Single-leader avoids conflicts, multi-leader buys local writes at the cost of write-write conflicts, leaderless uses R+W>N quorums; resolve conflicts losslessly and reason with PACELC.

Map each lag-induced bug to its session guarantee and fix it with sticky routing or version tokens, instead of over-promising linearizability.

Range wins range scans but hotspots on sequential keys, hash spreads evenly but loses ranges, directory adds a flexible routing hop; map every dominant query to its partitions.

Hash mod N remaps nearly everything on resize; a ring with virtual nodes moves only ~1/N of keys, smooths load, and spreads rebalancing across many neighbors.

A good shard key is high-cardinality, evenly accessed, and query-aligned; a celebrity is one key on one node, so split the key, dedicate shards, or change the read pattern.

Avoid 2PC on the hot path (locks, blocks on coordinator failure); use sagas of local transactions with compensations, idempotency keys, and the outbox pattern.

Default to cache-aside reads plus invalidate-on-write, pick the write policy by its durability/latency trade, and always state the TTL and consistency window.

Layer the defenses: singleflight coalescing with a distributed lock, jittered TTLs and probabilistic early refresh, plus L1 caches and key replication for genuinely hot keys.

Shard by hash slots, replicate each shard with failover, tier L1-near plus L2-remote, treat the cache as disposable, and never bring a cold cache online under load.

An L1/L2/origin-shield hierarchy coalesces misses to ~1 fetch per object; version URLs instead of purging, normalize cache keys, and never cache authenticated bodies.

A dedicated search tier: inverted index plus analysis pipeline, BM25 with boosting, cached filters, shards and replicas, CDC-fed eventual consistency, search_after pagination.

Embeddings plus ANN give semantic recall, BM25 catches exact tokens; fuse by rank with RRF, rerank the top-k with a cross-encoder, and plan re-embedding migrations.

Encode points into cell keys (geohash/S2/H3) so 2D proximity becomes an indexable range query, ring-query neighbors, tune cell size to radius, and defuse hot cells.

Precompute rollups and sketches so reads are O(1) lookups, and feed timelines with hybrid fan-out: push for normal users, pull-and-merge for celebrities.

Never dual-write to a DB and a derived store: use the transactional outbox or log-based CDC through Kafka, with at-least-once delivery and idempotent, versioned consumers.