Horizontal scaling adds more instances or nodes instead of making one machine bigger, so work can be shared across multiple system units.
Horizontal scaling, also called scaling out, increases capacity by adding more instances, machines, or nodes that can share the workload.
A request may be routed to any healthy application instance, while a cluster or service coordinates the collection of machines. The system grows by adding units rather than by upgrading only one unit.
Core idea: use more cooperating machines to handle more work.
Think of a busy shop opening more locations instead of making one shop endlessly larger.
A shared brand and operating process can serve more customers across several locations. The benefit is more total capacity and a smaller blast radius when one location has trouble, but the locations must coordinate inventory, staffing, and customer records.
Horizontal scaling lets a system add parallel capacity when one instance is no longer enough.
Run more copies of a stateless application so requests can be processed in parallel.
Use a load balancer or routing layer to send work to healthy instances.
A failure can remove one instance while other instances continue serving traffic.
Add or remove units as demand changes instead of resizing one large machine.
More instances help only when the application, data, network, and dependencies can work with a distributed shape.
Horizontal scaling is common when demand varies, availability matters, or a workload can be split across independent workers.
Run multiple frontend or backend instances behind a load balancer during normal traffic and peaks.
Add queue consumers so image processing, notifications, or exports finish faster.
Scale the busiest service without giving every service a larger machine.
Use automated instance groups or containers to match capacity to a changing workload.
A meal-delivery app expects a large dinner-time spike. One application server is reaching its request and CPU limits, even though the database still has headroom.
Ask what work can safely be shared, whether users can land on any instance, and which state must be shared or reconstructed. Track p95 latency, error rate, successful orders, instance count, database saturation, and infrastructure cost. More instances improve capacity only if the next dependency is not already the limit.
The product gains parallel request capacity, but it also takes on routing, shared-state, deployment, and observability responsibilities.
Horizontal scaling changes the number of workers in the serving layer; it does not automatically scale the database, cache, queue, or external provider.
Add more instances or nodes. It can increase parallel capacity and reduce dependence on one machine, but it requires routing and coordination.
Make one machine bigger. It is often simpler for a single component, but it has a hardware ceiling and can remain a single point of failure.
Distribute requests across available instances. It helps horizontal scaling work, but it is a routing mechanism—not the whole scaling strategy.
Automatically add or remove capacity from rules or signals. Auto scaling can use horizontal or vertical changes; it is the policy, not the direction.
Keep service usable when components fail. Multiple instances can support availability, but only if health checks, dependencies, and recovery paths are designed too.
Make code, queries, or architecture more efficient. Optimization can delay the need to scale, but it is different from adding capacity.
Horizontal scaling answers “how many cooperating units?” It does not answer “how should state, failure, or data consistency work across them?”
| ENGLISH TERM | 中文 | SIMPLE ENGLISH | 中文小白理解 |
|---|---|---|---|
| Horizontal Scaling | 水平扩展 | Adding more machines or instances to share work. | 增加更多机器或实例，一起分担工作。 |
| Scaling Out | 向外扩展 | Another name for horizontal scaling. | 水平扩展的另一种说法。 |
| Instance | 实例 | One running copy of an application or service. | 正在运行的一份应用或服务副本。 |
| Node | 节点 | A machine or worker unit in a system. | 系统里的一台机器或一个工作单元。 |
| Load Balancer | 负载均衡器 | A component that routes traffic across healthy instances. | 把请求分配给健康实例的组件。 |
| Stateless | 无状态 | An instance does not need local memory of a user's previous request. | 实例不必把用户之前的请求记在自己本地。 |
| Shared State | 共享状态 | Information that multiple instances can read and update. | 多个实例都能读取和更新的信息。 |
| Health Check | 健康检查 | A test used to decide whether an instance can receive work. | 用来判断实例是否还能接收工作的测试。 |
| Blast Radius | 影响范围 | The amount of a system affected by one failure. | 一次故障会影响到系统多大范围。 |
| Auto Scaling | 自动扩缩容 | Automatically changing capacity from rules or signals. | 根据规则或指标自动调整资源数量或大小。 |
Horizontal scaling adds more instances or nodes instead of making one machine bigger.
It is useful for parallel web requests, background workers, and workloads that can be divided across independent units.
It can improve capacity and reduce reliance on one machine, but it introduces routing, shared-state, consistency, deployment, and dependency trade-offs.