Question
Design the active-active 'reactions' counter for a short-video platform where every video has like/heart/laugh counts that must accept increments in all 5 write regions simultaneously, with no cross-region coordination on the write path (an increment must commit in <10ms local). Scale: 400M DAU, viral videos take 200k reactions/sec for minutes. The displayed total may be eventually consistent (a few seconds stale is fine) but must never go backwards, must never lose increments even when a region is partitioned for 30 minutes, and a user who un-reacts must see their own action reflected immediately on their device. Pick the replicated data type and the convergence/anti-entropy mechanism, and explain how you avoid both lost updates and double-counting on partition heal.
Clarify scale and constraints first. Propose a clean component breakdown, then go deep on the hard parts — data model, bottlenecks, consistency, failure modes — and name the trade-offs you are making.