Goblin Core ARRAY Benchmark

Generated on a quiet dedicated benchmark host at 2026-07-16 23:02:24 UTC.

Summary

This report compares Goblin Core's Classic and real-time ARRAY implementations with the Redis-compatible servers in the benchmark suite, using only native AR* commands over RESP/TCP. It loads 500,000 16-byte values into dense and stride-16 sparse arrays, measures population throughput and process RSS, then exercises eight read and mutation workflows at pipeline depth 256. A separate 500,000-append, pipeline-1 test reports P90 through P99.99 latency. Engines without native ARRAY commands are marked X; hashes are never used as a fallback.

Classic is the memory-oriented implementation for sparse or evolving index ranges. RT trades fixed dense leaves and additional capacity for a more predictable mutation path; its reserved row provisions and prefaults hard index, metadata, and byte limits before timing. Two results stand out: resident-locked reserved RT records 36.11 us P99.99 versus 54.48 us for Redis 8.8; and Classic uses 30.64 RSS-delta bytes/value on the sparse shape versus 169.57 for Redis 8.8.

Method

• 500,000 logical 16-byte values in both dense (stride=1) and sparse (stride=16) shapes.
• Population and throughput rows use RESP/TCP pipeline depth 256; individual-increment latency uses pipeline depth 1.
• Goblin Classic uses adaptive sparse/dense leaves and a growable directory. Goblin RT uses fixed-size dense leaves, a fixed-depth directory, and a separate 2x value-arena growth policy.
• Every server receives native ARSET, ARMSET, ARGET, ARMGET, ARCOUNT, ARLEN, ARDEL, and ARINSERT requests. There is no hash-command fallback.
• Operation rates are medians of 3 native C++ client rounds. Churn counts one delete + restore pair as one logical operation.
• Servers run one at a time on one serving core; Dragonfly uses one proactor and mini-redis-go uses GOMAXPROCS=1.
• X means the server returned a command error for the required AR operation. Error replies are not timed.
• RSS is sampled from the launched PID (ps for mini-redis-go; /proc for the others), never from a server-reported RSS field.
• Incumbents are exercised strictly through RESP as black boxes; their source is not inspected.
• Goblin rows were refreshed at 2026-07-16 23:02:24 UTC; unchanged incumbent rows are retained from the same-host 2026-07-16 21:17:44 UTC run.

Individual Increment Latency

Each of the 500,000 samples is one dense ARSET key index value append that advances a fresh array's logical length by one. Values are 16 bytes, indexes are sequential, connection setup is excluded, and RESP/TCP pipeline depth is 1. At this sample count, the P99.99 tail contains approximately 50 commands. The goblin-rt-reserved row issues GOBLIN.RT.ARRESERVE before timing, prefaulting its complete index, value-slot metadata, and encoded-byte budget. Reservation time and connection setup are excluded; exhaustion would fail instead of allocating in a measured command. Immediately after timing, /proc/<pid>/status reported goblin-classic-resp-tcp 26.12 MiB locked / 0.00 MiB swapped; goblin-rt-resp-tcp 25.66 MiB locked / 0.00 MiB swapped; goblin-rt-reserved-resp-tcp 28.27 MiB locked / 0.00 MiB swapped. This verifies the benchmarked mappings were resident-locked and not swapped.

Dense Indexes

Population And Memory

Operations

Sparse Indexes, Stride 16

Population And Memory

Operations

Interpretation

Classic and RT solve different deployment problems. Classic is the memory-oriented choice for sparse or evolving index ranges. RT fixes each touched leaf at its final size and never rebuilds its fixed-depth directory; an out-of-range write fails. Its ordinary value arena grows 2x to reduce relocations. GOBLIN.RT.ARRESERVE goes further by allocating and prefaulting the declared index, metadata, and byte capacity before serving, then failing closed at any bound instead of allocating from a timed mutation.

All comparison rows exercise native AR commands. A row marked X does not expose the required command; the harness never substitutes hashes or times an error reply as useful work.