Capabilities

Capabilities are declarative variables you can add to the agents, to allow developers to select the right agent for their pipeline (official documentation).

Disctinct the agents by capabilities. For examples:

• A pool of X64 agents, and a pool of ARM64 agents
• A pool of agents with GPU, and a pool of agents without GPU
• A pool of agents with low performance (standard usage), and a pool of agents with high performance (IA training, intensive C/Rust/GraalVM compilation, …), with distinct Kubernetes Node pool, scaling to 0 when not used (AKS documentation)

Example: ARM64 agents

Take the assumption we want to host a specific instance pool to ARM servers.

# values.yaml
pipelines:
  poolName: private_kube
  capabilities:
    - arch_arm64

extraNodeSelectors:
  kubernetes.io/arch: arm64

Deploy the Helm instance:

❯ helm upgrade --install agent-arm64 clemlesne-blue-agent/blue-agent -f values.yaml

Update the Azure Pipelines file in the repository to use the new pool:

# azure-pipelines.yaml
pool:
  name: private_kube
  demands:
    - arch_arm64

stages: ...

Example: Use different agents on specific jobs

In that example:

• We are using a default agent on ARM64
• Semgrep, our SAST tool, is not compatible with ARM64, let’s use X64 pool
• Our devs are working on a Java project, built with GraalVM, and the container is built locally with BuildKit: we need a system lot of RAM and multipe CPUs for building the application

Our problematic:

• Is it possible to reconcile the efficiency of these different architectures, without restricting ourselves?
• Do we necessarily have to install high performance agents when the use of these large constructions is only a small part of the total execution time (tests, deployments, monitoring, rollback, external services, …)?

We decide to dpeloy these agents:

The developer can now use:

# azure-pipelines.yaml
pool:
  name: private_kube
  demands:
    - arch_arm64
    - perf_standard

stages:
  - stage: build
    jobs:
      - job: sast
        # Use X64 Linux agent because Semgrep is not available on ARM64
        # See: https://github.com/returntocorp/semgrep/issues/2252
        pool:
          name: private_kube
          demands:
            - arch_x64
            - perf_standard
      - job: unit_tests
      - job: container
        # Use high performance agent as Java GraalVM compilation is complex
        pool:
          name: private_kube
          demands:
            - arch_x64
            - perf_high
  - stage: deploy
    jobs:
      - job: upgrade
      - job: dast
      - job: integration_tests