Historical report imported 2026-10-10 from docs/experiments/benchmarks/block-tail-run-01.md. Original SHA-256: 0965aea39bbdea97052f29da07fd573ab8ade17412f35a244273df605d661d95. Measurements and outcomes retain their stated revisions. Links and local environment labels are normalized; $FABRIC_DATA is the mounted experiment drive. See evidence policy.

Block tail run 01: the product's walk over a canonical block tail

Status: Exploratory. Measured on 2026-10-03 for ADR-0024, part 3, on the product tree. No protocol was registered; nothing here is Measured in the evidence-state sense. It is the server-path run that hypothesis A3 of the suite asked for.

Setup

As query walk run 01: one release build of the product tree, query_plan=scan against query_plan=walk, the real-text 64 MiB unsealed tail (149,585 entries, 619,170 records), the fourteen L-21 shapes, eight repetitions, the median, in ms, a 200-query random differential with pages, CPUs 0 and 1. The walk now reads every complete tail block (4,096 records, closed at a frame boundary, Zstd-compressed FOB1 with a trigram filter over its log bodies) through decode_view, and only the entries after a file's last complete block through their OTLP bytes (product_blocks.json (local evidence omitted from stage), product_walk.py (local evidence omitted from stage)). The 64-Segment state of the same run was sealed before part 2 and holds no text filters; its figures are in the JSON and do not bear on part 3.

Results

Shape	scan	walk without blocks (query walk run 01)	walk with the block tail	block tail against scan
logs limit 50	631	26.0	26.2	0.042
logs limit 1,000	625	34.5	33.2	0.053
logs limit 10,000	715	160.9	125.4	0.175
one node, limit 50	569	14.9	16.9	0.030
one metric, limit 100	580	22.7	15.1	0.026
text, 35 % of lines	604	28.5	27.5	0.046
text, 1 %	614	40.5	30.8	0.050
text, 0.1 %	601	202.3	65.7	0.109
text, no match	614	549.7	27.7	0.045
text, no match, last 60 s	582	34.0	13.5	0.023
rate, whole window	596	213.0	175.1	0.294
logs, last 10 s	621	13.3	11.4	0.018
one node, last 60 s	595	13.1	17.8	0.030
empty window in the past	633	11.2	10.7	0.017
200 random queries with pages (s)	255.08	12.92	11.42
peak RSS (MiB)	279	62	113

Differential: 0 mismatches in 400 random queries with pages (200 on the tail, 200 on the Segments).

Verdict

A3's registered statistic is met through the server. The empty-window shape answers at 0.017 of the scan, limit 50 at 0.042 and the no-match text search at 0.045, all under one fifth, with a clean differential. The no-match search, which the walk could not stop (550 ms), now costs 28 ms: every block's filter rejects the needle, so only the index and the few entries outside complete blocks are read. That is the A3 mechanism together with C5's filter per block, as optimality run 01 found was needed; blocks alone measured 0.26 to 0.28 there. The rare token falls from 202 to 66 ms, the 10,000-row page from 161 to 125 ms and the metric shape from 23 to 15 ms, because a block decodes at about 0.13 µs per record against 4.7 µs per OTLP entry. The whole-window rate (175 ms) still reads every metric point, which is its floor.

Memory: peak resident 113 MiB against 62 MiB for the walk without blocks and 281 MiB for the scan; the same tail compressed to 4.1 MB in 65,536-record blocks in optimality run 01, and at this run's 4,096-record blocks its size and the share of the rise due to building them on the first query were not measured separately.

Limits

• One host, warm page cache, one tail; the target profile is not measured.
• The blocks are derived in memory and rebuilt after a restart (the first query after start builds the index and the blocks); their build cost was not separated from the index's in this run.
• Correctness evidence: a_block_tail_answers_as_the_scan_does (small blocks over a mostly unsealed history from three Spindles at identical instants, page for page against the scan, with a full drain; an injected entry-boundary defect in block reading fails it), the unit test that a block the codec refuses is not made, part 1's tests, and this differential.