Benchmarks

Benchmarks

Date: 2026-08-30 Stack: Go 1.26 + Fiber + mattn/go-sqlite3, Node 22 + hyper-express + better-sqlite3, Bun 1.4 + Bun.serve + bun:sqlite Environment: Mac M4 (localhost), 2-4 nodes per runtime, 50ms batch interval, batch-size 10000

Dual-Writer Throughput (Point-to-Point)

10 runs × 100 concurrent requests per node (200 total per run). Both nodes receive writes simultaneously.

Runtime	QPS median	QPS min	QPS max	Integrity
Go	9,149	3,705	15,063	10/10 PASS
Bun	10,476	3,004	14,159	10/10 PASS
Node	8,393	3,458	10,068	10/10 PASS

Integrity: both nodes have equal item count, 0 pending changes, 0 dead letter after sync settles.

Sync Does Not Block Writes

Having a peer does not slow down writes. Sync runs in a separate timer + HTTP POST.

Runtime	Single (no peer)	With peer	Difference
Go	12,668 QPS	13,436 QPS	+6% (noise)
Bun	8,099 QPS	8,578 QPS	+6% (noise)
Node	14,764 QPS	14,701 QPS	-0.4% (noise)

Sync Delay = Batch Interval

Interval	Sync p50	Burst sync (100 writes)	Use case
10ms	12ms	20ms	Local/LAN
50ms (default)	52ms	25ms	Remote/WAN
100ms	100ms	22ms	Conservative

Full Mesh (4 nodes all-to-all)

Runtime	QPS median	QPS min	QPS max	Integrity
Go	14,853	6,108	23,570	5/5 PASS
Node	14,414	3,165	18,487	5/5 PASS
Bun	13,175	1,872	18,887	5/5 PASS

Dedicated Hub (1 Go hub + 3 edges)

Runtime (edges)	QPS median	QPS min	QPS max	Integrity
Go	16,268	5,706	17,516	5/5 PASS
Bun	19,750	12,931	27,683	5/5 PASS
Node	13,193	148	21,660	5/5 PASS

Convergence Speed

Items	Before (fixed LIMIT 100)	After (batch 10K + drain)	Speedup
10K	~6s	<1s	6x
100K	~60s	~2s	30x

Volume Stress Test

Writes 10K, 100K, and 500K items via batch endpoint, then verifies convergence, consistency, and persistence (kill + restart).

Volume	Runtime	Write time	Converge time	Consistency	Persistence
10K	Go	96ms	1s	✅ PASS	✅ PASS
10K	Bun	58ms	1s	✅ PASS	✅ PASS
10K	Node	50ms	1s	✅ PASS	✅ PASS
100K	Go	891ms	1s	✅ PASS	✅ PASS
100K	Bun	769ms	1s	✅ PASS	✅ PASS
100K	Node	618ms	1s	✅ PASS	✅ PASS
500K	Go	4.2s	5s	✅ PASS	✅ PASS
500K	Bun	10.7s	5s	✅ PASS	✅ PASS
500K	Node	10.2s	7s	✅ PASS	✅ PASS

9/9 PASS. Zero data loss, zero dead letters across all volumes and runtimes.

Split-Brain Safety

36 checks across all 3 runtimes — 36/36 PASS. See Split-Brain Safety for details.

Real Network Test (OVH → 1TIM)

10,000 items in single batch, real network (2.7ms RTT, 289 Mbps):

Metric	Result
Converge time	<1s
Data loss	0
Dead letter	0

Crash Recovery

Write 50 items → kill node A → verify 50 pending in _changes → restart → sync resumes → both nodes converge. PASS.

Dead Letter Queue

Connection errors retry indefinitely — no data loss, no dead letter. Only ACK mismatch (protocol error) moves to _dead_letter. PASS.

hook-sync vs Postgres (100K writes, real network)

Both configured with equivalent durability (no fsync per write):

System	Durability	QPS median	Replica converge	Integrity
hook-sync	synchronous=NORMAL	6,065	~2s (batch 10K + drain)	100K, 0 pending, 0 dead letter
Postgres	synchronous_commit=off	6,238	~3s (WAL streaming)	100K

At equivalent durability, throughput is tied — Postgres +2.8% (within noise).

Batch Mode: HTTP Overhead Bypass

Batch size	hook-sync (SQLite)	Postgres	SQLite advantage
1 (single)	6,065 QPS	6,238 QPS	tie (HTTP dominates)
100	27,366 QPS	22,703 QPS	+20.5%
1,000	31,429 QPS	23,682 QPS	+32.7%
10,000	31,558 QPS	8,278 QPS	+3.8x

At batch 10K, SQLite’s single-transaction insert bypasses HTTP overhead per row. Postgres degrades because pgx pool handles rows individually.

Where Each System Wins

Metric	hook-sync	Postgres	Winner
Write throughput (fair durability)	6,065	6,238	Tie
Sync overhead	~0% (background)	WAL sender overhead	hook-sync
Replica lag	~2s	~3s	hook-sync
Cross-runtime	Go, Bun, Node	Go-only (pgx)	hook-sync
Topology	P2P, mesh, hub	Primary-replica only	hook-sync
Multi-writer	Yes (UUID, idempotent)	No (primary-only)	hook-sync
Operational complexity	Single binary + SQLite file	Cluster + replication config	hook-sync

Compression Analysis (NOT implemented)

Tested gzip on 290 Mbps link: CPU cost (20-47ms) exceeds transfer save (0.5-50ms). Compression NOT worth it on fast links. Would only help on slow links (<50 Mbps).