Skip to content

Latest commit

 

History

History
78 lines (65 loc) · 3.57 KB

TODO.md

File metadata and controls

78 lines (65 loc) · 3.57 KB
Raw

TODO

The estimated complexity and priority of a feature/task is defined like this:

  • Priority scale: High (1), Medium (2) and Low (3). This is marked as Px in the task description.

  • Complexity scale: C1, C2, C4 and C8. The complexity scale is exponential, with complexity 1 being the lowest complexity. The estimated complexity is a function of both task 'complexity' and task 'scope'. If you think the complexity is >C8, then that is a good hint that the task needs to be split into smaller pieces.

Reference Plugins (this repository)

  • admin/repo
    • get proper permissions so that self-hosted runner scripts can create/destroy runner C1
  • cache
    • rework/split up, make the core functionality be usable for other types of plugins, C4
    • get rid of CRI-specific representation/interfaces, NRI/CRI pod, container conversion, C4
    • eliminate need for saving cache to disk, C2
  • config
    • switch to using CRDs (from ConfigMaps), C2
    • consider usig 1 CRD per policy as opposed to a single 'union' CRD, C2
    • create helm chart support for the resource policies, C1
  • policies
    • remove support for multiple policies in a single binary, C1
    • balloons: implement topology hint support, C2
    • balloons: make sure (just) enough permissions for cpufreq control from within containers, C2
    • balloons: implement node resource topology export, C4
    • topology-aware: legacy block I/O, RDT support if needed, C2
    • topology-aware: cleanup/refactor (rewrite nodes, supply, request, grant), C4
  • misc/infra/other
    • rework pkg hierarchy (with co-hosted plugins of other 'classes' in mind), C8
    • eliminate/replace resmgr in other user-visible 'artifacts' where appropriate, C1
    • check and unify annotation naming for consistency, C1
    • structural logging (with better configurability), check what Patrik did, C2
    • agent usage should be optional and controllable, C2
    • fix crun+cgroupv2 support (ineffective/broken for CPU and memory)
      • set cgroup parameters using v2/unified notation if possible
      • check if this fixes the problems with crun+cri-o
  • instrumentation
    • switch metrics collection to opentelemetry from opencensus
    • make sure default go runtime metrics gets properly exported
    • export policy-agnostic metrics data (maybe same as for node resource topology)
    • (maybe) remove any resulting duplicated data export from policy-specific data
  • testing
    • more unit tests, C4
    • unit tests in GH actions, C4
    • e2e-tests on merge by CI/self-hosted runners, C4
    • migrate shell based e2e-tests into Ansible, C4
    • make topology-aware e2e test07-mixed-allocations more reliable, C1
    • allow e2e tests to use the upstream containerd or cri-o instead of using binaries compiled from local sources, C1
  • documentation
    • minimal README with instructions about how to build/deploy/try the plugins, C2

NRI Core (https://github.com/containerd/nri)

  • Darwin support, C4
  • Windows support, C8
  • add warning to documentation about socket 'access security', C1
  • openshift: (manual) bootstrapping of enabling NRI support, C4
  • check if runtime-side extra adaptation code could be generalized and moved to NRI, C4
  • less protocol stutter:
    • add minimal pod/container cache to stub, C4
    • replace pod/container ID in most protocol messages, C1
  • add trace spans in adaptation for request/event handling
  • add trace spans to stub for request/event handling

Containerd integration

CRI-O integration