A pure-Python implementation of the HTTP/2 priority tree
Priority: A HTTP/2 Priority Implementation
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.
>>> 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:
>>> 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 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.
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.
- Throw PriorityLoop when inserting or reprioritising a stream that depends on itself.
- Throw BadWeightError when creating or reprioritising a stream with a weight that is not an integer between 1 and 256, inclusive.
- Throw PseudoStreamError when trying to reprioritise, remove, block or unblock stream 0.
- Add a new PriorityError parent class for the exceptions that can be thrown by priority.
- Allow insert_stream to be called with exclusive=True but no explicit depends_on value.
- Allow insertion of streams that have parents in the idle or closed states. This would previously raise a KeyError.
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.
- 2.5x performance improvement by swapping from queue.PriorityQueue to heapq.
- Throw DuplicateStreamError when inserting a stream that is already in the tree.
- Throw MissingStreamError when reprioritising a stream that is not in the tree.
- Initial release.
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, Size & Hash SHA256 Hash Help||File Type||Python Version||Upload Date|
(11.7 kB) Copy SHA256 Hash SHA256
|Wheel||py2.py3||Jan 27, 2017|
(13.8 kB) Copy SHA256 Hash SHA256
|Source||None||Jan 27, 2017|