Our notes from KubeCon + CloudNativeCon Europe 2022 in Valencia
KubeCon + CloudNativeCon Europe 2022: what a fantastic event.
KubeCon Europe 2022 brought 7,000 people to Valencia in person, with another 10,000 following from home. Three things stood out across the week. Kubernetes is now the de facto operating system of cloud-native applications. Cost and resource efficiency moved from a side conversation to a main one. And the best-attended sessions were about the limits of autoscaling, not the basics. Adobe reported 30-50% savings from HPA and 5-15% from VPA. Lombard Odier described the opposite constraint: thousands of pods and no time to tune any of them. Both ran into the same wall. The JVM does not cooperate with Kubernetes autoscaling mechanisms.
Below are my notes from the event. They are my personal perspective, with a focus on optimizing Kubernetes applications.
We also had the pleasure to finally meet customers and partners, people, in person after such a long time. And what a nice city Valencia is.
Kubernetes is hot: 5.6 million developers and counting
Well, no surprise here.
Kubernetes has become the de facto operating system of modern cloud-native applications. The CNCF 2021 Cloud Native Survey reported that about 5.6 million developers use Kubernetes worldwide. That is 31% of all backend developers. Kubernetes was first accepted into the CNCF program in 2016 and is now at Graduate level, so that adoption curve is remarkable.
Every organization relying on IT for its business, which these days means every company, needs to treat Kubernetes as a foundational technology. That has real implications for skills, tools and cost control strategies, as I discuss below.
The atmosphere at KubeCon in Valencia reminded me of the days when virtualization boomed with VMware. On a personal level, it brought back memories of VMworld Europe in Cannes in 2008. Another nice city by the sea, by the way.
VMworld established itself as the conference for anything related to virtualization, not just VMware technology. KubeCon + CloudNativeCon now plays that role for cloud-native. Great job from the CNCF in making such an event possible and packing it with learning and networking moments.
Sessions we liked: Kubernetes autoscaling in the real world
We attended a lot of interesting sessions, too many to list them all here. My personal bias is for anything related to scalability, performance and optimization.
Adobe on the horizontal and vertical pod autoscaler
A speech that stood out for me came from Carlos Sanchez, Cloud Engineer at Adobe. He described how Adobe uses standard Kubernetes capabilities to build custom solutions that minimize infrastructure resource waste without adding effort for dev teams.
To bring down costs, Adobe uses Kubernetes autoscalers extensively. The biggest benefits came from autoscaling clusters and pods. In Carlos’s numbers, HPA delivered 30-50% savings, but making the horizontal pod autoscaler work correctly is not easy. CPU autoscaling proved problematic; throughput-based autoscaling was better suited. The vertical pod autoscaler achieved only 5-15% savings. Adobe used VPA in development only, not considering it suitable for production. They also reported 60-80% savings from hibernating development environments.
One of the main challenges is optimizing Java workloads. The JVM, or any other runtime for that matter, does not play nice with Kubernetes autoscaling mechanisms. Startup behavior makes it worse. Carlos’s recommendation was direct. Do not stick to JVM default settings, and set GC type and memory heap according to configured resources. On this topic I would recommend the Dynatrace PurePerformance podcast with our CTO on busting four Java tuning myths.
Lombard Odier on tuning thousands of pods with no time to tune
Another interesting talk came from Vincent Sevel, Architect / Platform Ops at Lombard Odier. He described the challenge of managing large-scale, dynamic environments. There is simply no time to fine-tune tens of workloads across thousands of pods when dozens of releases reach production every day.
As with Adobe, Lombard Odier’s goal is to avoid resource waste without sacrificing SLOs and without workloads crashing. It also has to work without asking anything of dev teams. The core problem is familiar. Dev teams raise CPU requests hoping to fix issues, so CPU is always insufficient, clusters are too small and resources stay overallocated. On this I would recommend our blog on how resource and cost optimization made us love Kubernetes.
Vincent also described several problems we know well, and have helped our customers address. The JVM sizing dilemma is one. Avoiding startup problems without sacrificing cluster stability to CPU throttling is another. So is the well-known challenge of balancing performance against cost. On that last point, watch the Is it observable episode on how to optimize your Kubernetes application.
Moviri on the service efficiency autoscalers will not give you
Finally, I would like to mention the session from Moviri, Akamas strategic partner. I may be biased, but the session delivered by Mauro Pessina, Manager at Moviri Performance Engineering, was very well received. More than 700 people attended it. Pictures are below.
