Performance & Benchmarks
We time the HUD the same way we judge a newsroom deadline: can it deliver before the coffee cools? Hereβs how we benchmark the feed endpoint and what the numbers mean.
Methodology
- Hardware β Apple M2 Pro (16 GB RAM, macOS 14).
- Runtime β Node v22.17.1, npm 10.9.2, Next 15.5.3, React 19.1.0.
- Dataset β ~1,000
itemsspanning 10 sources (RSS + X) with 30-day retention. - Tooling β
autocannonto hammer/api/feedafter priming the in-memory cache.
How to run it yourself
npx autocannon -d 20 -c 20 http://localhost:3000/api/feed
Example run (local dev)
| Metric | Value | Why it matters |
|---|---|---|
| Req/s | ~750 | Cache keeps hot cursors in-memory, no extra DB hops |
| P95 latency | ~48 ms | JSON mapping only; no SSR or heavy transforms |
| CPU peak | ~65 % | Single Node process cruising without bottlenecks |
Note Treat these numbers as relative. Rerun the same command before and after refactors to prove you sped things up.
Visual pulse
Req/s
800 |βββββββββββββββββββββββ
700 |ββββββββββββββββββββ
600 |ββββββββββββββββββ
500 |βββββββββββββββ
+----+----+----+----+
5 10 15 20 s
Reading the tea leaves
- The feed endpoint scales nicely because itβs stateless and responds with lean JSON payloads.
- Indexes (
items_published_at_idx,sources_status_idx) matter; dropping them tanks throughput. - Access patterns improve if you avoid
select *in Supabase queriesβpull only what you render.
Limitations
Local benches arenβt production truths. Re-run on a Vercel preview using the same dataset to catch network nuances.