Full Wikimedia typed-layout and RLE matrix — September 22, 2026
Completed successfully at 2026-09-22 14:42:37 UTC (10:42:37 a.m. EDT). All 12 full replays passed verification: 80,798,328 members, a score sum of 1,483,700,913, identical normalized mapping and ordered digests, and zero command errors. Both input checksums were unchanged, and the controller recorded JOB-PASS with exit status zero. See the completed comparison for results.
The remainder records the launch configuration. The preceding two-million-command replay also passed full-state verification in every variant, with 279,458 members and a score sum of 2,000,000.
Configuration
The matrix contains all six typed packed B+ tree layouts, each with score RLE disabled and enabled, for 12 runs:
| Member type | Score types | RLE modes |
|---|---|---|
| INT32 | FLOAT32, FLOAT64 | off, on |
| INT64 | FLOAT32, FLOAT64 | off, on |
| UUID | FLOAT32, FLOAT64 | off, on |
Every variant uses the same frozen Release executable, merge exponent 0.5, and ordinary reply-per-command redis-cli over its own Unix socket. All 12 run concurrently from empty sets without explicit CPU pinning on naamah (Threadripper PRO 5995WX, 64 cores / 128 threads, approximately 991 GiB RAM). The full feeds began at 2026-09-22T04:10:53Z. Each receives 1,483,700,913 increments, with 80,798,328 members expected at completion.
The CLI selects --zset-implementation packed-<member>-<score> and either --packed-zset-score-rle or --no-packed-zset-score-rle. Each run verifies the resulting GOBLIN.MEMORY score_rle flag. Goblin also opens a loopback TCP listener; the harness assigns distinct unused ports, recorded in run/tcp-ports.tsv. The existing Redis service on port 6379 remains running.
RSS is sampled every 300 seconds and at feed completion. Final RSS and GOBLIN.MEMORY allocation/compression counters are recorded before digest verification, without a final forced compaction. The report compares RLE on/off within each layout. One concurrent trial includes client/protocol costs and does not establish statistical significance or maximum throughput.
Frozen inputs and executable
The numeric input is a read-only Btrfs copy-on-write snapshot of the previous verified full run. Its complete command count and checksum were checked before launch. The UUID file was generated before timing by zero-extending each numeric page ID to 128 bits; for example, page ID 16 becomes 00000000-0000-0000-0000-000000000010. This preserves numeric tie ordering. UUID commands are longer, so cross-layout timings include that wire-size difference. Digests decode UUIDs back to numeric IDs before comparison.
| Input | Bytes | SHA-256 |
|---|---|---|
input/numeric.cmds |
33,245,740,607 | 234506a2217a83580d4b47a644818a9f41a51e342d2e0acadfb0ed2e6ced66b1 |
input/uuid.cmds |
75,668,746,563 | 0a95548ca88e39809bae4bde146a0fe8fc159a30b091e60f242825a3d6479a7c |
The frozen workspace source includes the RLE implementation. Its Git base is e907e4fc705be3896b6d3a380a99689e7f075c3e; the snapshot also includes uncommitted changes, documented by workspace.patch and per-file hashes. Source archive SHA-256: ee480f36691b5f7c40bb5db9c19ea649282a91a3ba952a0d012a21a62dec7735.
Goblin executable SHA-256: 7340c28aa67d763e24553325b391ff77be6f4903c88211025a0b0023494148e6. Redis CLI 8.8.0 executable SHA-256: c45604bc44ce56c426a75723f0652831a93a0f9373d9bc1484ab9be5e2a1d8ca.
The Linux build used GCC 16.1.0 and CMake 4.2.3, Release with native CPU targeting, and Kafka, RDMA, TLS, HTML docs, and benchmark targets disabled. Core, packed-zset, packed-tree, and replication tests all passed. The frozen source, harness, and executable are read-only, so later workspace edits and development builds cannot change this run.
Verification and completion
The controller gated the full replay on the successful two-million-command prefix. That prefix was compared against an independent Python counter of the source commands. After every full feed finishes, the controller streams all final members in 65,536-member rank pages and compares all 12 normalized digests with the previous verified full-run numeric digest. Checks include member count, score sum, integral scores, numeric tie ordering, mapping hash, and ordered hash. Both input checksums are checked again afterward.
The controller automatically writes run/comparison.md, including feed times, RSS, object allocation, compression counters, paired RLE savings, and verification results. Final success requires JOB-PASS and exit.status equal to zero. A run/COMPLETE marker alone only means the replay/verification stage ended.
Locations
- Remote project:
/home/adam/wiki2/wikimedia-rle-matrix-20260922T035401Z. - Detached tmux session:
wiki-rle-full-20260922T035401Z. - Controller log:
controller.log; current stage:phase. - Prefix results:
smoke-2000000/comparison.mdandverification.txt. - Full results:
run/summary.tsv,comparison.md,verification.txt,results/,samples/, anddigests/. - Local startup archive:
benchmark-results/wikimedia-rle-matrix-20260922T035401Z/.
ssh adam@naamah 'tail -n 30 /home/adam/wiki2/wikimedia-rle-matrix-20260922T035401Z/controller.log'
ssh adam@naamah 'cat /home/adam/wiki2/wikimedia-rle-matrix-20260922T035401Z/run/comparison.md'
This job survives SSH disconnects. It does not resume after a host restart or automatically retry; a new attempt needs a fresh project directory.