Goblin Core ZSET Benchmark

Generated on a quiet dedicated benchmark host at 2026-07-16 19:23:35 UTC.

Summary

After loading and packing 1,000,000 members, Goblin Core uses 52.32 RSS-delta bytes/member versus 55.26 for the leanest incumbent. RESP/TCP same-member ZADD reaches 1.06M operations/s, 1.13x the fastest incumbent; typed SBE over 2 MiB request and reply rings reaches 3.39M operations/s.

Method

1,000,000 distinct 16-byte members in one sorted set, with deterministic scattered unsigned 32-bit integer scores.
Population uses ZADD batches of 128. Every timed operation uses pipeline depth 256.
Goblin runs GOBLIN.OPTIMIZE key 0.97 after population and before memory/operation measurements.
Goblin Core is measured over RESP/TCP and typed SBE over 2 MiB request and reply rings; every incumbent uses RESP/TCP.
Fixed rows are medians of 3 native C++ client rounds. ZREM + ZADD counts as one logical churn operation.
One serving core per engine. Dragonfly uses one proactor; mini-redis-go uses GOMAXPROCS=1.
A server without sorted-set support remains in the matrix with n/a cells; error replies are not timed.
RSS comes from the launched server PID (ps for mini-redis-go, /proc for the others). Server-reported RSS is not used.
Incumbent implementations are black boxes; no incompatible source is inspected.

Population And Memory

Engine	Transport	members/s	RSS MiB	RSS delta MiB	RSS delta B/member	key-reported B/member
`goblin-resp-tcp`	`resp`	1.08M	56.02	49.90	52.32	51.34
`goblin-sbe-ring`	`sbe`	1.42M	64.22	50.07	52.51	51.34
`redis-7.2.4`	`resp`	925K	111.64	104.68	109.76	99.59
`redis-8.8`	`resp`	905K	86.84	78.85	82.68	95.73
`valkey-9.1`	`resp`	910K	88.31	80.64	84.56	78.27
`dragonfly`	`resp`	1.24M	74.77	52.70	55.26	54.30
`mini-redis-go-55178df`	`resp`	n/a	n/a	n/a	n/a	n/a

Operations

Logical operations per second.

Operation	`goblin-resp-tcp`	`goblin-sbe-ring`	`redis-7.2.4`	`redis-8.8`	`valkey-9.1`	`dragonfly`	`mini-redis-go-55178df`
`ZADD existing member`	1.06M	3.39M	946K	920K	900K	452K	n/a
`ZSCORE hit`	1.70M	3.77M	1.64M	1.34M	1.45M	1.07M	n/a
`ZRANK`	1.08M	1.61M	809K	765K	797K	663K	n/a
`ZRANGE first 16`	740K	2.80M	575K	560K	504K	380K	n/a
`ZCARD`	2.96M	9.01M	2.40M	1.99M	1.87M	1.46M	n/a
`ZREM + ZADD`	495K	760K	424K	372K	386K	323K	n/a

Interpretation

The RESP rows isolate the sorted-set implementation under the same wire protocol. The ring row answers a different deployment question: how much transport and parsing work remains when a native client submits the same pipelined operations as typed SBE messages without a syscall or ASCII conversion. It is reported alongside, never substituted for the protocol-parity result.

The scattered score generator crosses signed 32-bit range, so Goblin stores this population at f64 score width. Workloads whose scores fit i16 or signed i32 can use less memory.

Tested Servers

goblin-resp-tcp
goblin-sbe-ring
redis-7.2.4
redis-8.8
valkey-9.1
dragonfly
mini-redis-go-55178df