Standard zset rescore fix — September 7, 2026

The full run under /home/adam/wiki2/wikimedia-zset-bplus-20260905T023755Z/runs/full-1483700913-20260905T024352Z on naamah was stopped at the user's request. Its input, source, binary, samples, and incumbent results are preserved; the run is marked STOPPED-BY-USER.

Standard zset positioning used a score-only binary search followed by a linear walk through equal-score blocks. The allocation-safe rescore path repeated that walk for the old and new positions. The fix binary-searches block maxima by (score, lexicographic member), reserves before transferring an entry, and tracks positions through splits/removal without repeated global searches. Member-score preparation now precedes the transfer, so failure restores the member snapshot without a reverse rescore. Optional source-block merging may be deferred on allocation failure. Redis member ordering is unchanged.

The fixed build and fresh diagnostic results are under /home/adam/wiki2/wikimedia-zset-rescore-fix-20260907T182911Z on naamah.

Focused score-index measurement

../zset_tie_rescore.cpp was compiled with g++ -std=c++23 -O3 -DNDEBUG against the preserved original header and the fixed header. Each run performs 50,000 random member lookups followed by 50,000 score increments. Setup is excluded. This measures the index directly, not end-to-end Redis command or full Wikimedia replay throughput.

All members initially have score 2 in the tied case:

At one million distinct initial scores, rescore time was 0.395309 µs before and 0.391482 µs after. The raw files also include 2,048-score-bucket cases. Each variant/shape/size was measured once, so small timing differences should not be treated as statistically significant.

Every completed measurement independently verifies the final member/score mapping, cardinality, uniqueness, and lexicographic score order. All passed. The original header's layout validator failed on the 100k and 1M tied cases: 64-slot capacity rounding could exceed the default 1,125-slot block limit. That pre-existing issue is recorded as layout_valid=0 in those raw results; the fixed version caps the rounding and passes every layout check.

Raw files: results/verified-{before,after}-{10000,100000,1000000}-{tied,bucketed,unique}.tsv.

Regression coverage

• Equal-score churn, decimal lexicographic ordering, shared prefixes, embedded NULs, negative/fractional/infinite scores, and absent/deleted entries.
• Moves in both directions, source draining, destination splitting, merging, empty-block removal, and all three rank-cache modes.
• Rejected allocations during index growth, and Store rollback with/without cached score text; rounded capacity at the default block size.
• Core and focused tests under AddressSanitizer/UndefinedBehaviorSanitizer, plus a Redis differential check (10,127 commands, seed 90210).

Bounded Wikimedia replay check

The fixed Release core and focused tests also passed on naamah. A fresh replay of the first 1,000,000 commands completed in 16.596 s for Goblin standard and 16.985 s for Redis 8.8, with zero command errors. Both final digests verified 165,385 members, score sum 1,000,000, maximum score 2,747, and the same page-id/edit-count mapping. Artifacts are in smoke-1000000 under the fixed project. The validation servers exited afterward; the full benchmark remains stopped. These short-replay timings do not establish full-run throughput.