Cohesive.Adapters.AzureML 0.1.0-alpha.64

Cohesive.Adapters.AzureML

Azure Machine Learning adapters for model training and dataset registry workflows.

Install

dotnet add package Cohesive.Adapters.AzureML

Use When

• You want to run Cohesive training requests against Azure Machine Learning.
• You need dataset and model registry integration backed by Azure ML workspace resources.
• You want model training to remain behind Cohesive.AI provider-neutral contracts.

Example

using Cohesive.Adapters.AzureML;

var azureMl = new AzureMLOptions
{
    SubscriptionId = configuration["AzureML:SubscriptionId"],
    ResourceGroupName = configuration["AzureML:ResourceGroupName"],
    WorkspaceName = configuration["AzureML:WorkspaceName"],
    RegistryName = configuration["AzureML:RegistryName"]
};

var trainer = azureMl.GetModelTrainer();
var registry = azureMl.GetRegistry();

Training submission recovery

AzureMLModelTrainer realizes TrainingJobSubmission with a deterministic Azure ML job name derived from the complete logical submission identity. The job retains the original logical identity and the versioned exact-request fingerprint in cohesive.submissionId and cohesive.requestFingerprint properties.

Submitting the same identity and request returns the existing job. Reusing the identity with different request content, or colliding with a provider job that lacks Cohesive submission evidence, fails with TrainingJobSubmissionConflictException; the adapter never treats that job as accepted for the caller.

ReconcileSubmissionAsync performs a direct lookup by deterministic job name:

• a matching job returns Accepted;
• an authoritative not-found result returns ConfirmedAbsent;
• an Azure request failure returns Unresolved with portable error and retryability evidence; and
• mismatched identity or fingerprint evidence throws TrainingJobSubmissionConflictException.

This is stable-identity idempotency, not a claim of physical exactly-once delivery. The calling durable workflow retains attempt, timeout, cancellation, and retry authority and must reuse the exact same TrainingJobSubmission through reconciliation and any admitted retry.

Training cancellation recovery

AzureMLModelTrainer implements ICancellableModelTrainer through Azure ML's command-job cancellation operation. It first observes the exact provider job:

• Completed, Failed, or Cancelled returns AlreadyTerminal without issuing cancellation;
• CancelRequested returns Accepted without redundant dispatch;
• authoritative absence returns NotFound; and
• any other observed state is eligible for Azure ML cancellation.

A successful Azure cancellation request returns Accepted, which remains non-terminal evidence. If dispatch fails, the adapter observes the job again before classifying the attempt. A terminal race or CancelRequested state wins over the transport failure. Absence after a cancellation 404 is NotFound; a still-observable job plus the same 404 is Unresolved because the provider evidence conflicts. Throttling, timeout, conflict, transport, and server failures remain retryable Unresolved evidence, while deterministic authorization or validation refusal is Rejected.

Azure ML's CancelRequested status maps to TrainingJobStatus.CancellationRequested, never ordinary Running. The calling workflow must continue status observation until the provider reports terminal state. Cancelling the method's CancellationToken stops only the local wait and does not prove that the Azure ML job stopped.

Related Packages

• Cohesive.AI for training, dataset, and model registry contracts.
• Cohesive.Adapters.AzureStorage for Azure Blob-backed training artifacts and dataset streams.