cannot accept chunked requests.

    Nonetheless, this is the only way to send request where you don’t know the size of the body ahead of time, so if that’s the situation you find yourself in then you might as well try it and hope.

On responses, things are a bit more subtle. There are effectively two cases:

  • Content-Length: ...: You’re going to send exactly the specified number of bytes. h11 will keep track and signal an error if your EndOfMessage doesn’t happen at the right place.

  • Transfer-Encoding: chunked, or, neither framing header is provided: These two cases are handled differently at the wire level, but as far as the application is concerned they provide (almost) exactly the same semantics: in either case, you’ll send a variable / not yet known number of bytes. The difference between them is that Transfer-Encoding: chunked works better (compatible with keep-alive, allows trailing headers, clearly distinguishes between successful completion and network errors), but requires an HTTP/1.1 client; for HTTP/1.0 clients the only option is the no-headers approach where you have to close the socket to indicate completion.

    Since this is (almost) entirely a wire-level-encoding concern, h11 abstracts it: when sending a response you can set either Transfer-Encoding: chunked or leave off both framing headers, and h11 will treat both cases identically: it will automatically pick the best option given the client’s advertised HTTP protocol level.

    You need to watch out for this if you’re using trailing headers (i.e., a non-empty headers attribute on EndOfMessage), since trailing headers are only legal if we actually ended up using Transfer-Encoding: chunked. Trying to send a non-empty set of trailing headers to a HTTP/1.0 client will raise a LocalProtocolError. If this use case is important to you, check Connection.their_http_version to confirm that the client speaks HTTP/1.1 before you attempt to send any trailing headers.

Re-using a connection: keep-alive and pipelining

HTTP/1.1 allows a connection to be re-used for multiple request/response cycles (also known as “keep-alive”). This can make things faster by letting us skip the costly connection setup, but it does create some complexities: we have to keep track of whether a connection is reusable, and when there are multiple requests and responses flowing through the same connection we need to be careful not to get confused about which request goes with which response.

h11 considers a connection to be reusable if, and only if, both sides (a) speak HTTP/1.1 (HTTP/1.0 did have some complex and fragile support for keep-alive bolted on, but h11 currently doesn’t support that – possibly this will be added in the future), and (b) neither side has explicitly disabled keep-alive by sending a Connection: close header.

If you plan to make only a single request or response and then close the connection, you should manually set the Connection: close header in your request/response. h11 will notice and update its state appropriately.

There are also some situations where you are required to send a Connection: close header, e.g. if you are a server talking to a client that doesn’t support keep-alive. You don’t need to worry about these cases – h11 will automatically add this header when necessary. Just worry about setting it when it’s actually something that you’re actively choosing.

If you want to re-use a connection, you have to wait until both the request and the response have been completed, bringing both the client and server to the DONE state. Once this has happened, you can explicitly call Connection.start_next_cycle() to reset both sides back to the IDLE state. This makes sure that the client and server remain synched up.

If keep-alive is disabled for whatever reason – someone set Connection: close, lack of protocol support, one of the sides just unilaterally closed the connection – then the state machines will skip past the DONE state directly to the MUST_CLOSE or CLOSED states. In this case, trying to call start_next_cycle() will raise an error, and the only thing you can legally do is to close this connection and make a new one.

HTTP/1.1 also allows for a more aggressive form of connection re-use, in which a client sends multiple requests in quick succession, and then waits for the responses to stream back in order (“pipelining”). This is generally considered to have been a bad idea, because it makes things like error recovery very complicated.

As a client, h11 does not support pipelining. This is enforced by the structure of the state machine: after sending one Request, you can’t send another until after calling start_next_cycle(), and you can’t call start_next_cycle() until the server has entered the DONE state, which requires reading the server’s full response.

As a server, h11 provides the minimal support for pipelining required to comply with the HTTP/1.1 standard: if the client sends multiple pipelined requests, then we the first request until we reach the DONE state, and then next_event() will pause and refuse to parse any more events until the response is completed and start_next_cycle() is called. See the next section for more details.

Flow control

Presumably you know when you want to send things, and the send() interface is very simple: it just immediately returns all the data you need to send for the given event, so you can apply whatever send buffer strategy you want. But reading from the remote peer is a bit trickier: you don’t want to read data from the remote peer if it can’t be processed (i.e., you want to apply backpressure and avoid building arbitrarily large in-memory buffers), and you definitely don’t want to block waiting on data from the remote peer at the same time that it’s blocked waiting for you, because that will cause a deadlock.

