Skip to content

Latest commit

 

History

History
67 lines (60 loc) · 3.54 KB

CHANGELOG.md

File metadata and controls

67 lines (60 loc) · 3.54 KB

Changelog

All notable changes to the cluster-setup chart. The chart follows Semantic Versioning; while it is below 1.0, a minor version may break values or CRDs, and every such change is listed under Breaking with what to do.

0.3.0

The first public release. Earlier versions were internal and are not supported upgrade sources; install fresh.

Breaking

  • The operator's CRDs moved to the wingsmith.dev API group (syncedresources.wingsmith.dev, streamcheckpoints.wingsmith.dev). Helm does not install new CRDs on upgrade: apply crds/ by hand first.
  • appVersion is a release version, not latest. The operator and agent images default to it, with pullPolicy: IfNotPresent.
  • cluster.allowedOperations defaults to empty, so a read-only cluster performs no operations until you name them.
  • The operator's RBAC follows cluster.readOnly and cluster.allowedOperations: a read-only install gets no write access outside the release namespace.
  • cluster.hubUrl must be wss://. A ws:// URL is refused unless hub.allowInsecure: true, which also tells the operator to accept it; use it for a local development plane only. hub.caBundle, agent.gateway.serverName and agent.gateway.caBundle are new, for a plane whose certificates come from a private CA.
  • Ingress NetworkPolicies for the operator, serving agent, inference frontends and NATS are on by default (networkPolicy.enabled). Anything else that must reach a frontend or NATS goes in networkPolicy.frontendFrom or networkPolicy.natsFrom.
  • serving.enabled requires an authenticating Traefik Middleware in serving.middlewares (or serving.allowUnauthenticated: true), and, with network policy on, the ingress controller in networkPolicy.frontendFrom.
  • frontend.enabled and operator.frontend.enabled are removed in favour of frontend.mode; the chart refuses either with migration instructions.
  • fabric.devices alone no longer means EFA. Set fabric.type: efa beside it (or rdma with fabric.resource); the chart refuses devices with no type. The operator now gets WINGSMITH_FABRIC and WINGSMITH_FABRIC_DEVICES instead of WINGSMITH_EFA_DEVICES.

Added

  • operator.images.allowedPrefixes and operator.benchmark.images limit the images the plane can have this cluster run: the operator refuses an engine, frontend or prefetch image outside the prefixes, and a benchmark harness not on the list. Both are empty by default, which keeps today's behaviour.
  • fabric.type (efa or rdma) and fabric.privileged, for an RDMA fabric that is not AWS EFA.
  • pools.gpuTolerationKey, pools.gpuTolerationValue and pools.cacheTaintValue, for pools tainted differently from the operator's defaults.
  • agent.terminationGracePeriodSeconds (default 120) sets the agent pod's grace period and its WINGSMITH_SHUTDOWN_GRACE together.
  • operator.benchmark.tokenSecret hands benchmark jobs a Secret's keys as environment, for a tokenizer in a gated repository.
  • networkPolicy.servingEgress closes egress from the serving pods to their own namespace, the release namespace, DNS, the API server and the peers you list. Off by default.
  • monitoring.remoteWrite.bearerToken sends a bearer token from a Secret you created instead of basic auth, for a receiver that takes one. Setting it together with any basic-auth credential is refused.
  • values.schema.json refuses a value of the wrong type before anything renders, such as cluster.readOnly: "false" as a string, which the templates would read as true.