INTERNET FUNDAMENTALSAI-REVIEWED

What is load balancing?

Useful first: HTTP · optional, not a requirement

Choose your depth

Stay with the essentials, or look closer.

Simple explanation
2 min read · estimate

YOUR GOALUnderstand why one service can use several servers.

Share incoming work

One service, several places to work

Imagine a help desk with two assistants. A coordinator sends each new question to one of them instead of making everyone wait for the same person. A load balancer plays a similar routing role for a group of servers.

For example, a website could run the same application on servers A and B. A load balancer chooses where to send an incoming request. The selected server does the application work; the balancer does not need to send that request to every server.

[1]

Equal turns can mean unequal work

In a simplified round-robin example, ordinary page requests take turns: A, B, A, B. Assume both servers are available, have equal weights, and no rule keeps a visitor on one server.

One request might be quick while another takes much longer. Sharing the number of requests therefore does not guarantee equal effort or instant responses. The help-desk analogy explains distribution, not a promise about speed.

[2]
THE IDEA TO KEEP

A load balancer shares incoming work across a server group.

ONE THING TO WATCH OUT FOR

A load balancer makes every server equally busy.

The routing policy distributes work using particular signals. Equal request counts can still produce unequal workloads.

A QUICK CHECK, NOT A TEST
A website has two application servers behind a load balancer. What is the balancer normally choosing for a new request?
Show the explanation

Which server should handle the request. Load balancing selects a destination for work. It does not by itself replicate application data or guarantee response speed.

Progress is saved only in this browser.

KEEP EXPLORINGHTTP
Was this explanation clear?
Report a mistake by email
Basic explanation
3 min read · estimate

YOUR GOALCompare routing policies and recognize the limits of health checks.

Choose a server and check its health

Round robin and weights

Round robin shares requests by taking turns. NGINX supports weights that adjust the relative shares. In a simplified stable pool with weights 3 for A and 1 for B, A is intended to receive roughly three requests for each one sent to B.

These are configured shares, not a measurement proving A is three times faster. Failures and session-persistence rules can change the observed distribution.

[2] [3]

Least connections measures connections

NGINX also offers least connections, choosing with active connection counts and server weights. With equal weights, two active connections on A and none on B can favor B for the next request.

This can help when connections last different amounts of time. It does not directly measure CPU usage or how expensive the next operation will be.

[3]

Health checks are evidence with limits

Active checks send periodic probes. AWS Application Load Balancer, for example, checks a configured path and uses response criteria and failure thresholds. Passive checks instead learn from real traffic failures; NGINX can temporarily avoid a server after configured unsuccessful attempts.

A hypothetical /health response can succeed while a different operation fails. Detection also takes time. Do not assume every product always excludes every unhealthy server: AWS ALB routes to its unhealthy targets if all registered targets in the group are unhealthy.

[3] [4]
THE IDEA TO KEEP

Match the routing signal and health check to the work you need to serve.

ONE THING TO WATCH OUT FOR

Passing a health check proves every feature of the application works.

It proves only that the configured check passed. Other operations can fail, and routing behavior depends on the product and settings.

A QUICK CHECK, NOT A TEST
A health check for /health passes, but readers report that /search fails. What follows?
Show the explanation

The health check may not exercise the failing search operation. A probe tests its configured operation and success criteria. Passing one endpoint is not proof that a different endpoint works.

Progress is saved only in this browser.

KEEP EXPLORINGHTTP
Was this explanation clear?
Report a mistake by email
Advanced explanation
3 min read · estimate

YOUR GOALSeparate routing choices from session storage and safe retry behavior.

Connections, state, and recovery

A request is not always the balancing unit

Check what the balancer selects. NGINX HTTP upstreams distribute requests; its stream upstreams distribute connections. These are different scopes.

For a concrete application example, AWS ALB keeps an established WebSocket connection with its selected target. Each later message is not a fresh opportunity to choose another server. A failed connection needs application recovery; routing alone does not move its conversation history.

[2] [5] [6]

Affinity is not a copy of session data

Session affinity, or sticky sessions, tries to keep a client with one server. AWS ALB can use a cookie to remember that choice, but may select another target if the previous one becomes unhealthy or is removed.

Suppose a basket exists only in A's memory. Sending the next request to B does not give B that basket. An application could instead store the basket where both servers can retrieve it; that shared storage then has its own availability requirements. The routing cookie alone does not provide that storage.

[5] [8]

A missing response does not prove nothing happened

Imagine A accepts an order but its response is lost. Repeating the order at B could create a second order. Finding another server is therefore different from deciding that a retry is safe.

HTTP calls an operation idempotent when repeating it has the same intended server effect as doing it once. RFC 9110 prohibits proxies from automatically retrying non-idempotent requests. Applications need recovery rules that account for whether the original operation may already have happened.

[7]
THE IDEA TO KEEP

Design connection recovery, session storage and retry safety separately from server selection.

ONE THING TO WATCH OUT FOR

Sticky sessions preserve my in-memory data when a server fails.

Affinity controls routing. A different server does not automatically receive the failed server's local application state.

A QUICK CHECK, NOT A TEST
A visitor's basket exists only in server A's memory. A fails and the balancer sends the visitor to B. What is justified?
Show the explanation

Routing to B alone does not restore the basket; the application needs a way to recover its state. Choosing another destination does not copy A's memory. The application must define where state lives and how it survives a server failure.

Progress is saved only in this browser.

KEEP EXPLORINGHTTP
Was this explanation clear?
Report a mistake by email
GO STRAIGHT TO THE SOURCE

Keep asking. Check the details.

Source-checked using AI (OpenAI Codex) on . The checks covered sources, examples, and quizzes for this article version.

No human review is claimed. AI review can miss errors; check the original sources below.

  1. AWS — What is Elastic Load Balancing?
  2. NGINX — HTTP load-balancing methods and session persistence
  3. NGINX — HTTP upstream weights, least_conn and failure accounting
  4. AWS — Application Load Balancer health checks and fail-open behavior
  5. AWS — Application Load Balancer sticky sessions and connection behavior
  6. NGINX — Stream upstream connection balancing
  7. RFC 9110 — HTTP idempotency and automatic retries, section 9.2.2
  8. AWS Well-Architected — Stateless systems and offloading session data
How explanations are prepared