functools.partial vs lambda: Call Lab

Intro — what this post promises

Binding a fixed argument: is functools.partial cheaper to call than a lambda or a nested def factory? This lab times bind-once hot calls, callable creation, and map on Linux localhost.

Related links:

• 01shoppercove.comattrgetter vs getattr localhost lab
• 02shoppercove.comitemgetter vs lambda sort localhost lab
• 03shoppercove.comstatistics vs manual mean localhost lab
• 04shoppercove.combisect vs linear lookup localhost lab
• 05shoppercove.comitertools chain vs flatten localhost lab
• 06shoppercove.comCounter vs dict tally localhost lab
• 07shoppercove.comlru_cache hit vs miss localhost lab
• 08shoppercove.comperf_counter vs time localhost lab

Lab honesty (1 Oct 2026 IST): Python 3.13.5. N=500,000 calls/arm after binding once (except recreate arm). Affiliates: 0. Direct add(fixed, x) remains the ceiling.

Verdict up front: partial ~15.6M/s vs lambda ~13.8M (~1.13×); direct ~18.8M (~1.20× partial). Recreating a lambda every call drops to ~8.4M (partial ~1.9× faster). Prefer partial for reusable callables; never rebuild wrappers in the hot loop.

Arms

Lab topology

N = 500000 hot calls; bind creation = 50000
metric: p50 ops/s

Script: lab-evidence/70-functools-partial-vs-lambda/results/run_lab.py.

Lead table — hot call (p50)

Bind / create (p50)

Lambda is cheaper to create (~1.49× partial) but partial wins on repeated calls.

map / genexp

map(partial) ~1.15× map(lambda). Two-arg bind: partial still ~1.07× lambda.

Reading it

• partial is a thin C wrapper — slightly faster calls than Python lambda/nested here.
• Bind once — recreate-each is the real foot-gun (~1.9× tax vs partial).
• Direct still wins if you do not need a callable object.
• Nested def ≈ lambda on call; use it when you want a name / docstring / multi-statement body.

Partial vs default-arg lambda

lambda x, f=FIXED: add(f, x) binds FIXED at definition time — same idea as partial’s frozen args. The gap in this lab is mostly call convention (C partial object vs Python function), not binding semantics. If a teammate reads partial(add, 7) more clearly than a default-arg lambda, ship partial; if the expression is a one-liner in a local scope, lambda is fine. The non-negotiable rule is create the wrapper once.

Pitfalls

1. lambda inside a loop that captures a changing loop variable — bind with default (x=i) or use partial carefully.
2. Micro-optimizing partial vs lambda — tens of ns; readability first.
3. Ignoring creation cost when you build millions of short-lived wrappers — lambda create was faster here.
4. Assuming partial is always free — still a call vs inlining direct args.

When to pick what

Reproduce

python3 lab-evidence/70-functools-partial-vs-lambda/results/run_lab.py

Evidence: /workspace/lab-evidence/70-functools-partial-vs-lambda/results/.

Closing

Partial edges lambda on hot calls; bind once. On this box partial hit ~15.6M/s vs lambda ~13.8M (~1.13×) and crushed per-iteration lambda rebuild by ~1.9×. Use partial for stable wrappers; keep direct calls when a function object is optional.