One complication here is that if you’re implementing a server, you have to be prepared to handle Requests that have an Expect: 100-continue header. You can read the spec for the full details, but basically what this header means is that after sending the Request, the client plans to pause and wait until they see some response from the server before they send that request’s Data. The server’s response would normally be an InformationalResponse with status 100 Continue, but it could be anything really (e.g. a full Response with a 4xx status code). The crucial thing as a server, though, is that you should never block trying to read a request body if the client is blocked waiting for you to tell them to send the request body.

Fortunately, h11 makes this easy, because it tracks whether the client is in the waiting-for-100-continue state, and exposes this as Connection.they_are_waiting_for_100_continue. So you don’t have to pay attention to the Expect header yourself; you just have to make sure that before you block waiting to read a request body, you execute some code like:

if conn.they_are_waiting_for_100_continue:
    do_send(conn, h11.InformationalResponse(100, headers=[...]))
do_read(...)

In fact, if you’re lazy (and what programmer isn’t?) then you can just do this check before all reads – it’s mandatory before blocking to read a request body, but it’s safe at any time.

And the other thing you want to pay attention to is the special values that next_event() might return: NEED_DATA and PAUSED.

NEED_DATA is what it sounds like: it means that next_event() is guaranteed not to return any more real events until you’ve called receive_data() at least once.

PAUSED is a little more subtle: it means that next_event() is guaranteed not to return any more real events until something else has happened to clear up the paused state. There are three cases where this can happen:

  1. We received a full request/response from the remote peer, and then we received some more data after that. (The main situation where this might happen is a server responding to a pipelining client.) The PAUSED state will go away after you call start_next_cycle().
  2. A successful CONNECT or Upgrade: request has caused the connection to switch to some other protocol (see Switching protocols). This PAUSED state is permanent; you should abandon this Connection and go do whatever it is you’re going to do with your new protocol.
  3. We’re a server, and the client we’re talking to proposed to switch protocols (see Switching protocols), and now is waiting to find out whether their request was successful or not. Once we either accept or deny their request then this will turn into one of the above two states, so you probably don’t need to worry about handling it specially.

Putting all this together –

If your I/O is organized around a “pull” strategy, where your code requests events as its ready to handle them (e.g. classic synchronous code, or asyncio’s await loop.sock_recv(...), or Curio), then you’ll probably want logic that looks something like:

# Replace do_sendall and do_recv with your I/O code
def get_next_event():
    while True:
        event = conn.next_event()
        if event is h11.NEED_DATA:
            if conn.they_are_waiting_for_100_continue:
                do_sendall(conn, h11.InformationalResponse(100, ...))
            conn.receive_data(do_recv())
            continue
        return event

And then your code that calls this will need to make sure to call it only at appropriate times (e.g., not immediately after receiving EndOfMessage or PAUSED).

If your I/O is organized around a “push” strategy, where the network drives processing (e.g. you’re using Twisted, or implementing an asyncio.Protocol), then you’ll want to internally apply back-pressure whenever you see PAUSED, remove back-pressure when you call start_next_cycle(), and otherwise just deliver events as they arrive. Something like:

class HTTPProtocol(asyncio.Protocol):
    # Save the transport for later -- needed to access the
    # backpressure API.
    def connection_made(self, transport):
        self._transport = transport

    # Internal helper function -- deliver all pending events
    def _deliver_events(self):
        while True:
            event = self.conn.next_event()
            if event is h11.NEED_DATA:
                break
            elif event is h11.PAUSED:
                # Apply back-pressure
                self._transport.pause_reading()
                break
            else:
                self.event_received(event)

    # Called by "someone" whenever new data appears on our socket
    def data_received(self, data):
        self.conn.receive_data(data)
        self._deliver_events()

    # Called by "someone" whenever the peer closes their socket
    def eof_received(self):
        self.conn.receive_data(b"")
        self._deliver_events()
        # asyncio will close our socket unless we return True here.
        return True

    # Called by your code when its ready to start a new
    # request/response cycle
    def start_next_cycle(self):
        self.conn.start_next_cycle()
        # New events might have been buffered internally, and only
        # become deliverable after calling start_next_cycle
        self._deliver_events()
        # Remove back-pressure
        self._transport.resume_reading()

    # Fill in your code here
    def event_received(self, event):
        ...

And your code that uses this will have to remember to check for they_are_waiting_for_100_continue at the appropriate time.

