Registry service benchmark

This page summarizes a set of performance tests conducted on the Registry Service in December 2025 to understand how it behaves under different data sizes and…

This page summarizes a set of performance tests conducted on the Registry Service in December 2025 to understand how it behaves under different data sizes and infrastructure configurations.

The goal is to give operators a practical sense of the throughput and latency they can expect under different data sizes and load levels.

The results below are based on sustained load tests using the open-source k6 tool and reflect steady-state behavior rather than short-lived spikes.

---

✨ At a Glance

🚀 ~2,200 requests/sec Peak throughput	⚡ ~42 ms p95 latency At peak throughput	📈 ~2x throughput gain 2 vCPU → 4 vCPU

---

🔬 Test Methodology

Benchmarks were run under controlled conditions using k6 virtual users against a live Registry Service instance.

Traffic Simulation

Parameter	Value
Concurrent Users	50–100 virtual users
Duration	10 minutes (R0: 2 minutes)

Infrastructure Configurations Tested

Config	Memory	vCPU
Small	4 GB	2
Medium	8 GB	2
Large	16 GB	4

Dataset Shapes Tested

Namespaces	Registries / Namespace	Records / Registry	Total Records
500	1	200	100,000
2,000	2	100	400,000
2,000	2	200	800,000
3,000	2	250	1,500,000
1,500	2	500	1,500,000

All reported values represent steady-state performance.

---

📈 Run Results

Reading the latency columns p90 = 90% of requests completed within that time. p95 captures the 95th percentile — the experience of nearly all requests, including slower ones.

#	Thread Count	Namespaces	Registries / Namespace	Records / Registry	Total Records	Duration	Memory	CPU	Iterations	TPS	Avg (ms)	Min (ms)	Med (ms)	Max (ms)	p90 (ms)	p95 (ms)
R0	50	500	1	200	100,000	2m	4GB	2	123,938	1032.34	48.18	0.31	46.83	286.86	70.71	78.63
R1	50	500	1	200	100,000	10m	4GB	2	656,173	1093.53	45.5	0.32	44.56	284.97	66.21	72.92
R2	50	500	1	200	100,000	10m	8GB	2	696,003	1159.94	42.88	0.29	42.27	175.11	61.88	67.86
R3	100	500	1	200	100,000	10m	8GB	2	693,374	1155.47	86.31	0.3	85.21	523.17	118.7	129.11
R4	50	2,000	2	100	400,000	10m	8GB	2	1,228,847	2048.01	24.17	0.27	23.22	317.04	36.45	40.82
R5	50	2,000	2	200	800,000	10m	8GB	2	676,575	1127.54	44.08	0.3	42.83	1035.82	63.83	70.69
R6	50	2,000	2	200	800,000	10m	16GB	4	1,337,820	2229.62	22.17	0.27	20.84	134.52	36.38	41.82
R7	50	3,000	2	250	1,500,000	10m	16GB	4	1,070,116	1783.44	27.78	0.28	26.56	208.58	43.78	49.46
R8	50	1,500	2	500	1,500,000	10m	16GB	4	575,289	958.75	51.89	0.25	46.45	301.22	87.21	103.98

---

🔍 Key Findings

⚡ Infra Scaling

Scaling from 2 to 4 vCPUs on an 800K-record dataset nearly doubled throughput while cutting tail latency by over 85%.

Metric	8 GB / 2 vCPU	16 GB / 4 vCPU	Improvement
Throughput	1,127 TPS	2,229 TPS	~2x 🚀
p95 Latency	70.7 ms	41.8 ms	41% faster 🚀
Max Latency	1,035 ms	134 ms	87% faster 🚀

📦 Dataset Scale

With well-distributed data, the Registry Service handles 1.5 million records efficiently on a 16 GB / 4 vCPU instance — sustaining ~1,800 TPS with p95 under 50 ms.

Records	Namespaces	TPS	p95 Latency
100,000	500	1,159	67.9 ms
800,000	2,000	2,229	41.8 ms
1,500,000	3,000	1,783	49.5 ms

The service scales gracefully across dataset sizes — even at 15× the baseline record count, throughput stays above 1,700 TPS.

---

🎯 Key Takeaways

• Highly Scalable: Throughput nearly doubles when scaling from 2 to 4 vCPUs — consistent, predictable gains with additional compute.
• Low Latency at Scale: Even at 800K records, the service sustains ~2,200 TPS with p95 latency of just ~42 ms on a 16 GB / 4 vCPU instance.
• Stable Under Sustained Load: No latency degradation observed over 10-minute test runs, demonstrating robust steady-state behavior.
• Handles Large Datasets: With well-distributed data, the service sustains ~1,800 TPS even at 1.5 million records.

---

Final note

These results highlight the Registry Service's ability to deliver consistent, predictable performance under sustained load across a wide range of dataset sizes and configurations. The service demonstrates clear scaling behavior, improving both throughput and latency as additional CPU and memory are made available.

While real-world performance will naturally vary based on data layout and usage patterns, the observed results show that the Registry Service can comfortably support high-throughput, low-latency workloads when appropriately sized.

Overall, the tests reinforce that the Registry Service is well-suited for production workloads, with performance characteristics that scale in a transparent and predictable way.

_{Benchmarking data collected using k6 against the NFH fabric Registry Service in December 2025.}