Episode 6. Caching, document, and graph. Day 6. Three engines, one pattern: pick the right data model.

Key facts. ElastiCache runs Valkey, Memcached, or Redis OSS: microsecond caching, pub sub, and session state. MemoryDB adds durability with Multi AZ persistence. DocumentDB is MongoDB compatible document storage. Neptune is a fully managed graph engine for Gremlin, openCypher, and SPARQL. Say: Cache, document, and graph are not interchangeable. The access pattern chooses the engine.

Scenario 1. Our cache hit rate collapsed after a deploy and the database is now overloaded. Help.

Understand. Ask: what changed in keys, TTLs, or eviction policy? Hit rate before and after, exact numbers. Is it one hot dataset or everything? Database load shape: reads, connections, or both?

Prepare. Say the data in order. ElastiCache hit rate, evictions, curr connections, and replication lag. Key inventory: TTL distribution and key cardinality. Deploy diff on serialization or key prefixes. Database Performance Insights showing which queries surged.

Say your translation line: Same instinct as buffer cache hit collapse: the working set moved or the keys changed shape.

Identify. Rank the hypotheses. One: key prefix or serialization change invalidating the working set. Two: TTL shortened or removed, churning hot keys. Three: a thundering herd on cold start with no request coalescing. Four: node loss or failover dropping the dataset. Fix: restore key format, stage TTL changes, add jittered TTLs and stampede protection, warm the cache before cutover.

Support. Success: hit rate back above target for seven days, database load at baseline, and alarms on hit rate plus evictions.

Scenario 2: DocumentDB versus DynamoDB versus Neptune. Say: I'd choose DynamoDB for key value scale, DocumentDB for document queries with flexible schema, and Neptune for traversals deeper than two hops. Then commit. Say: my recommendation is.

Your Day 6 lab: stand up an ElastiCache cluster, load a working set, change the key prefix, watch hit rate fall, then restore it. Delete the cluster the same day.

Tomorrow, Episode 7: networking. VPC, endpoints, and PrivateLink. The silent killer of database cases. Understand. Prepare. Identify. Support.