Skip to main content

Measure the graph, transport, and projection separately

Normalization reduces copied writes, but it does not make every query or render free. Measure entity write frequency, subscriber count, list joins, transport latency, persistence time, and bundle payload independently.

  • Keep lists as IDs so a canonical change does not rebuild copied rows.
  • Let RealtimeManager coalesce bursty updates; change its 16 ms interval only from observed latency or render evidence.
  • Use local view mode only when completeness is proven. Hybrid mode can provide immediate local results while remote data reconciles.
  • Tune stale and GC times to real navigation/reconnect behavior. Aggressive revalidation can hide a broken subscription while increasing I/O.
  • Keep CRDT, A2A, A2UI, Tauri, and UI packages optional so consumers pay only for selected capabilities.
  • Compare bundle sizes with the same minification, tree-shaking, compression, and peer-dependency methodology.

Lighthouse delivery budgets are mobile performance ≥90, accessibility/best practices/SEO ≥95, LCP ≤2.5 s, and CLS ≤0.1. They are T3 deployment gates, not the edit loop.