موازنة التحميل الموزونة والخوادم الاحتياطية

Real upstream pools often mix machines of different capacity. A new 16-core server should not get the same share as an old 4-core box.

Nginx lets you bias traffic with weights so stronger servers handle more requests.

Add weight=N to a server. The default weight is 1. A server with weight=3 receives roughly three times as many requests as a weight=1 server.

With weights 3 and 1, out of every 4 requests, 3 go to the first server and 1 to the second. Weights are relative, so 6 and 2 produce the same split as 3 and 1.

Weights work with round robin (the default) and least connections. They do not apply to ip_hash, which selects a server purely from the client IP.

A server marked backup receives no traffic while any primary server is available. It activates only when all primaries are down or unreachable.

Backups give you a graceful fallback, for example a maintenance page server or a secondary data center, without sending it normal traffic during healthy operation.

Use down to permanently remove a server from rotation, handy during planned maintenance without deleting the line.

max_fails sets how many failed attempts mark a server unavailable, and fail_timeout sets the window and how long it stays out.

You can mix weights for primaries and keep a single backup for the whole pool. The backup ignores weight until it is the only option.

The max_conns parameter caps simultaneous connections to a server, protecting a weaker box even if its weight is high.

Fire many requests and check your access logs per upstream to confirm the ratio matches the weights. Adjust weights based on observed CPU and latency.

A server is marked backup. When does it receive requests?

You learned to fine-tune traffic distribution:

• weight=N biases traffic toward stronger servers
• Weights are relative and apply to round robin and least_conn
• backup servers activate only when primaries fail
• down, max_fails, and max_conns control availability

Together these give you precise control over an uneven server fleet.