System Design Interviews

Resources

• AusDevs Wiki: Interviewing Resources
  • I actually help maintain this wiki. We aim to make it a great meta-resource!
• GitHub: donnemartin / system-design-primer
  • I’m largely studying this document, so my notes will retain a lot of the structure from it.

Distributed Systems Concepts

• Performance vs. Scalability
  • Performance: “How it is normally.”
  • Scalability: “It’s slow under heavy load.”
  • Scalability: “Adding resources improves performance linearly.”
• Latency vs. Throughput
• CAP Theorem (for distributed systems)
  • CAP Theorem: Can only guarantee two properties of:
    • Consistency: Every read receives the most recent write or an error.
    • Availability: Every request receives a non-error response. (Not necessarily the most recent.)
    • Partition Tolerance: The system continues to operate even with arbitrary network failures/delays.
  • Choices of guarantees:
    • CA: Consistency + Availability
      • Networks are never reliable. Almost never choose CA.
    • CP: Consistency + Partition Tolerance
      • Good if you need atomic reads/writes.
      • TODO: Atomic in what way?
    • AP: Availability + Partition Tolerance
      • Good if you need reliable access to the system.
      • TODO: It also talks about a business needing eventual consistency. Why would a business need it specifically?
• Consistency Patterns
  • Weak Consistency: After a write, reads may or may not see it.
    • Example: real-time use cases like VoIP and multiplayer games.
  • Eventual Consistency: After a write, reads will eventually see it. (Typically milliseconds.)
    • Useful if you need high availability.
    • Example: DNS and email.
  • Strong consistency: After a write, reads will always see it.
    • Useful if you need transactions.
    • TODO: What is a transaction?
    • Example: File systems and RDBMSes.
• Availability Patterns
  • Active-Passive Failover:
    • Active server normally manages traffic.
    • Only one “IP address” is needed.
      • TODO: Be more clear that “IP address” doesn’t necessarily mean a literal IP address?
    • If active server heartbeats stop, passive server takes over the “IP address”.
  • Active-Active Failover:
    • Load is spread between both servers.
    • Requires “two IP addresses”.
      • TODO: Be more clear that this doesn’t necessarily mean literal IP addresses?
  • Master-Slave Replication:
    • One master, many slaves.
    • Master server can read/write. Slave servers can only read.
    • If master fails, a slave can be promoted to master.
  • Master-Master Replication:
    • All servers are masters and can all read/write.
  • TODO: Talk more about the advantages/disadvantages of each approach?
• Availability also depends on whether components are in series or parallel.
  • 99.9% availability in series with 99.9% = 99.8%
  • 99.9% availability in parallel with 99.9% = 99.9999%