Documentation
¶
Overview ¶
Package schedule provides flate's dependency-driven reconcile scheduler: a re-entrant fixpoint engine that runs each node's reconcile body on a bounded task pool, parks a body that reports unsatisfied dependencies (Dispatcher OutcomeBlocked), and re-runs it when any of those dependencies advances.
Termination is STRUCTURAL, not timed. Every render emission is a synchronous store write on the body's own task goroutine, completing before the body returns and before the scheduler decrements its in-flight count. Therefore when no body is in flight and the runnable frontier is empty, no new object can ever appear — so any still-parked node's dependencies are provably unproducible. A draining sweep then terminalizes those nodes with the canonical "dependency not found" / cascade / "not ready" statuses. There is no per-dependency timeout and no shared quiescence counter, so the #666 transient-drain false-drop cannot occur: a parked node is never counted in flight, and nothing drops it except the fixpoint, which only fires when nothing is running.
The package depends only on pkg/manifest and pkg/task; all store and controller interaction is behind the Dispatcher seam, so the scheduler is unit-testable against a fake Dispatcher with no store or controllers.
Index ¶
Constants ¶
const ( // DrainNone: normal operation — an unsatisfiable dependency parks. DrainNone = 0 // DrainCascade: an absent dependency (and a never-true ReadyExpr) // terminalizes as a failure; a present-but-Pending dependency still // parks, so a dangling chain fails leaf-first and each level cascades // the child's real terminal message upward. DrainCascade = 1 // DrainForce: a present-but-Pending dependency ALSO terminalizes ("not // ready"). Reached only when a DrainCascade pass made no progress — i.e. // a cross-kind structural cycle the same-kind preflight detector cannot // represent; forcing the failure breaks it. DrainForce = 2 )
Drain levels passed to Dispatcher.Dispatch. 0 is normal operation; the scheduler escalates only at the structural fixpoint (nothing in flight, nodes still parked).
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type Dispatcher ¶
type Dispatcher interface {
// Dispatch invokes id's reconcile body synchronously on the calling
// goroutine (a task.Service worker) and reports back:
// - out: OutcomeTerminal or OutcomeBlocked.
// - blocked: unsatisfied dependency ids (non-nil only when Blocked).
// drainLevel is one of DrainNone/DrainCascade/DrainForce.
Dispatch(ctx context.Context, id NodeID, drainLevel int) (out Outcome, blocked []NodeID)
}
Dispatcher runs a node's reconcile body. The orchestrator supplies the concrete implementation, closing over the store and the three controllers; the scheduler never sees a store or controller type.
type NodeID ¶
type NodeID = manifest.NamedResource
NodeID identifies a schedulable node — a Kustomization, HelmRelease, or source CR — by its store identity.
type Outcome ¶
type Outcome int
Outcome is what one reconcile-body invocation reported via the Dispatcher.
const ( // OutcomeTerminal means the body wrote a terminal store status // (Ready/Skipped/Failed); the node is done unless a later store event // re-emits or resets it. OutcomeTerminal Outcome = iota // OutcomeBlocked means the body could not proceed because one or more // dependencies are unsatisfied; the scheduler parks the node keyed on the // returned ids and re-runs it when any advances. OutcomeBlocked )
type Scheduler ¶
type Scheduler struct {
// contains filtered or unexported fields
}
Scheduler is a re-entrant fixpoint reconcile driver. Construct with New, Seed the initial node set, wire store events to OnArrival/OnStatusWake, then call Run.
func New ¶
func New(tasks *task.Service, disp Dispatcher) *Scheduler
New constructs a Scheduler that runs bodies on tasks via disp.
func (*Scheduler) OnArrival ¶
OnArrival is called from the store's EventObjectAdded subscription (which fires only when an object's content actually changed — including a Refire's status reset). schedulable reports whether id is a node the scheduler runs (a Kustomization/HelmRelease/source) versus pure data (a ConfigMap/Secret): a data arrival must WAKE nodes parked on it (a KS waiting on a substituteFrom CM) but is never registered as a runnable node. A content-changed arrival of a terminal node re-dispatches it (the re-run re-reads and is idempotent via fingerprint dedup), which is what restores a Refired producer/source.
func (*Scheduler) OnStatusWake ¶
OnStatusWake is called from the store's EventStatusUpdated subscription. It acts only on a TERMINAL store status (Ready or Failed): a parked node gates on Ready/Failed, so intermediate Pending progress writes never change its gate answer and waking on them is pure churn that also widens race windows.
func (*Scheduler) Run ¶
Run drives the scheduler to a fixpoint, returning when every node is terminal (or ctx is canceled) after the in-flight bodies drain.
func (*Scheduler) Seed ¶
Seed registers the initial node set (file-loaded Kustomizations, HelmReleases, and source CRs from Bootstrap) as runnable, in deterministic id order. Duplicates and already-known ids are ignored.
func (*Scheduler) SetRerunAtDrain ¶
SetRerunAtDrain installs the predicate that decides whether a node re-runs at the structural fixpoint (a selector-only ResourceSet, which has no nameable input provider to park on). It is evaluated in the dispatch goroutine after each Dispatch, so it may read the store. Optional — nil means no node reruns.
Source Files