# Hyperliquid Node Benchmark: Sentry Peering Impact

## Overview

Comparison of `book_diffs` latency on a Hyperliquid non-validator node **with** and **without** direct sentry peering. All values in milliseconds.

---

## Side-by-Side Results (15-minute window)

| Metric | Without Sentry | With Sentry | Delta | Improvement |
| ------- | -------------: | ----------: | ----------: | ----------: |
| Samples | 10,232 | 10,931 | +699 | +6.8% |
| Min | 20.0 ms | 14.0 ms | −6.0 ms | −30.0% |
| **p50** | 126.0 ms | 113.0 ms | −13.0 ms | −10.3% |
| **p90** | 200.0 ms | 176.0 ms | −24.0 ms | −12.0% |
| **p95** | 815.0 ms | 182.0 ms | −633.0 ms | −77.7% |
| **p99** | 1,904.0 ms | 202.0 ms | −1,702.0 ms | −89.4% |
| Max | 2,647.0 ms | 487.0 ms | −2,160.0 ms | −81.6% |
| Mean | 208.6 ms | 120.7 ms | −87.9 ms | −42.1% |

---

## Detailed Breakdown by Duration

### 5-Minute Window

| Metric | Without Sentry | With Sentry | Improvement |
| ------ | -------------: | ----------: | ----------: |
| n | 3,481 | 3,645 | +4.7% |
| Min | 21.0 ms | 14.0 ms | −33.3% |
| p50 | 122.0 ms | 113.0 ms | −7.4% |
| p90 | 199.0 ms | 176.0 ms | −11.6% |
| p95 | 730.0 ms | 183.0 ms | −74.9% |
| p99 | 1,660.0 ms | 208.0 ms | −87.5% |
| Max | 2,549.0 ms | 487.0 ms | −80.9% |
| Mean | 197.0 ms | 121.3 ms | −38.4% |

### 10-Minute Window

| Metric | Without Sentry | With Sentry | Improvement |
| ------ | -------------: | ----------: | ----------: |
| n | 6,763 | 7,270 | +7.5% |
| Min | 21.0 ms | 14.0 ms | −33.3% |
| p50 | 126.0 ms | 112.0 ms | −11.1% |
| p90 | 201.0 ms | 175.0 ms | −12.9% |
| p95 | 762.0 ms | 182.0 ms | −76.1% |
| p99 | 1,713.0 ms | 205.0 ms | −88.0% |
| Max | 2,591.0 ms | 487.0 ms | −81.2% |
| Mean | 203.2 ms | 119.0 ms | −41.4% |

### 15-Minute Window

| Metric | Without Sentry | With Sentry | Improvement |
| ------ | -------------: | ----------: | ----------: |
| n | 10,232 | 10,931 | +6.8% |
| Min | 20.0 ms | 14.0 ms | −30.0% |
| p50 | 126.0 ms | 113.0 ms | −10.3% |
| p90 | 200.0 ms | 176.0 ms | −12.0% |
| p95 | 815.0 ms | 182.0 ms | −77.7% |
| p99 | 1,904.0 ms | 202.0 ms | −89.4% |
| Max | 2,647.0 ms | 487.0 ms | −81.6% |
| Mean | 208.6 ms | 120.7 ms | −42.1% |

---

## Key Takeaways

- **Median (p50) improvement is modest: ~10%** — going from 126 ms to 113 ms. For the typical update, sentry peering shaves about 13 ms.

- **Tail latency is where sentry peering dominates.** The p95 drops from 815 ms to 182 ms (−78%), and the p99 drops from 1,904 ms to 202 ms (−89%). Without sentry peering, 5% of updates arrive with 4–13× more delay than the median.

- **With sentry peering, the distribution is extremely tight.** The gap between p50 (113 ms) and p99 (202 ms) is only 89 ms. Without it, that gap is 1,778 ms — a 20× wider spread.

- **Max latency drops from 2.6s to 0.5s** — eliminating the worst-case spikes that can be fatal for latency-sensitive strategies.

- **Throughput slightly increases** — the sentry-peered node processed ~7% more book diffs in the same time window, likely because it wasn't stalling on delayed gossip hops.

- **Results are stable across time windows** — the with-sentry numbers barely move between 5, 10, and 15 minutes, while without-sentry tail latencies drift upward over time.