In modern database capacity planning, when should you consider sharding?

In modern database capacity planning, when should you consider sharding?

• A. Shard only when read workloads grow heavy.
• B. Shard as soon as data grows beyond 5 TiB.
• C. Shard only when write throughput consistently exceeds 10k TPS.
• D. Only consider sharding when write throughput consistently exceeds 10k TPS, data approaches 50 TiB, or geographic distribution is required.

Answer: (D)

The main idea being tested is when to scale a database horizontally by sharding. Sharding is most justified when the workload and data volume push a single node beyond its comfort zone, especially for writes and for latency-related concerns. Sharding should be considered when write throughput is consistently high, because writes are often the bottleneck that limits capacity and performance. If the system begins to struggle delivering low-latency writes, spreading the data across multiple nodes can relieve contention and improve throughput. It’s also a practical signal when the dataset grows to a size where managing it on one machine becomes unwieldy or expensive; approaching tens of terabytes (for example, around 50 TiB) can make single-node performance and maintenance troublesome. Lastly, if your users are geographically dispersed or regulatory requirements demand data locality, distributing data across regions is a compelling reason to shard. Relying only on read workloads as a trigger can be misleading, since reads can often be handled with replicas or caching rather than sharding. A fixed arbitrary size like a small threshold (such as 5 TiB) isn’t generally reliable across systems or workloads. And opting to shard solely based on a write threshold without considering data size or distribution needs can leave you unprepared for cross-region access or future growth. By covering multiple practical triggers—write throughput bottlenecks, large data volume, and geographic distribution needs—the approach in the answer reflects the real-world factors that typically prompt shard design and deployment.

The main idea being tested is when to scale a database horizontally by sharding. Sharding is most justified when the workload and data volume push a single node beyond its comfort zone, especially for writes and for latency-related concerns.

Sharding should be considered when write throughput is consistently high, because writes are often the bottleneck that limits capacity and performance. If the system begins to struggle delivering low-latency writes, spreading the data across multiple nodes can relieve contention and improve throughput. It’s also a practical signal when the dataset grows to a size where managing it on one machine becomes unwieldy or expensive; approaching tens of terabytes (for example, around 50 TiB) can make single-node performance and maintenance troublesome. Lastly, if your users are geographically dispersed or regulatory requirements demand data locality, distributing data across regions is a compelling reason to shard.

Relying only on read workloads as a trigger can be misleading, since reads can often be handled with replicas or caching rather than sharding. A fixed arbitrary size like a small threshold (such as 5 TiB) isn’t generally reliable across systems or workloads. And opting to shard solely based on a write threshold without considering data size or distribution needs can leave you unprepared for cross-region access or future growth.

By covering multiple practical triggers—write throughput bottlenecks, large data volume, and geographic distribution needs—the approach in the answer reflects the real-world factors that typically prompt shard design and deployment.