Python 2 desupport and removal

[Tagar]

Python 2 desupport / removal.

py4j has supported Python 3+ for some time (CI tests py3.9 – py3.13) but the codebase still carried the full py2 / py3 dual-support machinery: from __future__ imports in every module, a 130-line py4j.compat module exporting py2/3 shims, # -*- coding: UTF-8 -*- declarations, sys.version_info conditionals, the long(x) / unicode(s) / basestring / CompatThread aliases, packaging metadata (setup.cfg's universal=1, tox.ini's envlist=py27,py34,py35 + nosetests), IDE config (.pydevproject files specifying python 2.6), and a couple of docs references. None of that does anything on Python 3.

This change is strictly mechanical — function bodies are unchanged, no perf optimizations included. Each py2 / py3 alias is replaced inline with its py3 equivalent.

Scope

Area	Change
Source modules	Dropped from __future__, # -*- coding ..., dead from py4j.compat import (...) lines. Replaced basestring → str, long → int, unicode → str, bytestr → bytes, bytearray2 → bytes, CompatThread → threading.Thread, items(d) → list(d.items()), iteritems(self) → self.items(), isbytestr / ispython3bytestr / isbytearray → direct isinstance(...), from py4j.compat import Queue → from queue import Queue.
java_collections.py	Collapsed try/except ImportError for collections.abc (always present on py3.3+). Dropped sys.version_info.major < 3 ternary on __EMPTY_SET / __SET_TEMPLATE.
protocol.py	Slimmed smart_decode body to py3 types (isinstance(s, str) instead of isinstance(s, unicode), etc.). encode_bytearray / decode_bytearray / OUTPUT_CONVERTER LONG_TYPE lambda use native bytes / int. get_command_part uses isinstance(..., str/int/bytes/bytearray) directly. All smart_decode(...) call sites in source are preserved — the function itself still exists and behaves identically.
py4j.compat	Slimmed: removed the if version_info.major < 3: branch entirely; kept the py3 form as the only definition. Module-level docstring documents the migration path. Every export still resolves (unicode, basestring, long, range, bytestr, bytearray2, unichr, iteritems, items, isbytestr, ispython3bytestr, isbytearray, bytetoint, bytetostr, strtobyte, Queue, Empty, hasattr2, CompatThread, version_info, tobytestr) so downstream code that imports from it (from py4j.compat import unicode, etc.) continues to work. The module is a candidate for removal in a future major release.
Tests (17 files + 1 new)	Same header cleanup. long(N) → N; bytearray2 → bytes; isbytearray(a) or ispython3bytestr(a) → isinstance(a, (bytearray, bytes)); unicode(i) → str(i). Dropped two if sys.version_info < (3,): branches in java_gateway_test.py (py2-only sleep paths; the py3 proc.wait(5) branch is the only one that ran on supported Pythons). New protocol_test.py (30 unit tests, no JVM required) pinning the bytes/str boundary and the py4j.compat back-compat surface.
setup.py	Removed "Programming Language :: Python :: 2" and "... :: 3.8" (EOL) classifiers. Added ... :: 3 :: Only and 3.9 – 3.13. Set python_requires=">=3.9" to match CI. Removed distutils fallback + 2to3-era comments.
setup.cfg	Removed — only contained [bdist_wheel] universal=1, the py2/3 universal-wheel marker, which produces an incorrectly-tagged wheel on a py3-only project.
tox.ini	Was envlist=py27,py34,py35 calling nosetests. Now envlist=py{39,310,311,312,313} calling pytest, matching CI. nose.cfg deleted alongside.
.pydevproject (py4j-python and py4j-web)	python 2.6 → python 3.9.
py4j-web/advanced_topics.rst	Dropped the parenthetical (Python 2.x) or bytes (Python 3.x) qualifier on the byte-array return-type description; now reads simply bytes.

Java side

No changes. The Java side speaks pure UTF-8 over the wire and has zero Python-version-specific code (confirmed by grep -rn -i "python 2\|python2\|py2" py4j-java/src/). The clean separation between protocol-level and Python-version-level is one of py4j's good properties — the JVM side is oblivious to which Python is on the other end.

Validation

Local pytest on macOS arm64 + JDK 17: 205 passed, 104 deselected, 0 failed (skipping TLS / perf-framework / memory-leak / multithread suites that need extra setup, matching the main CI workflow's filter).

The CodSpeed workflow will pick this PR up automatically (path filter matches py4j-python/src/py4j/**/*.py) and post per-scenario perf deltas vs master. Expected deltas: near-zero, since this is mechanical alias substitution, not optimization.

Co-authored-by: Isaac

[codspeed]

Merging this PR will improve performance by 20.22%

⚡ 1 improved benchmark
✅ 15 untouched benchmarks
⏩ 7 skipped benchmarks¹

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	WallTime	test_m5b_encode_string	3.2 µs	2.7 µs	+20.22%

Tip

Curious why this is faster? Comment @codspeedbot explain why this is faster on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing py2-desupport (bee1e43) with master (0bde5ca)²}

Footnotes

1. 7 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

2. No successful run was found on master (2ac0f8a) during the generation of this report, so 0bde5ca was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[Tagar]

py4j#585