Cloud Profiler

Cloud Profiler is a statistical, low-overhead profiling service that continuously collects CPU and memory usage data from your production applications. Unlike traditional profilers that add significant overhead and are typically only used in development, Cloud Profiler is designed to run in production with negligible impact on performance — typically less than 0.5% CPU overhead.

---

Why Profile in Production?

Development and staging environments rarely replicate production traffic patterns. Workloads that perform well under test may exhibit completely different behaviour under real user traffic — different data distributions, different concurrency levels, and different code paths. Production profiling reveals where your application actually spends its time and memory.

Benefits

Benefit	Description
Find real bottlenecks	Identify CPU and memory hotspots under actual production load
Reduce costs	Optimise code to reduce CPU usage and lower compute costs
Improve latency	Identify functions that contribute most to request latency
Detect regressions	Compare profiles across versions to spot performance regressions
No reproduction needed	Analyse performance issues without reproducing them locally

---

How Cloud Profiler Works

Collection Process

1. The profiling agent runs inside your application process
2. Every 10 seconds (configurable), the agent collects a 10-second profile
3. Only one profile is active per minute per deployment to limit overhead
4. Profiles are uploaded to the Cloud Profiler service for analysis
5. The service aggregates profiles over time for statistical significance

Profile Types

Profile Type	What It Measures	Language Support
CPU time	Time spent executing CPU instructions (excludes I/O waits)	Go, Java, Node.js, Python
Wall time	Elapsed real time (includes I/O waits, sleeps, locks)	Go, Java, Node.js, Python
Heap	Memory currently allocated on the heap	Go, Java, Node.js, Python
Allocated heap	Total memory allocated over time (including freed memory)	Go, Java
Contention	Time threads spend waiting for locks	Go, Java
Threads	Number of active threads	Java

---

Setting Up Cloud Profiler

Go

import "cloud.google.com/go/profiler"

func main() {
    cfg := profiler.Config{
        Service:        "order-service",
        ServiceVersion: "1.2.3",
        ProjectID:      "my-project",
    }
    if err := profiler.Start(cfg); err != nil {
        log.Fatalf("Failed to start profiler: %v", err)
    }

    // Your application code here
}

Python

import googlecloudprofiler

def main():
    googlecloudprofiler.start(
        service='order-service',
        service_version='1.2.3',
        verbose=0,
    )

    # Your application code here

Java

# Download the profiler agent
wget https://storage.googleapis.com/cloud-profiler/java/latest/profiler_java_agent.tar.gz
tar xzf profiler_java_agent.tar.gz

# Start your application with the agent
java -agentpath:/opt/cprof/profiler_java_agent.so=\
  -cprof_service=order-service,\
  -cprof_service_version=1.2.3 \
  -jar my-application.jar

Node.js

require('@google-cloud/profiler').start({
  serviceContext: {
    service: 'order-service',
    version: '1.2.3',
  },
});

---

The Profiler Interface

The Cloud Profiler interface in the Cloud Console provides interactive flame graphs and analysis tools:

Flame Graphs

A flame graph visualises the call stack of your application. Each bar represents a function, and the width of the bar represents the proportion of time (or memory) consumed by that function and its callees:

|====================== main() =======================|
|========= handleRequest() ========|=== logMetrics() =|
|=== parseJSON() =|= queryDB() ===|                   |
|                  |= executeSQL() |                   |