Diagnose the exact anomaly (dirty read, lost update, write skew) and fix it surgically; snapshot isolation still allows write skew, which only Serializable or a lock prevents.

MVCC versions rows so readers and writers never block each other (at vacuum/bloat cost); go optimistic under low contention, pessimistic under high, and shard hot counters.

B-trees update pages in place for fast reads and range scans; LSM-trees append and compact for write throughput, paying read amplification and compaction stalls.

Order composite indexes equality-then-sort per the leftmost-prefix rule, make hot queries covering, and remember every index taxes every write.

Writes land in dirty buffer-pool pages flushed lazily at checkpoints; the fsync'd sequential WAL is the durability point and the crash-recovery source.

Encode everything you query on into a namespaced key, guard against hot keys, TTL cache data, and never treat a non-persistent cache as a source of truth.

Embed bounded read-together data, reference large or unbounded entities, respect the 16MB cap, and treat per-document atomicity as a design constraint.

Model one denormalized table per query: a partition key that spreads load and co-locates the read, clustering for the sort, time-bucketed bounded partitions, tunable quorum.

Index-free adjacency keeps deep traversals local while recursive SQL joins explode; a 1-2 hop adjacency table in SQL is often the simpler right choice.

Exploit append-only time-ordered data with columnar compression, time-partitioning, and downsampling tiers, and design against cardinality explosion from unbounded tags.

ANN search trades recall for latency and memory: HNSW for recall in RAM, IVF-PQ at billion scale, plus filtered ANN, hybrid BM25 fusion, and re-embedding plans.

Normalize by default for write integrity; denormalize only a named hot read path with a real read/write ratio, and use materialized views as the managed middle ground.

Enumerate access patterns first, then design composite keys so each pattern is a single-partition lookup, with write sharding for hot partitions and GSIs added per named pattern.

Avoid monotonic-key hotspots and UUIDv4 fragmentation with ULID/UUIDv7, use surrogate keys with unique natural attributes, and let DB constraints and types carry correctness.

Bytes go to object storage with only the key and metadata in the DB, moved via presigned URLs and multipart upload, with lifecycle tiering and a CDN for cost and latency.

Reason from decision drivers to a storage family, default to boring relational, adopt NewSQL only past one node when SQL+ACID still matter, and justify every polyglot store.