Streaming, partitioning, consumer-group debugging. Open-ended prompts; your written answer gets senior-reviewer feedback on the coached tier.
Your team runs a Kafka consumer group enriching payment events with fraud signals — about 2M events/day across 12 partitions and 6 consumer instances. Last sprint someone added a new ML model that takes ~80ms per event (up from ~5ms). Consumer- group lag is now spiking to 30+ minutes during peak. The on-call DE just paged you to figure it out.
You're designing the Kafka topic strategy for a new multi-tenant analytics product. ~200 tenants, traffic ranges from 1k events/day for the smallest to 5M events/day for the biggest. Each tenant's data must stay isolated for compliance. Two senior engineers are arguing in the design doc — one wants one topic per tenant, the other wants a single topic partitioned by `tenant_id`. Both are bad ideas in different ways. What's your actual design?
Your team is building a payments-events pipeline through Kafka. Money is involved. The engineering manager says "we need exactly-once delivery — anything less is unacceptable." The tech lead says "at-least-once with idempotent consumers is the actual production move — Kafka's 'exactly-once' is more or less a marketing term." Both have a point. You've been brought in to make the call and write up the design.
Production Kafka cluster, AVRO with Confluent Schema Registry. The product team needs to add a required field settlement_currency to the order_placed event, retroactively required across all consumers. The problem: some consumers are stale-but-running, some are mid-deploy, some won't deploy for 2 weeks (frozen for a release). Marketing wants this field live by next Tuesday. You can't take downtime — the cluster handles fraud detection 24/7. Design the rollout.
3 AM page. Your fraud-detection consumer group is lagging — but only on ONE consumer instance. The other 11 are idle. You have 12 partitions, 12 consumers. Lag on partition 7 is climbing fast (3M events behind and growing); other partitions are at 0. The CEO of the affected merchant is on Twitter complaining about transaction delays. You have ~20 minutes before this becomes a PR incident. Diagnose, then walk through what you'd actually do.