Skip to content
Back to Blog

Website Is Slow but CPU and Memory Are Low: Where Should You Investigate First?

A practical guide to diagnosing slow websites with low CPU and memory by analysing TTFB, databases, pools, networks, storage, caches, external APIs, and frontend performance.

Website Is Slow but CPU and Memory Are Low: Where Should You Investigate First?

A website that responds slowly while server CPU and memory stay low is a common situation. It does not mean the system is healthy: the bottleneck may be the database, network, storage, connection pool, locks, an external API, DNS, TLS, or the frontend.

A common mistake is adding CPU or memory before knowing where requests are waiting. The right investigation breaks latency down from the browser to the backend, its dependencies, and the data layer.

Core principle

Low CPU and memory do not prove the system is healthy. Measure waiting time at each hop, queue, and dependency before scaling infrastructure.

Investigation

Is the slowness in the frontend or the backend?

Browser timing separates DNS, TCP, TLS, TTFB, download, and rendering.

If TTFB is high, the problem usually sits in the backend or a dependency. If TTFB is low but the page is still slow, investigate JavaScript, images, fonts, hydration, and main-thread work.

Illustration: Is the slowness in the frontend or the backend

Classification

Signals to record

Practical checks for identifying and confirming the real bottleneck.

High TTFB

Measure this signal with metrics, traces, logs, and a reproducible request.

Slow DNS

Measure this signal with metrics, traces, logs, and a reproducible request.

Slow TLS

Measure this signal with metrics, traces, logs, and a reproducible request.

Slow download

Measure this signal with metrics, traces, logs, and a reproducible request.

Long JavaScript execution

Measure this signal with metrics, traces, logs, and a reproducible request.

Region-specific slowness

Measure this signal with metrics, traces, logs, and a reproducible request.

Investigation

Requests can queue while the CPU is idle

Pools, worker limits, and locks create waiting time without using much CPU.

An application may spend most of its time waiting on the database, file storage, a queue, or an external API. When the connection pool or worker pool is small, requests queue while CPU stays low.

Backend checks

Points to check

Practical checks for identifying and confirming the real bottleneck.

Worker saturation

Measure this signal with metrics, traces, logs, and a reproducible request.

Connection pool

Measure this signal with metrics, traces, logs, and a reproducible request.

Event-loop lag

Measure this signal with metrics, traces, logs, and a reproducible request.

Lock contention

Measure this signal with metrics, traces, logs, and a reproducible request.

Request backlog

Measure this signal with metrics, traces, logs, and a reproducible request.

Retry

Measure this signal with metrics, traces, logs, and a reproducible request.

Investigation

The database is often the main bottleneck

Queries, locks, pools, and I/O can be slow even when database CPU is low.

Illustration: The database is often the main bottleneck

Database diagnostics

Priority checks

Practical checks for identifying and confirming the real bottleneck.

Slow query

Measure this signal with metrics, traces, logs, and a reproducible request.

Lock wait

Measure this signal with metrics, traces, logs, and a reproducible request.

Pool saturation

Measure this signal with metrics, traces, logs, and a reproducible request.

N+1 query

Measure this signal with metrics, traces, logs, and a reproducible request.

Missing index

Measure this signal with metrics, traces, logs, and a reproducible request.

I/O wait

Measure this signal with metrics, traces, logs, and a reproducible request.

Replica lag

Measure this signal with metrics, traces, logs, and a reproducible request.

Investigation

Slow I/O or network does not raise CPU

Disk latency, packet loss, and cross-region calls can stretch requests.

Check disk queue, read/write latency, IOPS, throughput, fsync time, packet loss, TCP connect, TLS handshake, and the number of hops through proxies or a service mesh.

Infrastructure checks

Common signs

Practical checks for identifying and confirming the real bottleneck.

High disk await

Measure this signal with metrics, traces, logs, and a reproducible request.

IOPS ceiling

Measure this signal with metrics, traces, logs, and a reproducible request.

Packet loss

Measure this signal with metrics, traces, logs, and a reproducible request.

Cross-region call

Measure this signal with metrics, traces, logs, and a reproducible request.

Proxy chain

Measure this signal with metrics, traces, logs, and a reproducible request.

Log pressure

Measure this signal with metrics, traces, logs, and a reproducible request.

Investigation

External dependencies can dominate response time

Internal CPU stays low because the application is waiting on another system.

Payment, email, authentication, AI APIs, and other SaaS calls need timeouts, circuit breakers, caching, and fallbacks. Caches also need monitoring for hit ratio, hot keys, stampedes, and TTL.

Dependency checks

Questions to answer

Practical checks for identifying and confirming the real bottleneck.

P95 and P99

Measure this signal with metrics, traces, logs, and a reproducible request.

Timeout

Measure this signal with metrics, traces, logs, and a reproducible request.

Retry policy

Measure this signal with metrics, traces, logs, and a reproducible request.

Sequential calls

Measure this signal with metrics, traces, logs, and a reproducible request.

Cache hit ratio

Measure this signal with metrics, traces, logs, and a reproducible request.

Fallback

Measure this signal with metrics, traces, logs, and a reproducible request.

Investigation

A fast backend can still feel slow to users

Low server CPU says nothing about the CPU on the user's device.

If TTFB is good but LCP or INP is poor, look at the JavaScript bundle, hydration, third-party scripts, images, fonts, and the render waterfall.

