WeakValueDictionary vs dict Cache: Localhost Lab

Intro — what this post promises

Can weakref.WeakValueDictionary replace a strong dict as a cache-like map? This lab measures insert and lookup ops/s on Linux localhost, then shows entries vanishing after del + gc.collect() when only weak refs remain.

Related links:

• 01shoppercove.comlru cache hit miss localhost lab
• 02shoppercove.comthreading local vs dict localhost lab
• 03shoppercove.comsetdefault vs defaultdict localhost lab
• 04shoppercove.comdecimal vs float sum localhost lab
• 05shoppercove.compickle vs json roundtrip localhost lab
• 06shoppercove.comasyncio gather vs taskgroup localhost lab
• 07shoppercove.comcounter vs dict tally localhost lab
• 08shoppercove.comcsv reader vs split localhost lab

Lab honesty (1 Oct 2026 IST): Python 3.13.5. Values are small Payload objects with __weakref__. Affiliates: 0. No Docker. Not an LRU eviction policy lab (see lru_cache lab).

Verdict up front (n=10 000): strong insert ~6027.1 k/s vs weak ~960.1 k/s (~6.28×); strong lookup ~16.37 M/s vs weak ~7.64 M/s (~2.14×). After dropping all strong refs to values and gc.collect(), weak map went 10000 → 0 (dropped 10000).

Arms

Lab topology

n in {1000, 10000, 50000} · 7 rounds · p50
insert includes Payload construction
lookup holds strong refs so weak map stays full
GC demo at n=10000 with only-weak ownership

Script: lab-evidence/86-weakref-vs-dict-cache/results/run_lab.py.

Lead table — n=10 000 (p50)

Scale sketch

Insert tax for weak maps stayed roughly ~6× at 10 k; lookup closer to ~2×.

GC drop (the point of weak caches)

Scenario B — only the weak map holds values:

1. Insert 10 000 objects into WeakValueDictionary, keep a list of strong refs.
2. del the list; gc.collect().
3. len(weak) went 10000 → 0.

If a strong dict still owns the same objects, weak entries stay alive — weakrefs do not override other strong owners. That is why “cache” designs often pair weak value maps with an explicit size/TTL policy, not weakrefs alone.

Reading it

• Strong dict wins raw speed for insert/lookup microbenches.
• WeakValueDictionary buys auto-drop when nothing else references the value — pay ~2–6× on this box.
• Use weak value maps for identity caches / registries where leaking objects is worse than a miss; use functools.lru_cache / explicit eviction when you need bounded strong residency.

When weakrefs help

Pitfalls

• Storing objects that cannot be weakly referenced (some builtins/extension types).
• Expecting weak maps to bound memory while another structure still holds strong refs.
• Looking up after GC without handling KeyError / missing keys.
• Comparing to LRU without measuring hit-rate economics.

Reproduce

python3 lab-evidence/86-weakref-vs-dict-cache/results/run_lab.py

Evidence: summary.json, summary.txt.

Limits

One Linux box. Synthetic Payload with slots+__weakref__. Not multi-process. GC timing can vary under load.

Takeaway

Strong dict was about ~6.28× faster to fill and ~2.14× faster to read at n=10 k. WeakValueDictionary dropped all 10000 entries once strong refs vanished — that auto-cleanup is the product, not the Mops/s. Pick weak maps for leak resistance; pick strong maps (or LRU) for speed and explicit residency.