Notes from https://www.hellointerview.com/learn/system-design/core-concepts/consistent-hashing

• Outgrown single database
• need to split data across multiple independent servers
• Happens when hit storage limits
  • A single Postgres instance maxes out in the TB
  • Write throughput limits
    • Tens of thousands of writes per second
  • Read throughput that even replicas can’t handle
• Most important decision is shard key
  • Determines how data gets distributed
  • For user-centric apps
    • Instagram
    • Sharding by user_id means all of a user’s posts, likes, and comments live on one shard
    • User-scoped queries are fast because they only hit one shard
    • Global queries like “trending posts across all users” become expensive
      • Have to hit every shard and aggregate results
• Most systems use hash-basd sharding
  • Hash the shard key and use modulo to pick a shard
  • Distributes data evenly and avoids hot spots
• Range-based sharing
  • Can work if access patterns naturally partition
    • Like multi-tenant SaaS where each company only queries their own data
  • Easy to create hot spots if one range gets more traffic
• Directory-based sharding
  • Uses a lookup table to decide where data lives
  • Flexible
  • Adds dependency and latency to every request
  • Rarely worth it in interview
• Problems
  • Cross-shard transactions become nearly impossible
  • Design shard boundaries to avoid them
    • If user transfer in banking app requires updating accounts on different shards, need to distribute transactions or sagas
      • Comples and slow
  • Hot spots
    • Happen when one shard gets disproportionate traffic
  • Resharding is painful
    • Can’t just add a new shard without moving massive amounts of data around
• Mistakes
  • Sharding too early
  • Well-tuned single database with read replicas can handle way more than one might think
  • If proposing sharding, do capacity math
    • 10k write per second and 100GB data, don’t need sharding yet
    • Suggest if numbers justify
• Interviews
  • Only bring up with justification of why single database won’t work
  • Clearly state shard key choice
  • Explain tradeoff
    • Fast for X queries, slow for Y queries

Notes from https://www.hellointerview.com/learn/system-design/core-concepts/sharding