Path.glob vs listdir+fnmatch: Localhost Lab

Intro — what this post promises

Match files in one directory: Path.glob('*.py') vs os.listdir + fnmatch.filter. This lab reports matches/s on Linux localhost.

It is not glob-vs-rglob-vs-walk (lab 76) and not fnmatch-vs-re pattern microbenches (lab 93). Focus: Path.glob vs listdir+fnmatch on a single directory.

Related links:

• glob vs rglob vs walk localhost lab
• fnmatch vs re localhost lab
• tomllib vs json localhost lab
• stat vs path stat localhost lab
• heapq merge vs sorted localhost lab
• islice vs list slice localhost lab
• dataclass replace vs manual localhost lab
• mmap write vs write localhost lab

Lab honesty (1 Oct 2026 IST): Python 3.13.5. Affiliates: 0. Flat dirs only.

Verdict up front (medium 2000 entries / 400 *.py): listdir+endswith ~721109 matches/s; listdir+fnmatch ~526570; Path.glob ~294091; Path.iterdir+suffix ~61488.

Arms

Lab topology

small 200/40 · medium 2000/400 · large 8000/1600 · 7 rounds · p50
metric: matches/s = n_match / p50_s

Script: lab-evidence/113-path-glob-vs-fnmatch/results/run_lab.py.

Lead table — medium (p50 matches/s)

Scale sketch (Path.glob vs listdir+fnmatch)

Glob pattern cost

*.py is a cheap filter. Patterns with multiple * / ? classes push more work into fnmatch — still usually dominated by directory listing syscall cost on large folders. Re-measure if your matcher is ** recursive (again: lab 76).

Reading it

• Simple *.ext on one dir → listdir/scandir + endswith leads here.
• Need glob wildcards beyond a suffix → fnmatch or Path.glob.
• Path.glob stays convenient and mid-pack; fine for tooling scripts.
• Recursive trees → lab 76 (rglob / walk), not this post.

When Path.glob is still right

Tooling scripts that already speak pathlib, need brace-free relative patterns, or may later grow a single * directory segment often stay on Path.glob for clarity. The ~2× gap vs listdir+fnmatch on medium here rarely matters unless you scan directories in a tight loop.

endswith vs fnmatch

endswith(".py") is not a glob. It misses patterns like test_*.py or *.pyc exclusions. Use it only for fixed-suffix filters; otherwise fnmatch.filter / Path.glob keep semantics honest. Lab 93 covers pattern-engine speed in isolation — not directory IO.

Object tax

Path.iterdir() yielding Path objects paid a clear tax versus name-only listdir/scandir arms. If you only need filenames for a filter, keep strings until you must touch the filesystem again.

Pitfalls

• Using Path.iterdir lists when you only need names (Path object tax showed up here).
• Running rglob accidentally when one directory was intended.
• Treating fnmatch pattern speed (lab 93) as directory listing speed.
• Ignoring locale/case rules on case-insensitive filesystems.

Reproduce

python3 lab-evidence/113-path-glob-vs-fnmatch/results/run_lab.py

Evidence: summary.json, summary.txt, tree/.

For CI file collectors, scandir+endswith on a fixed suffix is a good default; graduate to fnmatch/Path.glob when patterns get expressive.

Always verify match counts against a known fixture before trusting a faster filter in production sweeps.

Limits

One Linux box. Flat synthetic directories. Pattern fixed to *.py. Not networked FS.

Takeaway

At 2000 entries, listdir+fnmatch ~526570 matches/s beat Path.glob ~294091; plain endswith ~721109 led for suffix-only filters. Use Path.glob for ergonomics; listdir+fnmatch/endswith when directory scans are hot.