The problem#
Mosaic's public site was a React SPA that fetched everything client-side. Time-to-interactive was 4+ seconds on mid-range mobile, and bounce rate tracked it almost one-to-one. The business wanted faster, but the team also wanted to stop hand-rolling cache invalidation.
The re-architecture#
We moved the site to server-rendered Next.js with three layers of caching: static pages pre-rendered at build time, dynamic pages cached at the edge with tag-based invalidation, and per-user data streamed in after first paint.
The load-bearing decision was the caching model:
export function middleware(req: NextRequest) {
const url = req.nextUrl;
// Public pages: cache at the edge, revalidate on tag invalidation.
if (isPublicRoute(url.pathname)) {
return NextResponse.next({
headers: {
"x-cache": "edge",
"cache-control": "public, max-age=0, s-maxage=300",
},
});
}
// Dashboards are dynamic but served stale-while-revalidate.
return NextResponse.next();
}The database, honestly#
Most of the win came from SQL, not from the framework. Three changes did the heavy lifting:
- Read replicas with query routing — 80% of queries never touch the primary.
- Materialized views for dashboard aggregates, refreshed on a 30-second
cadence instead of per-request
COUNT(*). - Connection pooling tuned per-service — pool exhaustion was the #1 cause of the old site's p95 tail.
CREATE MATERIALIZED VIEW dashboard_rollups AS
SELECT
account_id,
date_trunc('hour', created_at) AS hour,
count(*) AS events,
sum(amount_cents) AS volume_cents
FROM events
WHERE created_at > now() - interval '90 days'
GROUP BY 1, 2
WITH NO DATA;
CREATE UNIQUE INDEX ON dashboard_rollups (account_id, hour);Edge caching is not free
Tag-based invalidation is elegant until a tag fans out to 40k pages. We learned to model tags at the section level (a feed, a profile) rather than per-entity, and to treat "eventually consistent by 60s" as a product decision, not an engineering accident.
What stuck#
The lesson I repeat most from Mosaic: performance work compounds when you make it visible. We shipped a latency budget into CI — every route declares a p95 budget and a synthetic check enforces it. Regressions get caught in review instead of in the on-call channel.