Pre-warm Flow // pixelmarket · architecture

Volunteer compute isn’t in the hot path. It runs ahead of demand, batch-filling the cache for trending prompts. Live traffic only ever hits CDN or cloud GPU.
Cache hit rate
68%
with pre-warming  ·  34% without
● User prompt request HTTP · edge ▩ CDN Cache lookup(hash) ~40ms · global ◆ Cloud GPU on-demand infer ~2.0s · paid ⇧ Trend Analyzer top-100 / hour embeddings · kNN ☰ Job Queue batch · FIFO no SLA ♦ Worker Pool volunteer GPUs redundant · untrusted rewarded in PX ✓ Verifier 3x redundancy CLIP match hit · 300ms miss → gpu → warm · 2.0s pre-warm · async

Three concurrent paths

  • Live hit 300 ms
  • Live miss → cloud GPU 2.0 s
  • Pre-warm (volunteers) async

Path details

Pre-warm path
Volunteer workers run in background, batch-generating variations of trending prompts (the top 100 each hour). Results are verified via redundancy and written to cache. When a real user requests a similar prompt, it’s already warm. Volunteers get PX proportional to verified jobs completed. Zero latency pressure on volunteer nodes.