Full Wikimedia benchmark — launched September 8, 2026

Status as of 2026-09-08 19:17:24 UTC: RUNNING, not a completed result. All six servers and full-file clients are alive and progressing; startup client/error logs are empty. Final performance and correctness results are pending.

Run configuration

• Host: naamah; six engines run concurrently without explicit CPU pinning.
• Input: 1,483,700,913 page-ID edit increments, 33,245,740,607 bytes.
• Engines: Goblin standard, Goblin packed INT32/FLOAT32, Redis 8.8, Redis 7.2.4, Valkey 9.1 and Dragonfly (one proactor thread).
• Packed merge exponent: 0.5. Numeric packed member tie ordering is unchanged; standard/incumbents retain lexicographic member ordering.
• Both Goblin modes use the same verified optimized Release binary from the successful limited comparison.
• Ordinary reply-per-command redis-cli feeds over independent Unix sockets; the timed replay path and existing Redis/Valkey defragmentation settings are unchanged. Memory samples are recorded every 300 seconds.
• Feeds started at 2026-09-08T19:14:27Z.

The completed rebuilt source file at /home/adam/wiki2/redis.cmds was copied to an independent read-only Btrfs copy-on-write snapshot inside this project. Its metadata was unchanged across the copy, no source writers were present, and the complete frozen file was counted and hashed before launch. Later edits to the user's original cannot change this run's input. Earlier Wikimedia and Lichess artifacts/writeups remain untouched.

Input SHA-256: 234506a2217a83580d4b47a644818a9f41a51e342d2e0acadfb0ed2e6ced66b1.

Goblin executable SHA-256: 9513ff0e26399b0c68d6b84032a95e4724cdb6e0d2f4c9def984033f5c3b4e8d.

Incumbent/client executable hashes match the limited run. The project keeps its own executable and frozen headers, so subsequent development builds do not affect this benchmark.

Automatic verification and completion

After all feeds finish, the verifier reads each zset in 65,536-member rank pages, bounding server replies and client buffers for the full dataset. The resulting ordered member/score stream is identical to single-reply verification. This path was checked on the saved two-million-command input: the complete result digest matched exactly, including the partial final page. That preflight run is a verification check, not a new comparative timing.

Final checks compare all mapping digests, member counts, each representation's tie ordering, integral scores and the expected 1,483,700,913 score sum. The limited-run digest is deliberately not used as a full-run reference. The controller rechecks the input checksum after the benchmark and records the final status. A run/COMPLETE marker alone is not proof of a passing run; check JOB-PASS/JOB-FAILED, exit.status, and run/verification.txt.

The job survives SSH disconnects in a detached tmux session. It does not checkpoint/resume across a host restart and does not automatically retry. Any new attempt must use a new project directory.

Locations and status

• Project: /home/adam/wiki2/wikimedia-full-20260908T190700Z.
• Frozen input: redis.cmds inside that project.
• Live results/logs: run/ inside that project.
• Controller output: controller.log.
• Completion markers: JOB-RUNNING, then JOB-PASS or JOB-FAILED, plus exit.status and finished.utc.
• Detached session: wiki-full-20260908T190700Z.
• Local launch archive: benchmark-results/wikimedia-full-20260908T190700Z (startup snapshot only).

ssh adam@naamah 'tmux list-sessions'
ssh adam@naamah 'tail -n 30 /home/adam/wiki2/wikimedia-full-20260908T190700Z/controller.log'

The active run owns benchmark ports 16501, 16502 and 16409. Do not launch another copy of the harness using those ports while it is running.