AWS DVA-C02: Refactoring Practice Questions

Amazon MSK runs Apache Kafka as a managed service, so existing Kafka producer/consumer code works unchanged, and Amazon OpenSearch Service is the managed successor to self-managed Elasticsearch. Kinesis Data Streams requires rewriting all Kafka client code, and CloudWatch Logs Insights is not a drop-in replacement for an Elasticsearch search cluster. SQS FIFO has different semantics (no replayable log, no consumer groups) and Athena queries S3, not a live search index. Amazon MQ supports ActiveMQ/RabbitMQ protocols, not Kafka, and Redshift is a data warehouse, not a log search engine.

A cache-aside Redis cache with a 5-minute TTL serves repeated identical queries from memory, read replicas horizontally scale read traffic away from the writer, and API Gateway response caching returns full responses without touching the backend at all — all three fit the 5-minute staleness tolerance. A Multi-AZ standby in RDS (non-cluster) does not serve read traffic; it is only for failover. Adding storage increases capacity and baseline IOPS for gp2 but does nothing about thousands of identical repeated queries and is the least targeted option.

EventBridge rules match on event content (detail-type, detail fields) and route only matching events to each target; rules are managed independently of the producer, which simply puts events on the bus. Per-type SQS queues push routing logic into the producer, violating the decoupling requirement. Kinesis shards are a scaling construct, not a routing construct — the partition key does not guarantee type isolation and every consumer reads the whole stream. An SNS topic with client-side filtering means every consumer receives every message (SNS message filtering policies would help, but plain client-side filtering does not and wastes delivery).

Bounded timeouts with a limited retry policy make calls fail fast instead of blocking threads indefinitely, and an ALB with health checks automatically stops routing to unhealthy targets, which is the managed equivalent of tripping a circuit. More CPU and memory only lets the service block on more threads before failing. Removing timeouts is the opposite of failing fast and causes cascading exhaustion. CPU-based auto scaling scales out a service whose threads are idle-blocked on I/O, which neither detects nor isolates the failing dependency.

RDS Proxy pools and multiplexes database connections, so thousands of concurrent Lambda environments share a small pool. Reserved concurrency caps how many concurrent executions can exist, bounding the number of connections. Opening and closing a connection per invocation increases connection churn and still scales linearly with concurrency (it also wastes the container reuse optimization). Raising the timeout keeps connections open longer, worsening the problem. Removing the function from the VPC does not reduce connection count and exposes the database publicly.