A pure-Python implementation of the HTTP/2 priority tree
Priority is a pure-Python implementation of the priority logic for HTTP/2, set
out in RFC 7540 Section 5.3 (Stream Priority)
_. This logic allows for clients
to express a preference for how the server allocates its (limited) resources to
the many outstanding HTTP requests that may be running over a single HTTP/2
connection.
Specifically, this Python implementation uses a variant of the implementation
used in the excellent H2O
_ project. This original implementation is also the
inspiration for nghttp2's
_ priority implementation, and generally produces a
very clean and even priority stream. The only notable changes from H2O's
implementation are small modifications to allow the priority implementation to
work cleanly as a separate implementation, rather than being embedded in a
HTTP/2 stack directly.
While priority information in HTTP/2 is only a suggestion, rather than an enforceable constraint, where possible servers should respect the priority requests of their clients.
Priority has a simple API. Streams are inserted into the tree: when they are inserted, they may optionally have a weight, depend on another stream, or become an exclusive dependent of another stream.
.. code-block:: python
>>> p = priority.PriorityTree()
>>> p.insert_stream(stream_id=1)
>>> p.insert_stream(stream_id=3)
>>> p.insert_stream(stream_id=5, depends_on=1)
>>> p.insert_stream(stream_id=7, weight=32)
>>> p.insert_stream(stream_id=9, depends_on=7, weight=8)
>>> p.insert_stream(stream_id=11, depends_on=7, exclusive=True)
Once streams are inserted, the stream priorities can be requested. This allows the server to make decisions about how to allocate resources.
Iterating The Tree ~~~~~~~~~~~~~~~~~~
The tree in this algorithm acts as a gate. Its goal is to allow one stream "through" at a time, in such a manner that all the active streams are served as evenly as possible in proportion to their weights.
This is handled in Priority by iterating over the tree. The tree itself is an iterator, and each time it is advanced it will yield a stream ID. This is the ID of the stream that should next send data.
This looks like this:
.. code-block:: python
>>> for stream_id in p:
... send_data(stream_id)
If each stream only sends when it is 'ungated' by this mechanism, the server will automatically be emitting stream data in conformance to RFC 7540.
Updating The Tree ~~~~~~~~~~~~~~~~~
If for any reason a stream is unable to proceed (for example, it is blocked on
HTTP/2 flow control, or it is waiting for more data from another service), that
stream is blocked. The PriorityTree
should be informed that the stream is
blocked so that other dependent streams get a chance to proceed. This can be
done by calling the block
method of the tree with the stream ID that is
currently unable to proceed. This will automatically update the tree, and it
will adjust on the fly to correctly allow any streams that were dependent on
the blocked one to progress.
For example:
.. code-block:: python
>>> for stream_id in p:
... send_data(stream_id)
... if blocked(stream_id):
... p.block(stream_id)
When a stream goes from being blocked to being unblocked, call the unblock
method to place it back into the sequence. Both the block
and unblock
methods are idempotent and safe to call repeatedly.
Additionally, the priority of a stream may change. When it does, the
reprioritize
method can be used to update the tree in the wake of that
change. reprioritize
has the same signature as insert_stream
, but
applies only to streams already in the tree.
Removing Streams ~~~~~~~~~~~~~~~~
A stream can be entirely removed from the tree by calling remove_stream
.
Note that this is not idempotent. Further, calling remove_stream
and then
re-adding it may cause a substantial change in the shape of the priority
tree, and will cause the iteration order to change.
Priority is made available under the MIT License. For more details, see the LICENSE file in the repository.
Priority is maintained by Cory Benfield, with contributions from others. For more details about the contributors, please see CONTRIBUTORS.rst in the repository.
.. _RFC 7540 Section 5.3 (Stream Priority): https://tools.ietf.org/html/rfc7540#section-5.3 .. _nghttp2's: https://nghttp2.org/blog/2015/11/11/stream-scheduling-utilizing-http2-priority/ .. _H2O: https://h2o.examp1e.net/
API Changes
PriorityLoop
when inserting or reprioritising a stream that
depends on itself.BadWeightError
when creating or reprioritising a stream with a
weight that is not an integer between 1 and 256, inclusive.PseudoStreamError
when trying to reprioritise, remove, block or
unblock stream 0.PriorityError
parent class for the exceptions that can be
thrown by priority.Bugfixes
insert_stream
to be called with exclusive=True
but no explicit
depends_on
value.Bugfixes
Security Fixes
CVE-2016-6580: All versions of this library prior to 1.2.0 are vulnerable to a denial of service attack whereby a remote peer can cause a user to insert an unbounded number of streams into the priority tree, eventually consuming all available memory.
This version adds a TooManyStreamsError
exception that is raised when
too many streams are inserted into the priority tree. It also adds a keyword
argument to the priority tree, maximum_streams
, which limits how many
streams may be inserted. By default, this number is set to 1000.
Implementations should strongly consider whether they can set this value
lower.
Bugfixes
queue.PriorityQueue
to
heapq
.API Changes
DuplicateStreamError
when inserting a stream that is already in the
tree.MissingStreamError
when reprioritising a stream that is not in the
tree.