fnmatch vs re Name Filter: Localhost Lab

Intro — what this post promises

Filter a large list of basenames with shell globs (fnmatch.filter / fnmatchcase) vs re.compile. This lab reports names/s on ~52700 synthetic names on Linux localhost.

It is not a filesystem walk (glob vs rglob vs walk) and not general string search (re vs str). Patterns run against in-memory names only.

Related links:

• 01shoppercove.comglob vs rglob vs walk localhost lab
• 02shoppercove.compython re vs str localhost lab
• 03shoppercove.comscandir vs listdir localhost lab
• 04shoppercove.comfutures as completed vs wait localhost lab
• 05shoppercove.comtarfile vs zipfile localhost lab
• 06shoppercove.comstr translate vs replace localhost lab
• 07shoppercove.comtextwrap fill vs manual localhost lab
• 08shoppercove.comsecrets vs urandom localhost lab

Lab honesty (1 Oct 2026 IST): Python 3.13.5. Affiliates: 0. No Docker. Regexes are hand-translated equivalents of the globs (\Z-anchored).

Verdict up front (*.py): fnmatch.filter ~9.03 Mnames/s vs re.match ~8.28 M vs re.search ~3.07 M. Multi-glob any-of: re alternation ~1.95 M/s vs fnmatch loop ~1.65 M/s.

Arms

Lab topology

~52700 basenames · 7 rounds · p50
patterns: *.py, test_*, *.log, app_*.json, pkg.module_*.py
metric: names/s = n / p50_s

Script: lab-evidence/93-fnmatch-vs-re/results/run_lab.py.

Lead table — names/s (p50, millions)

Multi-pattern (any of *.py / *.json / test_*)

Reading it

• fnmatch.filter is the convenience winner for one shell glob over a list — often tied with or ahead of a careful re.match.
• Per-name fnmatchcase loops pay Python call overhead (~2× slower than filter here).
• re.search with leading .* can lose to match / fnmatch — compile thoughtfully.
• Many globs: one compiled alternation beat a Python any-of fnmatch loop (~1.19×).

Convenience vs hot path

Use fnmatch.filter in tools and one-off filters — readable and fast enough. Reach for re.compile when patterns are dynamic, numerous, or already regex; keep the compiled object outside the loop.

translate() note

fnmatch.translate can turn a glob into a regex string for re.compile. This lab used hand-written equivalents for clarity. If you generate regex from globs at runtime, compile once outside the name loop — the compile tax will otherwise dominate.

Batch vs per-name

fnmatch.filter walks the list in C and appends matches — that is why it beat a Python fnmatchcase loop by ~2× on every pattern here. If you already iterate names for other reasons, fnmatchcase inline can still be fine; for “give me all *.py,” call filter.

Match counts sanity

All arms agreed on match counts per pattern (e.g. *.py and app_*.json hits matched across fnmatch and re). Throughput differences are engine cost, not divergent semantics on these fixtures.

Pitfalls

• Translating globs to regex incorrectly (forgetting to escape dots).
• Using re.search(r".*\.py") without \Z / fullmatch semantics.
• Walking the filesystem with fnmatch on every os.walk name without measuring — different problem than this list filter.
• Case folding: fnmatch vs fnmatchcase differ on case-insensitive platforms.

Reproduce

python3 lab-evidence/93-fnmatch-vs-re/results/run_lab.py

Evidence: summary.json, fixture_names.txt sample.

Limits

One Linux box. Synthetic basenames. Not pathlib.match / recursive glob. Regexes hand-written, not fnmatch.translate microbench.

Takeaway

On ~52700 names, fnmatch.filter ~9.03 Mnames/s for *.py matched re.match ~8.28 M. For multi-glob, compiled alternation ~1.95 M/s edged the fnmatch any-loop. Default to fnmatch for shell-shaped filters; compile re when the hot path is many patterns or regex features.