Avoid making a copy of large frames.

This isn't very significant compared to the cost of compression.

It can make a real difference for decompression.

docs/project/changelog.rst

@@ -61,6 +61,20 @@ Backwards-incompatible changes
     Aliases for deprecated API were removed from ``__all__``. As a consequence,
     they cannot be imported e.g. with ``from websockets import *`` anymore.
 
+.. admonition:: :attr:`Frame.data <frames.Frame.data>` is now a bytes-like object.
+    :class: note
+
+    In addition to :class:`bytes`, it may be a :class:`bytearray` or a
+    :class:`memoryview`.
+
+    If you wrote an :class:`extension <extensions.Extension>` that relies on
+    methods not provided by these new types, you may need to update your code.
+
+Improvements
+............
+
+* Sending or receiving large compressed frames is now faster.
+
 .. _13.1:
 
 13.1

src/websockets/extensions/permessage_deflate.py

@@ -129,16 +129,22 @@ def decode(
         # Uncompress data. Protect against zip bombs by preventing zlib from
         # decompressing more than max_length bytes (except when the limit is
         # disabled with max_size = None).
-        data = frame.data
-        if frame.fin:
-            data += _EMPTY_UNCOMPRESSED_BLOCK
+        if frame.fin and len(frame.data) < 2044:
+            # Profiling shows that appending four bytes, which makes a copy, is
+            # faster than calling decompress() again when data is less than 2kB.
+            data = bytes(frame.data) + _EMPTY_UNCOMPRESSED_BLOCK
+        else:
+            data = frame.data
         max_length = 0 if max_size is None else max_size
         try:
             data = self.decoder.decompress(data, max_length)
+            if self.decoder.unconsumed_tail:
+                raise PayloadTooBig(f"over size limit (? > {max_size} bytes)")
+            if frame.fin and len(frame.data) >= 2044:
+                # This cannot generate additional data.
+                self.decoder.decompress(_EMPTY_UNCOMPRESSED_BLOCK)
         except zlib.error as exc:
             raise ProtocolError("decompression failed") from exc
-        if self.decoder.unconsumed_tail:
-            raise PayloadTooBig(f"over size limit (? > {max_size} bytes)")
 
         # Allow garbage collection of the decoder if it won't be reused.
         if frame.fin and self.remote_no_context_takeover:
@@ -176,11 +182,15 @@ def encode(self, frame: frames.Frame) -> frames.Frame:
 
         # Compress data.
         data = self.encoder.compress(frame.data) + self.encoder.flush(zlib.Z_SYNC_FLUSH)
-        if frame.fin and data[-4:] == _EMPTY_UNCOMPRESSED_BLOCK:
-            # Making a copy is faster than memoryview(a)[:-4] until about 2kB.
-            # On larger messages, it's slower but profiling shows that it's
-            # marginal compared to compress() and flush(). Keep it simple.
-            data = data[:-4]
+        if frame.fin:
+            # Sync flush generates between 5 or 6 bytes, ending with the bytes
+            # 0x00 0x00 0xff 0xff, which must be removed.
+            assert data[-4:] == _EMPTY_UNCOMPRESSED_BLOCK
+            # Making a copy is faster than memoryview(a)[:-4] until 2kB.
+            if len(data) < 2048:
+                data = data[:-4]
+            else:
+                data = memoryview(data)[:-4]
 
         # Allow garbage collection of the encoder if it won't be reused.
         if frame.fin and self.local_no_context_takeover:

src/websockets/frames.py

@@ -7,7 +7,7 @@
 import secrets
 import struct
 from collections.abc import Generator, Sequence
-from typing import Callable
+from typing import Callable, Union
 
 from .exceptions import PayloadTooBig, ProtocolError
 
@@ -117,7 +117,8 @@ class CloseCode(enum.IntEnum):
 }
 
 
-BytesLike = bytes, bytearray, memoryview
+# Change to bytes | bytearray | memoryview when dropping Python < 3.10.
+BytesLike = Union[bytes, bytearray, memoryview]
 
 
 @dataclasses.dataclass
@@ -139,7 +140,7 @@ class Frame:
     """
 
     opcode: Opcode
-    data: bytes
+    data: BytesLike
     fin: bool = True
     rsv1: bool = False
     rsv2: bool = False
@@ -160,7 +161,7 @@ def __str__(self) -> str:
         if self.opcode is OP_TEXT:
             # Decoding only the beginning and the end is needlessly hard.
             # Decode the entire payload then elide later if necessary.
-            data = repr(self.data.decode())
+            data = repr(bytes(self.data).decode())
         elif self.opcode is OP_BINARY:
             # We'll show at most the first 16 bytes and the last 8 bytes.
             # Encode just what we need, plus two dummy bytes to elide later.
@@ -178,7 +179,7 @@ def __str__(self) -> str:
             # binary. If self.data is a memoryview, it has no decode() method,
             # which raises AttributeError.
             try:
-                data = repr(self.data.decode())
+                data = repr(bytes(self.data).decode())
                 coding = "text"
             except (UnicodeDecodeError, AttributeError):
                 binary = self.data

tests/extensions/test_permessage_deflate.py

@@ -1,4 +1,5 @@
 import dataclasses
+import os
 import unittest
 
 from websockets.exceptions import (
@@ -167,6 +168,16 @@ def test_encode_decode_fragmented_binary_frame(self):
         self.assertEqual(dec_frame1, frame1)
         self.assertEqual(dec_frame2, frame2)
 
+    def test_encode_decode_large_frame(self):
+        # There is a separate code path that avoids copying data
+        # when frames are larger than 2kB. Test it for coverage.
+        frame = Frame(OP_BINARY, os.urandom(4096))
+
+        enc_frame = self.extension.encode(frame)
+        dec_frame = self.extension.decode(enc_frame)
+
+        self.assertEqual(dec_frame, frame)
+
     def test_no_decode_text_frame(self):
         frame = Frame(OP_TEXT, "café".encode())