Closing connections

h11 represents a connection shutdown with the special event type ConnectionClosed. You can send this event, in which case send() will simply update the state machine and then return None. You can receive this event, if you call conn.receive_data(b""). (The actual receipt might be delayed if the connection is paused.) It’s safe and legal to call conn.receive_data(b"") multiple times, and once you’ve done this once, then all future calls to receive_data() will also return ConnectionClosed():

In [22]: conn = h11.Connection(our_role=h11.CLIENT)

In [23]: conn.receive_data(b"")

In [24]: conn.receive_data(b"")

In [25]: conn.receive_data(None)

(Or if you try to actually pass new data in after calling conn.receive_data(b""), that will raise an exception.)

h11 is careful about interpreting connection closure in a half-duplex fashion. TCP sockets pretend to be a two-way connection, but really they’re two one-way connections. In particular, it’s possible for one party to shut down their sending connection – which causes the other side to be notified that the connection has closed via the usual socket.recv(...) -> b"" mechanism – while still being able to read from their receiving connection. (On Unix, this is generally accomplished via the shutdown(2) system call.) So, for example, a client could send a request, and then close their socket for writing to indicate that they won’t be sending any more requests, and then read the response. It’s this kind of closure that is indicated by h11’s ConnectionClosed: it means that this party will not be sending any more data – nothing more, nothing less. You can see this reflected in the state machine, in which one party transitioning to CLOSED doesn’t immediately halt the connection, but merely prevents it from continuing for another request/response cycle.

The state machine also indicates that ConnectionClosed events can only happen in certain states. This isn’t true, of course – any party can close their connection at any time, and h11 can’t stop them. But what h11 can do is distinguish between clean and unclean closes. For example, if both sides complete a request/response cycle and then close the connection, that’s a clean closure and everyone will transition to the CLOSED state in an orderly fashion. On the other hand, if one party suddenly closes the connection while they’re in the middle of sending a chunked response body, or when they promised a Content-Length: of 1000 bytes but have only sent 500, then h11 knows that this is a violation of the HTTP protocol, and will raise a ProtocolError. Basically h11 treats an unexpected close the same way it would treat unexpected, uninterpretable data arriving – it lets you know that something has gone wrong.

As a client, the proper way to perform a single request and then close the connection is:

  1. Send a Request with Connection: close
  2. Send the rest of the request body
  3. Read the server’s Response and body
  4. conn.our_state is h11.MUST_CLOSE will now be true. Call conn.send(ConnectionClosed()) and then close the socket. Or really you could just close the socket – the thing calling send will do is raise an error if you’re not in MUST_CLOSE as expected. So it’s between you and your conscience and your code reviewers.

(Technically it would also be legal to shutdown your socket for writing as step 2.5, but this doesn’t serve any purpose and some buggy servers might get annoyed, so it’s not recommended.)

As a server, the proper way to perform a response is:

  1. Send your Response and body
  2. Check if conn.our_state is h11.MUST_CLOSE. This might happen for a variety of reasons; for example, if the response had unknown length and the client speaks only HTTP/1.0, then the client will not consider the connection complete until we issue a close.

You should be particularly careful to take into consideration the following note fromx RFC 7230 section 6.6:

If a server performs an immediate close of a TCP connection, there is a significant risk that the client will not be able to read the last HTTP response. If the server receives additional data from the client on a fully closed connection, such as another request that was sent by the client before receiving the server’s response, the server’s TCP stack will send a reset packet to the client; unfortunately, the reset packet might erase the client’s unacknowledged input buffers before they can be read and interpreted by the client’s HTTP parser.

To avoid the TCP reset problem, servers typically close a connection in stages. First, the server performs a half-close by closing only the write side of the read/write connection. The server then continues to read from the connection until it receives a corresponding close by the client, or until the server is reasonably certain that its own TCP stack has received the client’s acknowledgement of the packet(s) containing the server’s last response. Finally, the server fully closes the connection.

