Functional requirements (FR)

• Add functional requirements here.

Non-functional requirements (NFR)

• Add non‑functional requirements here.

The prompt

Design a distributed, leaderless key-value storage engine modeled on Amazon Dynamo and Apache Cassandra. The system provides high-throughput point lookups (get and put) across hundreds of commodity server nodes, handles hardware and network partition failures gracefully, and allows client applications to tune consistency levels according to their business needs.

Functional requirements

• put(key, value, [ttl]): Store or update an arbitrary byte or string payload for a given key.
• get(key): Retrieve the stored value associated with a given key.
• delete(key): Mark a key as deleted using a tombstone.
• Support tunable read and write consistency levels (ONE, QUORUM, ALL).
• Support configurable data replication factors across independent availability zones and regions.
• Provide deterministic conflict resolution using Last-Write-Wins (LWW) or Vector Clocks.

Non-functional requirements

• Scale to 10 Billion records and 100 TB of data across 100 nodes.
• Read throughput: 1,000,000 QPS with $p99 \le 5\text{ ms}$; Write throughput: 200,000 QPS with $p99 \le 3\text{ ms}$.
• High availability target: 99.999% uptime adhering to an AP (Availability & Partition Tolerance) architecture.
• Cluster supports elastic node addition and removal with minimal data migration and zero downtime.
• Background anti-entropy repair processes reconcile divergent replica data without taking nodes offline.
• Storage engine maximizes sequential disk I/O efficiency using Log-Structured Merge (LSM) trees and Bloom filters.
• Zero single point of failure (SPOF); all coordination is distributed via peer-to-peer gossip protocols.

Discussion prompts

• How does consistent hashing with virtual nodes prevent hot-spot partitions and smooth out rebalancing?
• What are the write-path dynamics: Write-Ahead Log (WAL), MemTable, SSTables, and compaction?
• How do Bloom filters optimize read paths by eliminating unnecessary disk seeks on missing keys?
• What are the mathematical trade-offs of quorum configurations ($W + R > N$) vs. latency and availability?
• What is the difference between Hinted Handoff for transient node downtime and Merkle tree anti-entropy for permanent repair?