Inter-Service Communication Lab (Interactive)
Run a request through REST or gRPC hops, then flip to Kafka and compare latency, availability, and propagation. Synchronous chains sum every hop latency and multiply availability; asynchronous chains return after a broker ack and let consumers proceed independently.
Sync RPC Waterfall vs Async Event Stream
Same checkout, four collaborators. Switch paradigms and see latency summation and availability multiplication change.
Client-visible path (blocking waterfall)
T_total = Σ Tᵢ = 92ms and A = 99.9%^4. Every hop is a chance the whole user transaction dies.
Eventual-consistency trade
Production rule: gRPC synchronously for reads that need an immediate answer (JWT validation, fraud score), Kafka asynchronously for mutations and side effects (OrderCreated events). The hybrid checkout acks in ~30ms while four consumers reconcile eventually.
How It Works Under the Hood
Synchronous HTTP/gRPC couples callers temporally: total latency is the sum of hops, availability is the product of hops, and every failure path needs timeout, retry, and circuit-breaker logic. Asynchronous messaging through Kafka converts the chain into a relay race: the producer pays only for a durable broker acknowledgment, often under 10 ms, and slow consumers buffer behind it instead of failing the user request. The price is eventual consistency, event-schema versioning, and consumers that must be idempotent because at-least-once delivery duplicates messages.
Core Architectural Principles
- Sync mode sums per-hop latency across the chain; async mode charges only broker ack plus fan-out.
- Availability composes as A^N for sync calls but stays A^1 for the produce-and-ack path.
- REST at 25 ms and gRPC at 3 ms per hop change the crossover where async clearly wins.
For "when async?" answer with three triggers: work can happen later, you need burst absorption, or you want fan-out to many consumers. Then show the trade side: sync is right when the user is blocked on the answer and you need immediate validation. Name the async responsibilities you take on: idempotent consumers and a versioned event schema registry.
Async buys latency isolation and fan-out at the cost of eventual consistency; sync gives an immediate answer with multiplied failure coupling.