Switching protocols

h11 supports two kinds of “protocol switches”: requests with method CONNECT, and the newer Upgrade: header, most commonly used for negotiating WebSocket connections. Both follow the same pattern: the client proposes that they switch from regular HTTP to some other kind of interaction, and then the server either rejects the suggestion – in which case we return to regular HTTP rules – or else accepts it. (For CONNECT, acceptance means a response with 2xx status code; for Upgrade:, acceptance means an InformationalResponse with status 101 Switching Protocols) If the proposal is accepted, then both sides switch to doing something else with their socket, and h11’s job is done.

As a developer using h11, it’s your responsibility to send and interpret the actual CONNECT or Upgrade: request and response, and to figure out what to do after the handover; it’s h11’s job to understand what’s going on, and help you make the handover smoothly.

Specifically, what h11 does is pause parsing incoming data at the boundary between the two protocols, and then you can retrieve any unprocessed data from the Connection.trailing_data attribute.

Support for sendfile()

Many networking APIs provide some efficient way to send particular data, e.g. asking the operating system to stream files directly off of the disk and into a socket without passing through userspace.

It’s possible to use these APIs together with h11. The basic strategy is:

  • Create some placeholder object representing the special data, that your networking code knows how to “send” by invoking whatever the appropriate underlying APIs are.
  • Make sure your placeholder object implements a __len__ method returning its size in bytes.
  • Call conn.send_with_data_passthrough(Data(data=<your placeholder object>))
  • This returns a list whose contents are a mixture of (a) bytes-like objects, and (b) your placeholder object. You should send them to the network in order.

Here’s a sketch of what this might look like:

class FilePlaceholder:
    def __init__(self, file, offset, count):
        self.file = file
        self.offset = offset
        self.count = count

    def __len__(self):
        return self.count

def send_data(sock, data):
    if isinstance(data, FilePlaceholder):
        # socket.sendfile added in Python 3.5
        sock.sendfile(data.file, data.offset, data.count)
    else:
        sock.sendfile(data)

placeholder = FilePlaceholder(open("...", "rb"), 0, 200)
for data in conn.send_with_data_passthrough(Data(data=placeholder)):
    send_data(sock, data)

This works with all the different framing modes (Content-Length, Transfer-Encoding: chunked, etc.) – h11 will add any necessary framing data, update its internal state, and away you go.

Identifying h11 in requests and responses

According to RFC 7231, client requests are supposed to include a User-Agent: header identifying what software they’re using, and servers are supposed to respond with a Server: header doing the same. h11 doesn’t construct these headers for you, but to make it easier for you to construct this header, it provides:

h11.PRODUCT_ID

A string suitable for identifying the current version of h11 in a User-Agent: or Server: header.

The version of h11 that was used to build these docs identified itself as:

In [26]: h11.PRODUCT_ID
Out[26]: 'python-h11/0.7.0+dev'

Chunked Transfer Encoding Delimiters

New in version 0.7.0.

HTTP/1.1 allows for the use of Chunked Transfer Encoding to frame request and response bodies. This form of transfer encoding allows the implementation to provide its body data in the form of length-prefixed “chunks” of data.

RFC 7230 is extremely clear that the breaking points between chunks of data are non-semantic: that is, users should not rely on them or assign any meaning to them. This is particularly important given that RFC 7230 also allows intermediaries such as proxies and caches to change the chunk boundaries as they see fit, or even to remove the chunked transfer encoding entirely.

However, for some applications it is valuable or essential to see the chunk boundaries because the peer implementation has assigned meaning to them. While this is against the specification, if you do really need access to this information h11 makes it available to you in the form of the Data.chunk_start and Data.chunk_end properties of the Data event.

Data.chunk_start is set to True for the first Data event for a given chunk of data. Data.chunk_end is set to True for the last Data event that is emitted for a given chunk of data. h11 guarantees that it will always emit at least one Data event for each chunk of data received from the remote peer, but due to its internal buffering logic it may return more than one. It is possible for a single Data event to have both Data.chunk_start and Data.chunk_end set to True, in which case it will be the only Data event for that chunk of data.

Again, it is strongly encouraged that you avoid relying on this information if at all possible. This functionality should be considered an escape hatch for when there is no alternative but to rely on the information, rather than a general source of data that is worth relying on.