- For each shard/partition, having a single copy has a few disadvantages
- A single instance can be easily overwhelmed by a large number of requests
- A single instance can be close to some users but far from others, who will experience greater latencies.
- A single instance provides low overall availability
- The improvement is to add copies/replicas, in different availability zones
- need to consider replicas consistent(accuracy, latency, availability)
- Example DynamoDB, only two levels
- Strong Consistency - A read is guaranteed to return the latest committed write
- Accuracy
- increase latency
- Reduce availability during failure
- Eventual Consistency - Reads may temporarily return old data right after a write.
- Latency
- Availability: keeps working even if some replicas are behind