Datadog Report Details Increased Adoption of Containers
A Datadog report found the container landscape appears to be evolving as more organizations embrace serverless computing frameworks and use of graphical processing units (GPUs) increases thanks mainly to the rise of artificial intelligence (AI) models built using containers.
Based on an analysis of more than 2.4 billion containers run by thousands of organizations, the report found that 46% of organizations now run serverless containers, up from 31% two years ago.
At the same time, the report notes there has been a 58% year-over-year increase in the GPU compute time consumed by containers that are most likely being used to train AI models.
Yrieix Garnier, vice president of products for Datadog, said overall, the report makes it clear that the usage of containers is both expanding and maturing.
For example, 40% of organizations are now using version 1.25 of Kubernetes or higher, all of which have been available for less than a year. That suggests organizations are more committed to staying current to address cybersecurity vulnerabilities as they arise, noted Garnier. A year ago, only 5% of organizations were running versions of Kubernetes that were less than a year old, the report noted. The report also found more than 40% of organizations running more than 1,000 hosts are using a service mesh.
In addition, the report noted 53% of organizations are also currently using containerd, compared to 23% a year ago. As more organizations have migrated to newer versions of Kubernetes that no longer support Dockershim, there has also been a decline in the usage of Docker, which fell to 65% from 88% over the past year.
More than 42% of organizations are now also hosting databases using containers, the report found. That suggests the number of stateful applications built using containers has substantially increased in recent years. Node.js (43%) is the most widely used programming language.
Additionally, more than 80% of organizations have enabled Horizontal Pod Autoscaling (HPA), with 45% using it everywhere they run Kubernetes clusters.
However, there is still much progress to be made in terms of efficiency and cost control. More than 65% of Kubernetes workloads use less than half of their requested CPU and memory. There is still a tendency to overprovision workloads despite the ability of Kubernetes clusters to scale applications up and down, noted Garnier. Less than 1% of respondents are using the Vertical Pod Autoscaler (VPA) capability enabled by Kubernetes that recommends CPU and memory requests and limits based on past container resource usage. One reason for that lack of usage is VPA is not recommended for use alongside HPA.
There has, however, been increased adoption of Arm-based processors (7%) as a more efficient option for running Kubernetes clusters.
As more instances of Kubernetes are deployed in production environments, the processes used to manage fleets of clusters are evolving. Managing a small number of Kubernetes clusters running a handful of applications is one thing; managing a portfolio of cloud-native applications that can now easily span hundreds of clusters is quite another challenge.
In the meantime, IT teams should expect that the bulk of the applications being deployed from here on out are going to be based on microservices constructed using containers.
There is, of course, no shortage of tools for managing those applications and the clusters they run on, but the expertise required to manage Kubernetes clusters at scale is still hard to find and retain. Most IT teams will, in the coming year, need to continue to invest in additional training for their existing IT staff to fill a void that is only going to grow larger with each new cluster deployed.
