Internals

Read from the source at commit 7110813. Every claim here points at a file and line.

Code map

Path	Responsibility
pkg/scheduler/framework/	Session, Statement, and the plugin registry: the heart of scheduling
pkg/scheduler/actions/	enqueue, allocate, backfill, preempt, reclaim, gangpreempt, gangreclaim, shuffle; registered in factory.go
pkg/scheduler/plugins/	The comparison/predicate/score plugins: gang, drf, proportion, capacity, binpack, predicates, nodeorder, numaaware, and more
pkg/scheduler/api/	The scheduler's in-memory domain model, distinct from the CRD types
pkg/scheduler/cache/	Informer to internal-type conversion and the async bind worker
pkg/controllers/	CRD reconcilers (job, podgroup, queue, jobflow, cronjob, hypernode, gc)
pkg/webhooks/	Validating and mutating admission handlers

Core data structures

• Session (pkg/scheduler/framework/session.go:65): the full snapshot for one cycle. It holds Jobs, Nodes, Queues, HyperNodes, and dozens of function maps (jobOrderFns, predicateFns, nodeOrderFns, preemptableFns, allocatableFns, jobReadyFns, and others) that plugins fill in OnSessionOpen. Its lifetime is a single cycle.
• JobInfo (pkg/scheduler/api/job_info.go:363): one PodGroup. It carries MinAvailable, a TaskStatusIndex (tasks indexed by status), Allocated/TotalRequest, SubJobs, and NetworkTopology. The gang min-member check reads this.
• TaskInfo (pkg/scheduler/api/job_info.go:118): one pod. It carries Resreq/InitResreq, DRAResreq (dynamic resource allocation), a TransactionContext (the current assigned node and status), Priority, Preemptable, BestEffort, and SchGated.
• Resource (pkg/scheduler/api/resource_info.go:60): the quantity model holding CPU, memory, and ScalarResources (GPU/NPU and the like). Its LessEqual and future-idle comparisons are the basis of every fit decision.
• Statement (pkg/scheduler/framework/statement.go): the list of tentative operations. It records Allocate (:246), Pipeline (:140), and Evict in memory, then Commit (:392), Discard, or RecoverOperations confirms, drops, or restores them. This is the dry-run transaction.

A path worth tracing

Statement.Allocate does not write to the API server. It updates the session's in-memory state and records an operation for later replay.

// pkg/scheduler/framework/statement.go:246
func (s *Statement) Allocate(task *api.TaskInfo, nodeInfo *api.NodeInfo) (err error) {
    // ...
    job.UpdateTaskStatus(task, api.Allocated)   // :262, in-memory only
    // ... node.AddTask(task), then append an "allocate" operation
}

The real bind happens only when the action commits the statement, and even then it is queued, not synchronous:

allocate.go:314  if stmt != nil && ssn.JobReady(job) { stmt.Commit() }
statement.go:392 Commit()         -> replays recorded operations
statement.go:317 allocate()       -> cache.AddBindTask(bindContext)   (:319)
cache.go:874     go wait.Until(sc.processBindTask, ...)               (async worker)
cache.go:231     DefaultBinder.Bind(kubeClient, tasks)                -> API server

The gang guarantee falls out of the ssn.JobReady(job) gate at allocate.go:314. The gang plugin registers a JobReady that returns true only when the job's allocated task count reaches MinAvailable. If it does not, Commit never runs and the tentative allocations are dropped, so the cluster never sees a half-placed gang.

Things that surprised me

• Almost all scheduling logic is memory-only. Allocation, pipelining, and eviction mutate session state and append operations; nothing touches the API server until commit, and the bind itself is deferred to the async worker (pkg/scheduler/cache/cache.go:874). This makes the scheduler easy to reason about as a pure function of the snapshot.
• HyperNode search is brute force over clones. Network-topology placement clones a per-hyperNode worksheet, runs a dry-run allocation, scores it, discards it (pkg/scheduler/actions/allocate/allocate.go:501), and recovers only the best result with RecoverOperations (:447). A TODO in the code notes that the per-task check sequence overlaps with the normal path and should eventually merge (allocate.go:656).
• A nomination fast path exists. allocateFromNomination (pkg/scheduler/actions/allocate/allocate.go:595) takes a node that preempt or reclaim already nominated and tries it first, falling back to the normal path if the nomination no longer validates.
• Preempt and reclaim ship disabled. The default ConfigMap runs only enqueue, allocate, backfill (pkg/scheduler/util.go:38), so priority-based eviction is opt-in.