CloudPilot AI Roadmap

The goal of CloudPilot AI is to build an extremely elastic infrastructure, ensuring the lowest cost while maintaining stability. From the underlying computing resources to the operating system and up to the upper-layer applications, it achieves optimal cost and extreme elasticity through technologies such as elasticity and scheduling from bottom to top.

They are aligned with the roadmap direction for the next half year. As for the detailed roadmap, we’ll update and list below:

2026 First Quarter Roadmap

This iteration mainly focuses on troubleshooting and fixing, performance enhancements, and intelligent configurations, providing users with ultimate flexibility and significantly reducing their maintenance operations.

• Anomaly detection and automatic repair: Automatically detect cluster anomalies, issue alert notifications, and perform automatic repairs, such as automatically fixing abnormal PDB configurations that prevent cluster resource consolidation, or automatically upgrading disk sizes when node disks are insufficient.
• Shutdown-based node startup acceleration: By pre-warming nodes, directly starting nodes from a shutdown state, the node startup time is reduced from 50s to 20s.
• P2P image pull acceleration: Using P2P technology to accelerate the pulling of new Pod images, mainly targeting large image operations.
• Intelligent configuration: Currently, for different environments (such as development and production), users need to fine-tune many configurations, which can be burdensome. CloudPilot AI is optimizing this part, with the goal of fully intelligent and automated configuration to reduce user configuration tasks.
• microVM-based live migration: CRIU, which runc uses, has many limitations, and there are a lot of restrictions during hot migration, such as when the container uses fsnotify. CloudPilot AI will use microVM, bypassing CRIU, to achieve a general live migration. Ultimately, users can leverage this technology to achieve ultra-fast startup (even for Java), allowing the workloads to migrate from one node to another without interruption (for Job types). Combined with CloudPilot AI, this approach reduces costs and improves performance.

Workload Autoscaler

Supports DaemonSet Workloads

We are currently adding support for DaemonSet workloads, and this capability is expected in an upcoming release. Due to the unique characteristics of DaemonSets, we must ensure that optimizations never interfere with their per-node guarantees or their lifecycle.

DaemonSets will be supported via OnCreate and InPlace update modes only, ensuring that optimization does not trigger unwanted Pod recreation.

HPA Compatibility

VPA and HPA compatibility has long been a difficult challenge, as both attempt to control Pod resource usage but in fundamentally different ways, inevitably leading to conflicts. CloudPilot AI will introduce improved mechanisms to enable seamless coexistence between HPA and the Workload Autoscaler in future releases.

2025 Fourth Quarter Roadmap

This iteration mainly focuses on completing the basic functions and optimizing the application layer, providing customers with a comprehensive basic functionality experience.

• CloudPilot AI Universal: Supports optimizing the Pod resource requests/limits of any K8s cluster, including on-premises and public clouds, with functionalities consistent with the Workload autoscaler.
• GKE support: CloudPilot AI will provide GKE optimization support.
• Organization management: CloudPilot AI will support organization management, allowing multiple organization members to be assigned different permissions for unified cluster management.

Workload Autoscaler

Better Java Workload Handling

Java workloads have always been challenging in Kubernetes environments. This is because Java applications typically adjust their internal memory allocation and garbage collection behavior based on the resource settings available at startup. As a result, when Requests and Limits are dynamically modified, the JVM may fail to properly adapt, leading to performance degradation or stability issues.

Due to the JVM’s unique design, VPA-like solutions are almost incapable of tuning Java workloads effectively. CloudPilot AI has recognized this problem and plans to introduce specialized Java workload handling in future releases.

We will use a non-intrusive approach (such as eBPF) to automatically collect JVM metrics from all Java workloads across the cluster. By leveraging indicators such as JVM heap usage and GC latency, we can more accurately assess resource requirements and perform fully automated Java optimization—continuously monitoring GC behavior, CPU throttling, and other signals to balance performance and efficiency.

Workload Startup Boost

Java workloads often require multiple times more resources during startup compared to their steady state. We refer to this as the Startup Spike problem. With the help of InPlace Update, CloudPilot AI ensures that Java workloads receive sufficient resources during startup and then automatically adjusts their Requests and Limits to the appropriate runtime values, eliminating over-provisioning.

For example: If a Java workload needs 1Gi to run but requires 3Gi during startup, Startup Boost will temporarily raise the memory request to 3Gi at startup, and automatically reduce it back to 1Gi after the workload becomes ready.

Previously, users had two undesirable choices:

• statically configure high Requests to accommodate startup needs → wasting resources,
• or configure low Requests to save resources → risking startup failure.

Startup Boost fully automates this process and eliminates the trade-off.

Auto Limit Policy

Configuring Limit policies is difficult for many users because it’s unclear how to balance performance and cost. Even worse, different workloads may require different limit strategies, increasing operational complexity.

CloudPilot AI will provide an Auto Limit Policy, which intelligently adjusts Limits based on cluster and node resource conditions. This simplifies configuration, improves user experience, and supports defining separate CPU and Memory limit policies.

Workload Container Application Type & Language Detection

In Kubernetes, different types of workloads have varying resource needs and runtime behaviors. CloudPilot AI plans to introduce Application Type & Language Detection, allowing the Workload Autoscaler to automatically identify the workload’s application category and programming language—and optimize resources accordingly.

We plan to support detection of major programming languages, including Java, Go, Python, Node.js, .NET, Ruby, PHP, and others.

In addition, we will support automatic detection of ApplicationType, covering more than 50 common Kubernetes applications, such as MySQL, Envoy, Kafka, Redis, Elasticsearch, Prometheus, and many more.