Frontend checks

Common issues

Practical checks for identifying and confirming the real bottleneck.

Large bundle

Measure this signal with metrics, traces, logs, and a reproducible request.

Hydration cost

Measure this signal with metrics, traces, logs, and a reproducible request.

Unoptimised images

Measure this signal with metrics, traces, logs, and a reproducible request.

Third-party script

Measure this signal with metrics, traces, logs, and a reproducible request.

Font loading

Measure this signal with metrics, traces, logs, and a reproducible request.

Render waterfall

Measure this signal with metrics, traces, logs, and a reproducible request.

Investigation

Without tracing it is hard to see where a request is slow

Aggregated infrastructure metrics are not enough.

Signals

Minimum observability set

Practical checks for identifying and confirming the real bottleneck.

Distributed tracing

Measure this signal with metrics, traces, logs, and a reproducible request.

RED metrics

Measure this signal with metrics, traces, logs, and a reproducible request.

Database metrics

Measure this signal with metrics, traces, logs, and a reproducible request.

Infrastructure metrics

Measure this signal with metrics, traces, logs, and a reproducible request.

Correlated logs

Measure this signal with metrics, traces, logs, and a reproducible request.

Deployment markers

Measure this signal with metrics, traces, logs, and a reproducible request.

Investigation

Where should the investigation start?

Move from symptom to evidence instead of optimising by intuition.

Illustration: Where should the investigation start

Step 1

Narrow down the symptom

Practical checks for identifying and confirming the real bottleneck.

Endpoint

Measure this signal with metrics, traces, logs, and a reproducible request.

Time window

Measure this signal with metrics, traces, logs, and a reproducible request.

Region

Measure this signal with metrics, traces, logs, and a reproducible request.

Percentile

Measure this signal with metrics, traces, logs, and a reproducible request.

Recent change

Measure this signal with metrics, traces, logs, and a reproducible request.

Step 2

Break down latency

Practical checks for identifying and confirming the real bottleneck.

Browser timing

Measure this signal with metrics, traces, logs, and a reproducible request.

Proxy timing

Measure this signal with metrics, traces, logs, and a reproducible request.

Application spans

Measure this signal with metrics, traces, logs, and a reproducible request.

Database spans

Measure this signal with metrics, traces, logs, and a reproducible request.

External calls

Measure this signal with metrics, traces, logs, and a reproducible request.

Step 3

Confirm the bottleneck

Practical checks for identifying and confirming the real bottleneck.

Reproduce

Measure this signal with metrics, traces, logs, and a reproducible request.

Compare

Measure this signal with metrics, traces, logs, and a reproducible request.

Controlled load test

Measure this signal with metrics, traces, logs, and a reproducible request.

Change one thing

Measure this signal with metrics, traces, logs, and a reproducible request.

Verify

Measure this signal with metrics, traces, logs, and a reproducible request.

Investigation

Fixes that usually do not work

Optimising the wrong place makes the system more complex.

Anti-patterns

Mistakes to avoid

Practical checks for identifying and confirming the real bottleneck.

Adding CPU first

Measure this signal with metrics, traces, logs, and a reproducible request.

Only looking at averages

Measure this signal with metrics, traces, logs, and a reproducible request.

Not separating frontend and backend

Measure this signal with metrics, traces, logs, and a reproducible request.

Ignoring pools and queues

Measure this signal with metrics, traces, logs, and a reproducible request.

No tracing

Measure this signal with metrics, traces, logs, and a reproducible request.

Changing many things at once

Measure this signal with metrics, traces, logs, and a reproducible request.

Investigation

How FlowNexa helps investigate website performance

Find the real bottleneck before scaling infrastructure.

Services

Support scope

Practical checks for identifying and confirming the real bottleneck.

Performance Assessment

Measure this signal with metrics, traces, logs, and a reproducible request.

Observability Setup

Measure this signal with metrics, traces, logs, and a reproducible request.

Database Review

Measure this signal with metrics, traces, logs, and a reproducible request.

Load Testing

Measure this signal with metrics, traces, logs, and a reproducible request.

Infrastructure Review

Measure this signal with metrics, traces, logs, and a reproducible request.

Remediation

Measure this signal with metrics, traces, logs, and a reproducible request.

FlowNexa's view

Low CPU and memory do not prove the system is healthy. Measure waiting time at each hop, queue, and dependency before scaling infrastructure.

Investigation

Frequently asked questions

Answers focused on practical investigation steps.

FAQ

Quick answers

Practical checks for identifying and confirming the real bottleneck.

Does low CPU mean no overload?

Measure this signal with metrics, traces, logs, and a reproducible request.

What to check first?

Measure this signal with metrics, traces, logs, and a reproducible request.

Low database CPU but slow queries?

Measure this signal with metrics, traces, logs, and a reproducible request.

Should we add servers right away?

Measure this signal with metrics, traces, logs, and a reproducible request.

Which metrics matter?

Measure this signal with metrics, traces, logs, and a reproducible request.

Can FlowNexa investigate production?

Measure this signal with metrics, traces, logs, and a reproducible request.

FlowNexa Performance & Observability

Find the bottleneck before adding infrastructure

Practical checks for identifying and confirming the real bottleneck.

Book a performance consultation

Measure this signal with metrics, traces, logs, and a reproducible request.

Request a Performance Assessment

Measure this signal with metrics, traces, logs, and a reproducible request.