ddia-systems

Match the data model to your access patterns: relational excels at many-to-many relationships and ad-hoc queries, document models fit self-contained aggregates with one-to-many data, and graph models suit recursive relationship traversals. Evaluate read/write ratio, consistency needs, and scaling path rather than popularity.

LSM trees optimize write throughput using sequential appends and background compaction, suiting write-heavy workloads like time-series ingestion. B-trees update pages in place with predictable read latency, making them better for mixed OLTP workloads. LSM trees have higher write amplification; B-trees suffer fragmentation.

Asynchronous replication creates lag between the leader and followers, causing read-your-writes and monotonic-read violations. Fix this by reading recently written data from the leader, pinning users to consistent followers, or tracking replication log positions to ensure followers have caught up.

Hot keys occur when one key receives disproportionate traffic, which hashing alone cannot fix. Mitigate by splitting the hot key with random suffixes and fanning out reads, adding a caching layer, buffering writes, or using databases with automatic hotspot detection like DynamoDB adaptive capacity.

Use sagas for operations spanning multiple services or partitions, since two-phase commit is slow, fragile, and makes the coordinator a single point of failure. Design single-partition transactions where possible and use compensating actions, like refunds, to roll back multi-step operations.

Most databases default to read committed or snapshot isolation, not serializable, which permits anomalies like write skew. Check your database documentation, test for write skew and phantom reads, and use explicit locking such as SELECT FOR UPDATE where stronger guarantees are needed.