System Design Basics

# System Design Basics

## Scale up vs out
- **Vertical** (bigger box): simple, limited, single point of failure.
- **Horizontal** (more boxes): requires stateless services + load balancer + shared state.

## Load balancing
- L4 (TCP) vs L7 (HTTP, content-aware routing).
- Algorithms: round-robin, least connections, consistent hashing (for sticky cache nodes).

## Caching layers
- Client (browser, ETag, Cache-Control)
- CDN (static + edge cache)
- Reverse proxy (Varnish, Nginx)
- Application (Redis/Memcached)
- DB query cache
Patterns: cache-aside, write-through, write-back.

## Database scaling
- **Read replicas**: scale reads, eventual consistency.
- **Sharding**: by user ID hash, range, or geo. Resharding is hard -> use consistent hashing.
- **CQRS**: separate read model from write model.

## Async via queues
- Smooth bursty traffic, decouple producers/consumers.
- At-least-once delivery is the norm -> consumers must be idempotent.
- Kafka (log, replayable, partitioned), RabbitMQ (smart broker), SQS (managed).

## CAP / PACELC
In a partition you pick Consistency or Availability. Even without partition you trade Latency vs Consistency.
- CP: HBase, ZooKeeper, etcd
- AP: Cassandra, DynamoDB (tunable), Couchbase

## Consistency models
Strong > Linearizable > Sequential > Causal > Read-your-writes > Eventual.

## Availability math
- 99.9% -> 8.76h/year downtime
- 99.99% -> 52.6m/year
- Each "nine" ~10x harder.

## Building blocks checklist
LB | API GW | Auth | App tier | Cache | DB primary+replicas | Queue | Object storage | CDN | Search index | Metrics/logs/traces