Short Wikimedia comparison — September 8, 2026

The fixed Goblin standard build completed this short replay 2.2–3.0% sooner than Redis/Valkey. Dragonfly was fastest. Packed int32/float32 used less memory than Goblin standard, but took 3.63 times as long.

Workload

One concurrent run of all six engines on naamah, using the first 2,000,000 commands from /home/adam/wiki2/redis.cmds. Each command increments a page's edit count. The run used the existing Wikimedia harness and ordinary reply-per-command redis-cli feeds over Unix-domain sockets. Feed time excludes server startup and final digest generation. Memory was sampled every ten seconds; the table reports final process RSS, including the server baseline.

Goblin uses the tested Release build containing the standard-zset rescore fix. The packed lane uses packed-int32-float32, merge exponent 0.5, and numeric member tie order. Standard Goblin and the incumbents retain lexicographic member tie order. Redis/Valkey retain the harness's active-defragmentation settings and disabled persistence; Dragonfly runs with one proactor thread.

Results

Engine Feed seconds Commands/s Final RSS (MiB)
Dragonfly 34.256 58,384 36.8
Goblin standard 37.024 54,019 35.7
Valkey 9.1 37.872 52,809 28.2
Redis 8.8 37.895 52,777 30.4
Redis 7.2.4 38.174 52,392 36.3
Goblin packed int32/float32 134.304 14,892 31.7

Goblin's internal zset allocation reports were 10,661,794 bytes for standard and 7,678,944 bytes for packed, including the packed member-to-score Swiss table and ordered tree. These object allocation counts differ from whole process RSS and are not directly comparable with incumbent process RSS.

All six lanes passed final verification: 279,458 members, score sum 2,000,000, maximum score 11,792, identical page-id/edit-count mapping digests, and zero command errors. Each lane's own tie ordering was checked. All benchmark clients and servers exited after verification; the earlier full replay remains stopped.

This is one short run. The small standard-versus-Redis/Valkey margin is an observation, not a statistically established speed advantage or a measurement of full-dataset scaling.

Provenance and artifacts

The directories contain the harness, software hashes, metadata, samples, per-engine memory reports, logs, summary, and digests. Lichess data and reports and the earlier Wikimedia runs were preserved.