What we shipped on Cloudflare Workers, what broke, and what we'd reach for again.
We moved our default production target to Cloudflare Workers in late 2023. Two years and eleven projects later, here is the honest accounting.
What worked: cold starts are no longer a category we think about. Global latency dropped by a factor of three on average. The pricing model rewards the kind of architecture we already favoured.
What did not: filesystem assumptions break in surprising places. The Node compatibility layer is good but not invisible. Long-running tasks need a different home.
Where we'd start fresh today: Workers for the edge, Durable Objects for state, and a single Postgres for everything that resists distribution. Boring, fast, cheap.