Finding Network Bottlenecks in Production

"The network is slow" is a symptom report. It does not identify the constrained resource, the delayed stage, or even whether the network caused the delay.

A request can wait for DNS, a connection pool, a TCP handshake, TLS, a proxy queue, an application worker, a database, a congested link, a receiver, or a retry. Several of those delays appear to the client as time spent waiting on the network.

A bottleneck is the resource or serialized dependency that limits the performance objective under the current workload. Removing it should produce a measurable change in latency, throughput, or capacity. High utilization alone does not establish that a resource is the bottleneck.

Finding the bottleneck requires a disciplined sequence. Define the affected operation, decompose its end-to-end performance, localize the first abnormal boundary, inspect resources and queues there, and test one causal hypothesis. The investigation ends when the original workload shows the predicted improvement without unacceptable downstream effects.