Moviri’s session was titled “Getting the optimal service efficiency that autoscalers won’t give you”. It described a customer story where Akamas AI-powered optimization was adopted to address Kubernetes resource overprovisioning and growing cloud bills. Other customers documented the same problem at KubeCon. You can find the full story in this case study on delivering high-quality SaaS services at lower cloud cost and carbon footprint.
In my humble opinion, the most interesting part came next. Mauro showed how Akamas automatically identified optimal configurations across two layers at once. On the Kubernetes side, that meant CPU and memory limits and requests. On the JVM side, it meant memory heap size and GC type. Both were tuned together to hit cost reduction goals while still matching SLOs for transaction throughput, response time and error rates. That is exactly the challenge other customers reported on stage.
Solutions we liked on the KubeCon show floor
While at KubeCon we did not miss the opportunity to meet some of our technology partners, namely Dynatrace, Gremlin, Google, AWS and RedHat. We also explored the show floor for other interesting solutions. As startup founders, Stefano Doni (Akamas CTO) and I always enjoy those pure innovation discussions with other players at their booths.
Kubernetes cost and troubleshooting tools we saw
We saw several solutions focused on managing the complexity of Kubernetes cloud-native applications. Two categories stood out.
Both categories are perfectly complementary to Akamas. Cost control tools in particular can serve Akamas as additional data sources, so cost metrics feed directly into optimization goals and constraints.
CNCF open-source projects we liked
We also noticed a positive mood around all CNCF open-source projects. These are the community-powered initiatives we liked most.
- Keptn, whose booth was packed. We believe that owes something to our friends Andi Grabner and Henrik Rexed, who are always up for an interesting conversation about technology. On that note, you may be interested in our integration with Keptn and Dynatrace Cloud Automation.
- Pixie, an interesting sandbox project for Kubernetes observability. It provides fine-grained microservices and database performance tracing using Linux kernel eBPF, with no service mesh to set up.
- OpenTelemetry, which has now reached more than 250 contributors. We expect Pixie to grow the same way.
- LitmusChaos, a CNCF incubating project on the chaos engineering side. We are pleased to see it getting a lot of traction from the community.
A note on the Italian presence at KubeCon Valencia
Please also allow me a final note on the Italian presence at KubeCon in Valencia, in exhibitors, attendees and contributors to open-source projects. It was really a pleasure to meet Federico Soncini Sessa and Giulio Roggero from Mia Platform, Gabriele Bartolini and the EDB team, and Jacopo Nardiello from Sighup. It was equally good to see the SparkFabrik and Clastix teams, and Loris Degioanni from Sysdig. Italy is generally less known for IT technology than for other areas of excellence. It remains a great place to start innovative companies and develop visionary technologies.
Conclusions
This event confirmed something for me. The challenges of running Kubernetes cloud-native applications are now well understood and well documented. We were pleased by the huge interest in improving service quality and cost efficiency for Kubernetes microservices applications.
We believe this is really an exciting time for any company like Akamas.
We look forward to meeting you in October at KubeCon + CloudNativeCon North America 2022 in Detroit, with many new customer stories and product updates to share.
FAQs
Where was KubeCon Europe 2022 held?
KubeCon + CloudNativeCon Europe 2022 was held in Valencia, Spain. It was the first European edition to run in person again after the pandemic, and it combined an on-site event with a virtual track.
How many people attended KubeCon Europe 2022?
About 7,000 people attended in person in Valencia. Another 10,000 followed the event remotely. That made it one of the largest cloud-native gatherings held in Europe at the time.
What savings did Adobe report from Kubernetes autoscaling?
In Carlos Sanchez’s session, Adobe reported 30-50% savings from the horizontal pod autoscaler and 5-15% from the vertical pod autoscaler. VPA was used in development only. Hibernating development environments delivered a further 60-80%.
Why is the JVM difficult to autoscale on Kubernetes?
The JVM does not cooperate with Kubernetes autoscaling mechanisms, and startup behavior makes it worse. Default JVM settings ignore the pod’s configured resources. GC type and memory heap have to be set to match the CPU and memory actually granted to the container.
Which CNCF projects stood out at KubeCon Europe 2022?
Four caught our attention: Keptn for continuous delivery and operations, Pixie for eBPF-based Kubernetes observability, OpenTelemetry with more than 250 contributors, and LitmusChaos for chaos engineering.
