Design a scalable, fault‑tolerant notification system that supports email, SMS, and push notifications for a global user base. How would you architect it to handle millions of messages per day while ensuring eventual consistency and low latency?

Context: Define scope and constraints. Identify: List key requirements (scalability, fault tolerance, eventual consistency, low latency). Recommend: Propose a microservices architecture with a message broker (Kafka/SQS), separate delivery services, and a retry/compensation layer. Clarify: Explain consistency model (eventual) and latency targets. List: Detail components – load balancer, API gateway, service registry, monitoring stack, and CDN for push. Evaluate: Discuss trade‑offs (CAP theorem, latency vs consistency). Summarize: Highlight how the design meets throughput, resilience, and observability.

To architect a notification system that scales to millions of messages daily, I would start with a microservices approach. A stateless API gateway fronts the system, routing requests to a notification orchestrator that publishes events to a Kafka cluster. Separate delivery services consume these events: EmailService, SmsService, and PushService, each with idempotent consumers and retry queues. For eventual consistency, each service writes to its own durable store and publishes a status event; a monitoring service aggregates these to provide a global delivery view. Load balancing across services and a CDN for push notifications ensure low latency. Observability is achieved with distributed tracing (Jaeger), metrics (Prometheus), and alerting (Grafana). This design satisfies scalability, fault tolerance, and low latency while adhering to the CAP theorem’s trade‑offs.