Local-first is the right default for tools people use on bad networks: a POS, a training log, a field form. The demo is "it works on a plane." The product is what you show when two replicas disagree.
CRDTs merge text. They do not merge "inventory is 1" and "two people checked out." That is a business rule wearing a sync engine.
Questions that have to be answered in writing
- Who wins a conflict: last write, first write, or a human?
- Can the user see they are offline, or do we pretend?
- When does a local write become "real" for money and stock?
- How do you debug a replica that diverged three days ago?
I built offline-first behavior into Gloves Up because a boxing session should not depend on a cell tower. That is a closed, personal dataset. A marketplace like RentaraH is the opposite: availability is a shared ledger. Pretending those are the same "local-first architecture" is how you double-book a vehicle.
A practical split
Use local-first for user-owned state (drafts, prefs, caches, session logs). Use a server as source of truth for shared scarce resources. Sync the former. Transaction the latter.
Libraries will get better. The product questions will not go away.
Takeaways
- Offline is a UX contract, not a library choice.
- Shared inventory is not a CRDT default.
- Write the conflict rule before you pick the sync engine.