Story of Cloudprober
Origin
I started building Cloudprober in 2016, while I was at Google, leading the Cloud Networking SRE team there. Google Cloud was just beginning to grow big, and we were still grappling with some early growth issue. Our biggest problem was that our customers were discovering problems before us, which resulted in bad experience for our customers and huge time sink for my team in debugging those issues¹.
Google’s existing monitoring tools didn’t work well in Cloud, necessitating the need to build things from ground up. And since probers are the cornerstone of monitoring and reliability at Google², that’s where we decided to start. Thus began the journey of Cloudprober.
Even though the primary goal of Cloudprober at that time was to discover and alert on Cloud Networking availability and performance problems, we decided to develop it as a generic prober that could be used to monitor a wide variety of systems and services. We also decided to make Cloudprober open source so that a wider community could trust it, contribute to it, and run it on their own systems.
Focus on Scale, Efficiency
For large-scale platforms like Google Cloud, horizontal scalability and efficiency become critical requirements, and for a monitoring software to be useful, reliability is super important as well. Keeping these requirements in mind, our goal for Cloudprober was for it to be able to reliably monitor 100s of 1000s of endpoints (IPs, Ports, HTTP/S URLs, etc) from each instance, while keeping the resource requirements and management overhead very low³.
Cloudprober maximizes resources utilization by relying heavily on Go concurrency (resource efficiency), supports probing large number of targets in parallel at a high frequency (each instance does more), minimizes the need of frequent updates by supporting dynamic targets discovery (ease of management), has native implementations for common probe types (efficiency), and so on.
Beyond Google and Open-Source
We open-sourced Cloudprober in 2017. That brought in a new phase in its evolution. We added many features over time to make it more useful to the wider community, such as first-class Kubernetes support, a built-in probe status UI, PostgreSQL and Cloudwatch surfacers, OAuth support, Validators, and most recently, built-in alerting capability.
We used the same codebase for the internal and open-source versions, which was more work but it created a huge advantage — our own extensive internal deployment provided a continuous testing platform for Cloudprober, particularly for its scalability and performance aspects, while we added all these features.
Move away from Google Github
I left Google in Nov 2021. To keep working on Cloudprober independently, I moved Cloudprober’s Github repository from github.com/google/cloudprober to github.com/cloudprober/cloudprober. This was a disruptive move and we lost a lot of Github stars in the process (1.4k — 😃), but there was really no other way.
The positive side of this separation was that Cloudprober became less focused on Google’s needs, allowing us to implement many new features that couldn’t be prioritized while I was at the company. Interestingly, it seems that usage of Cloudprober within Google has continued to grow since the transition.
Growth and stability
Throughout its journey, Cloudprober has continuously adapted and expanded to meet the evolving needs of its users⁴. To ensure that Cloudprober thrives and evolves robustly, we’ve been very diligent that it grows in a structured way, a commitment we’ll uphold in future as well.
Footnotes
- A customer-reported infrastructure issue is much harder to debug than an issue discovered by your own monitoring.
- Almost all of Google’s systems rely on probers to detect customer facing problems.
- Hostinger was able to probe 1.8M targets using a single instance: blog.
- I think it’s an essential trait for any software. Software that don’t evolve with time wither away.