System Design

Vocabulary dump (the words that vanish mid-interview)#

1. idempotent — same effect when repeated; backpressure — downstream slows upstream instead of dropping; fan-out — one event delivered to many consumers/feeds.
2. hot partition — one shard/key takes disproportionate load; thundering herd — many clients wake/retry at once (cache miss storm); split brain — two nodes both think they are leader.
3. quorum — majority (⌊N/2⌋+1) or R+W>N overlap; tail latency — p99/p99.9, what users actually feel; head-of-line blocking — one slow item blocks everything queued behind it.
4. circuit breaker — stop calling a failing dependency, probe later; bulkhead — isolate pools so one failure can't drain all resources; sidecar — per-service helper proxy/process (Envoy); service mesh — sidecars + control plane (Istio, Linkerd) for mTLS/retries/telemetry.
5. saga — chain of local transactions with compensations; outbox — write event in the same DB transaction, relay publishes it; CQRS — separate write model and read model(s).
6. materialized view — precomputed query result stored as a table; denormalization — duplicate data to avoid joins; write amplification — one logical write causes many physical writes (indexes, LSM compaction); read amplification — one logical read touches many structures (SSTables).
7. cold start — latency penalty of a fresh instance/function/cache; warm cache — pre-populated before traffic; stale-while-revalidate — serve stale now, refresh in background.
8. exponential backoff with jitter — retry delays double plus randomness to avoid synchronized retry storms; leader election — pick one coordinator via consensus (Raft/Paxos/Zab in etcd/Zookeeper) or lease.
9. gossip protocol — nodes periodically exchange state with random peers; membership/failure detection converges in O(log N) rounds; vector clock — per-node counters to detect concurrent vs causal versions; CRDT — data type that merges concurrent updates deterministically (counters, sets, text).
10. Bloom filter — probabilistic "definitely not / probably yes" set membership, no false negatives; HyperLogLog — count distinct in ~12 KB with ~1% error (Redis PFCOUNT); Merkle tree — hash tree to compare replicas and sync only differing ranges (Cassandra repair, Git, Dynamo).