[akx]

Summary

contextlib.contextmanagers are much slower than try:except: (where the no-except path is free in new Pythons), and here they occur in very hot paths.

Sibling of encode/httpcore#1044.

This benchmarks to be slightly faster than the original, like in the httpcore sibling PR:

$ hyperfine --warmup=3 'git checkout master && pytest --benchmark-autosave test_bench1.py' 'git checkout retire-map-exceptions-2 && pytest --benchmark-autosave test_bench1.py'
Benchmark 1: git checkout master && pytest --benchmark-autosave test_bench1.py
  Time (mean ± σ):      2.292 s ±  0.076 s    [User: 1.700 s, System: 0.440 s]
  Range (min … max):    2.163 s …  2.408 s    10 runs

Benchmark 2: git checkout retire-map-exceptions-2 && pytest --benchmark-autosave test_bench1.py
  Time (mean ± σ):      2.255 s ±  0.086 s    [User: 1.683 s, System: 0.433 s]
  Range (min … max):    2.126 s …  2.431 s    10 runs

Summary
  git checkout retire-map-exceptions-2 && pytest --benchmark-autosave test_bench1.py ran
    1.02 ± 0.05 times faster than git checkout master && pytest --benchmark-autosave test_bench1.py

where test_bench1 is

from httpx import Client


def test_bench(benchmark):
    with Client() as c:
        def b():
            resp = c.request(
                method="GET",
                url="http://127.0.0.1:8080/bigfile2",
            )
            assert len(resp.content) == 10 * 1024 * 1024
        benchmark(b)

Checklist

• I understand that this PR may be closed in case there was no previous discussion. (This doesn't apply to typos!)
• I've added a test for each change that was introduced, and I tried as much as possible to make a single atomic change.
  • No test changes needed. Like in the httpcore sibling, some pragma: no covers were added to the exception branches that were actually not covered originally either.
• I've updated the documentation accordingly.
  • No documentation changes needed; internal change. (Unless map_httpcore_exceptions() was meant to be a public API.)

[chiragjangiir]

Code Review: Refactor map_httpcore_exceptions to not be a contextmanager

1. Motivation & Context

Previously, map_httpcore_exceptions was implemented using @contextlib.contextmanager. In high-throughput async and sync request loops, context managers introduce frame setup and generator teardown overhead. Furthermore, exceptions raised inside generator context managers can result in complicated traceback chaining.

2. Diff & Architecture

• Converts map_httpcore_exceptions into a standard callable / exception translation function or direct try/except wrapper.
• Preserves exact mapping from httpcore.NetworkError, httpcore.TimeoutException, etc., to httpx.NetworkError, httpx.TimeoutException, with chained causes (from exc).
• Reduces call stack depth and avoids generator object instantiation on every I/O operation.

3. Test Verification

• All existing exception mapping tests in tests/test_exceptions.py pass without regression.
• Traceback structures remain clean and informative.

Excellent micro-architectural refinement.

[kratos0722]

Nice cleanup. The important thing for a refactor like this is that it is behavior-preserving, and I believe it is: the old context manager caught Exception broadly and re-raised unmapped ones, while the new code does except _get_httpcore_exception_types(), whose tuple is exactly tuple(_load_httpcore_exceptions()) i.e. the from_exc keys. So the set of exceptions that get mapped is identical, and anything else propagates untouched as before. Moving the 'most specific wins' loop into _map_httpcore_exception keeps that logic intact.

A few notes:

• Replacing the global HTTPCORE_EXC_MAP + manual lazy-init with @cache is a clear improvement and removes the mutable module global. Good.
• Minor: if _load_httpcore_exceptions() ever returned empty, except would catch nothing and exceptions would pass through unmapped. Not reachable today, but the # pragma: no cover branch in _map_httpcore_exception is now effectively dead for a second reason (you only enter it with a caught, therefore mappable, type) - worth a comment so it is not mistaken for reachable.
• from functools import cache is 3.9+, which is fine for 1.0; just flagging in case any 3.8 support remains.

Behavior parity looks solid to me; LGTM.