Imported from jeremylongshore/tons-of-skills-marketplace (
skills/.curated/hex-performance-tuning/SKILL.md). Install upstream withnpx skills add jeremylongshore/tons-of-skills-marketplace --skill hex-performance-tuning. Copyright stays with the author (MIT).
Hex Performance Tuning
Latency Benchmarks
| Operation | Typical Duration |
|---|---|
| ListProjects | 200-500ms |
| RunProject (trigger) | 500ms-2s |
| Project execution | 10s-30min (depends on queries) |
| GetRunStatus (poll) | 100-300ms |
Instructions
Cache Project Lists
import { LRUCache } from 'lru-cache';
const projectCache = new LRUCache<string, any>({ max: 50, ttl: 300000 }); // 5 min
async function getCachedProjects(client: HexClient) {
const cached = projectCache.get('projects');
if (cached) return cached;
const projects = await client.listProjects();
projectCache.set('projects', projects);
return projects;
}Parallel Independent Runs
// Run independent projects in parallel (respecting rate limits)
async function parallelRuns(client: HexClient, configs: Array<{ id: string; params: any }>) {
return Promise.allSettled(
configs.map(c => runWithRetry(client, c.id, c.params))
);
}Optimize Poll Interval
// Adaptive polling: start fast, slow down
async function adaptivePoll(client: HexClient, projectId: string, runId: string) {
let interval = 2000; // Start at 2s
while (true) {
const status = await client.getRunStatus(projectId, runId);
if (['COMPLETED', 'ERRORED', 'KILLED'].includes(status.status)) return status;
await new Promise(r => setTimeout(r, interval));
interval = Math.min(interval * 1.5, 30000); // Max 30s
}
}Overview
Tune run latency and throughput using safe sandbox projects and aggregate metrics. A gain is invalid if it expands data scope, compromises output correctness, exceeds quota, or makes rollback impossible.
Prerequisites
- Baseline latency/run metrics, safe fixture revision, approved error budget, and a rollback revision for cache, concurrency, parameters, and retry policy.
Output
Return a tuning receipt with baseline/canary percentile bands, cache/concurrency/parameter revisions, quota/error outcomes, aggregate assertion, owner approval, and rollback reference. Use aggregates only.
Error Handling
Roll back for quota saturation, increased errors, changed output assertion, access drift, or duplicate runs. Do not raise concurrency or cache duration to hide a failing dependency.
Examples
env=sandbox; p95=420ms->310ms; concurrency=2; cache=r4; quota=within-budget; assertions=pass; rollback=perf-r3 documents a safe canary.