Permalink
Fetching contributors…
Cannot retrieve contributors at this time
214 lines (165 sloc) 9.61 KB

Advanced Usage

This section of the documentation covers more advanced use-cases for hyper.

Responses as Context Managers

If you're concerned about having too many TCP sockets open at any one time, you may want to keep your connections alive only as long as you know you'll need them. In HTTP/2 this is generally not something you should do unless you're very confident you won't need the connection again anytime soon. However, if you decide you want to avoid keeping the connection open, you can use the :class:`HTTPConnection <hyper.HTTPConnection>` as a context manager:

with HTTPConnection('http2bin.org') as conn:
    conn.request('GET', '/get')
    data = conn.get_response().read()

analyse(data)

You may not use any :class:`HTTP20Response <hyper.HTTP20Response>` or :class:`HTTP11Response <hyper.HTTP11Response>` objects obtained from a connection after that connection is closed. Interacting with these objects when a connection has been closed is considered undefined behaviour.

Chunked Responses

Plenty of APIs return chunked data, and it's often useful to iterate directly over the chunked data. hyper lets you iterate over each data frame of a HTTP/2 response, and each chunk of a HTTP/1.1 response delivered with Transfer-Encoding: chunked:

for chunk in response.read_chunked():
    do_something_with_chunk(chunk)

There are some important caveats with this iteration: mostly, it's not guaranteed that each chunk will be non-empty. In HTTP/2, it's entirely legal to send zero-length data frames, and this API will pass those through unchanged. Additionally, by default this method will decompress a response that has a compressed Content-Encoding: if you do that, each element of the iterator will no longer be a single chunk, but will instead be whatever the decompressor returns for that chunk.

If that's problematic, you can set the decode_content parameter to False and, if necessary, handle the decompression yourself:

for compressed_chunk in response.read_chunked(decode_content=False):
    decompress(compressed_chunk)

Very easy!

Multithreading

Currently, hyper's :class:`HTTPConnection <hyper.HTTPConnection>` class is not thread-safe. Thread-safety is planned for hyper's core objects, but in this early alpha it is not a high priority.

To use hyper in a multithreaded context the recommended thing to do is to place each connection in its own thread. Each thread should then have a request queue and a response queue, and the thread should be able to spin over both, sending requests and returning responses. The stream identifiers provided by hyper can be used to match the two together.

SSL/TLS Certificate Verification

By default, all HTTP/2 connections are made over TLS, and hyper bundles certificate authorities that it uses to verify the offered TLS certificates. Currently certificate verification cannot be disabled.

Streaming Uploads

Just like the ever-popular requests module, hyper allows you to perform a 'streaming' upload by providing a file-like object to the 'data' parameter. This will cause hyper to read the data in 1kB at a time and send it to the remote server. You must set an accurate Content-Length header when you do this, as hyper won't set it for you.

Content Decompression

In HTTP/2 it's mandatory that user-agents support receiving responses that have their bodies compressed. As demonstrated in the quickstart guide, hyper transparently implements this decompression, meaning that responses are automatically decompressed for you. If you don't want this to happen, you can turn it off by passing the decode_content parameter to :meth:`read() <hyper.HTTP20Response.read>`, like this:

>>> resp.read(decode_content=False)
b'\xc9...'

Flow Control & Window Managers

HTTP/2 provides a facility for performing 'flow control', enabling both ends of a HTTP/2 connection to influence the rate at which data is received. When used correctly flow control can be a powerful tool for maximising the efficiency of a connection. However, when used poorly, flow control leads to severe inefficiency and can adversely affect the throughput of the connection.

By default hyper does its best to manage the flow control window for you, trying to avoid severe inefficiencies. In general, though, the user has a much better idea of how to manage the flow control window than hyper will: you know your use case better than hyper possibly can.

For that reason, hyper provides a facility for using pluggable window managers. A window manager is an object that is in control of resizing the flow control window. This object gets informed about every frame received on the connection, and can make decisions about when to increase the size of the receive window. This object can take advantage of knowledge from layers above hyper, in the user's code, as well as knowledge from hyper's layer.

To implement one of these objects, you will want to subclass the :class:`BaseFlowControlManager <hyper.http20.window.BaseFlowControlManager>` class and implement the :meth:`increase_window_size() <hyper.http20.window.BaseFlowControlManager.increase_window_size>` method. As a simple example, we can implement a very stupid flow control manager that always resizes the window in response to incoming data like this:

class StupidFlowControlManager(BaseFlowControlManager):
    def increase_window_size(self, frame_size):
        return frame_size

The class can then be plugged straight into a connection object:

HTTP20Connection('http2bin.org', window_manager=StupidFlowControlManager)

Note that we don't plug an instance of the class in, we plug the class itself in. We do this because the connection object will spawn instances of the class in order to manage the flow control windows of streams in addition to managing the window of the connection itself.

Server Push

HTTP/2 provides servers with the ability to "push" additional resources to clients in response to a request, as if the client had requested the resources themselves. When minimizing the number of round trips is more critical than maximizing bandwidth usage, this can be a significant performance improvement.

Servers may declare their intention to push a given resource by sending the headers and other metadata of a request that would return that resource - this is referred to as a "push promise". They may do this before sending the response headers for the original request, after, or in the middle of sending the response body.

In order to receive pushed resources, the :class:`HTTPConnection <hyper.HTTPConnection>` object must be constructed with enable_push=True.

You may retrieve the push promises that the server has sent so far by calling :meth:`get_pushes() <hyper.HTTP20Connection.get_pushes>`, which returns a generator that yields :class:`HTTP20Push <hyper.HTTP20Push>` objects. Note that this method is not idempotent; promises returned in one call will not be returned in subsequent calls. If capture_all=False is passed (the default), the generator will yield all buffered push promises without blocking. However, if capture_all=True is passed, the generator will first yield all buffered push promises, then yield additional ones as they arrive, and terminate when the original stream closes. Using this parameter is only recommended when it is known that all pushed streams, or a specific one, are of higher priority than the original response, or when also processing the original response in a separate thread (N.B. do not do this; hyper is not yet thread-safe):

conn.request('GET', '/')
response = conn.get_response()
for push in conn.get_pushes(): # all pushes promised before response headers
    print(push.path)
conn.read()
for push in conn.get_pushes(): # all other pushes
    print(push.path)

To cancel an in-progress pushed stream (for example, if the user already has the given path in cache), call :meth:`HTTP20Push.cancel() <hyper.HTTP20Push.cancel>`.

hyper does not currently verify that pushed resources comply with the Same-Origin Policy, so users must take care that they do not treat pushed resources as authoritative without performing this check themselves (since the server push mechanism is only an optimization, and clients are free to issue requests for any pushed resources manually, there is little downside to simply ignoring suspicious ones).

Nghttp2

By default hyper uses its built-in pure-Python HPACK encoder and decoder. These are reasonably efficient, and suitable for most use cases. However, they do not produce the best compression ratio possible, and because they're written in pure-Python they incur a cost in memory usage above what is strictly necessary.

nghttp2 is a HTTP/2 library written in C that includes a HPACK encoder and decoder. nghttp2's encoder produces extremely compressed output, and because it is written in C it is also fast and memory efficient. For this reason, performance conscious users may prefer to use nghttp2's HPACK implementation instead of hyper's.

You can do this very easily. If nghttp2's Python bindings are installed, hyper will transparently switch to using nghttp2's HPACK implementation instead of its own. No configuration is required.

Instructions for installing nghttp2 are available here.