flowyml 1.1.0__py3-none-any.whl → 1.3.0__py3-none-any.whl

flowyml/core/remote_orchestrator.py

@@ -0,0 +1,109 @@
+"""Remote Orchestrator - Executes pipelines on remote infrastructure."""
+
+from typing import Any, TYPE_CHECKING
+
+from flowyml.stacks.components import Orchestrator, ComponentType, ResourceConfig, DockerConfig
+from flowyml.core.execution_status import ExecutionStatus
+from flowyml.core.submission_result import SubmissionResult
+
+if TYPE_CHECKING:
+    from flowyml.core.pipeline import Pipeline
+
+
+class RemoteOrchestrator(Orchestrator):
+    """Base orchestrator for remote execution.
+
+    This orchestrator submits jobs to remote infrastructure and returns job IDs.
+    Cloud-specific orchestrators (AWS, GCP, Azure) inherit from this.
+    """
+
+    def __init__(self, name: str = "remote"):
+        super().__init__(name)
+
+    @property
+    def component_type(self) -> ComponentType:
+        return ComponentType.ORCHESTRATOR
+
+    def validate(self) -> bool:
+        """Validate remote orchestrator configuration."""
+        return True
+
+    def to_dict(self) -> dict[str, Any]:
+        return {
+            "name": self.name,
+            "type": "remote",
+        }
+
+    def get_run_status(self, job_id: str) -> ExecutionStatus:
+        """Get status of a remote pipeline run.
+
+        This should be overridden by cloud-specific orchestrators to query
+        the actual remote execution status.
+
+        Args:
+            job_id: The remote job identifier.
+
+        Returns:
+            The current execution status.
+        """
+        return ExecutionStatus.RUNNING
+
+    def fetch_step_statuses(self, job_id: str) -> dict[str, ExecutionStatus]:
+        """Get status of individual steps in a remote run.
+
+        Args:
+            job_id: The remote job identifier.
+
+        Returns:
+            Dictionary mapping step names to their execution status.
+        """
+        # Default implementation - override in subclasses
+        return {}
+
+    def stop_run(self, job_id: str, graceful: bool = True) -> None:
+        """Stop a remote pipeline run.
+
+        Args:
+            job_id: The remote job identifier.
+            graceful: If True, attempt graceful shutdown. If False, force kill.
+
+        Raises:
+            NotImplementedError: If stopping is not supported.
+        """
+        raise NotImplementedError(
+            f"{self.__class__.__name__} does not support stopping runs",
+        )
+
+    def run_pipeline(
+        self,
+        pipeline: "Pipeline",
+        run_id: str,
+        resources: ResourceConfig | None = None,
+        docker_config: DockerConfig | None = None,
+        inputs: dict[str, Any] | None = None,
+        context: dict[str, Any] | None = None,
+        **kwargs,
+    ) -> SubmissionResult:
+        """Submit pipeline to remote infrastructure.
+
+        This base implementation should be overridden by cloud-specific orchestrators
+        to submit to their respective services (AWS Batch, Vertex AI, Azure ML, etc.).
+
+        Args:
+            pipeline: The pipeline to run.
+            run_id: The unique run identifier.
+            resources: Resource configuration.
+            docker_config: Docker configuration.
+            inputs: Input data.
+            context: Context variables.
+            **kwargs: Additional arguments.
+
+        Returns:
+            SubmissionResult with remote job ID and optional wait function.
+
+        Raises:
+            NotImplementedError: Must be implemented by cloud-specific orchestrators.
+        """
+        raise NotImplementedError(
+            "RemoteOrchestrator.run_pipeline must be implemented by cloud-specific orchestrators",
+        )

flowyml/core/resources.py

@@ -39,7 +39,7 @@ class GPUConfig:
     @staticmethod
     def _is_valid_memory(memory: str) -> bool:
         """Check if memory string is valid (e.g., '16Gi', '32768Mi')."""
-        return bool(re.match(r"^\d+(\.\d+)?(Ki|Mi|Gi|Ti|K|M|G|T)$", memory))
+        return bool(re.match(r"^\d+(\.\d+)?(Ki|Mi|Gi|Ti|KB|MB|GB|TB|K|M|G|T)$", memory))
 
     def to_dict(self) -> dict[str, Any]:
         """Convert to dictionary representation."""
@@ -105,7 +105,7 @@ class GPUConfig:
         def to_bytes(mem: str) -> int:
             import re
 
-            match = re.match(r"^(\d+(?:\.\d+)?)(Ki|Mi|Gi|Ti|K|M|G|T)?$", mem)
+            match = re.match(r"^(\d+(?:\.\d+)?)(Ki|Mi|Gi|Ti|KB|MB|GB|TB|K|M|G|T)?$", mem)
             if not match:
                 return 0
             value, unit = float(match.group(1)), match.group(2) or ""
@@ -114,6 +114,10 @@ class GPUConfig:
                 "Mi": 1024**2,
                 "Gi": 1024**3,
                 "Ti": 1024**4,
+                "KB": 1000,
+                "MB": 1000**2,
+                "GB": 1000**3,
+                "TB": 1000**4,
                 "K": 1000,
                 "M": 1000**2,
                 "G": 1000**3,
@@ -236,8 +240,8 @@ class ResourceRequirements:
 
     @staticmethod
     def _is_valid_memory(memory: str) -> bool:
-        """Check if memory string is valid (e.g., '16Gi', '32768Mi')."""
-        return bool(re.match(r"^\d+(\.\d+)?(Ki|Mi|Gi|Ti|K|M|G|T|B)?$", memory))
+        """Check if memory string is valid (e.g., '16Gi', '32768Mi', '4GB')."""
+        return bool(re.match(r"^\d+(\.\d+)?(Ki|Mi|Gi|Ti|KB|MB|GB|TB|K|M|G|T|B)?$", memory))
 
     def to_dict(self) -> dict[str, Any]:
         """Convert to dictionary representation."""
@@ -258,6 +262,15 @@ class ResourceRequirements:
         """Check if GPU resources are requested."""
         return self.gpu is not None
 
+    def __getitem__(self, key: str) -> Any:
+        """Provide dict-style access for backwards compatibility."""
+        if not hasattr(self, key):
+            raise KeyError(key)
+        value = getattr(self, key)
+        if key == "gpu" and isinstance(value, GPUConfig):
+            return value.count
+        return value
+
     def get_gpu_count(self) -> int:
         """Get total number of GPUs requested."""
         return self.gpu.count if self.gpu else 0
@@ -295,7 +308,7 @@ class ResourceRequirements:
         import re
 
         def to_bytes(mem: str) -> int:
-            match = re.match(r"^(\d+(?:\.\d+)?)(Ki|Mi|Gi|Ti|K|M|G|T|B)?$", mem)
+            match = re.match(r"^(\d+(?:\.\d+)?)(Ki|Mi|Gi|Ti|KB|MB|GB|TB|K|M|G|T|B)?$", mem)
             if not match:
                 return 0
             value, unit = float(match.group(1)), match.group(2) or "B"
@@ -304,6 +317,10 @@ class ResourceRequirements:
                 "Mi": 1024**2,
                 "Gi": 1024**3,
                 "Ti": 1024**4,
+                "KB": 1000,
+                "MB": 1000**2,
+                "GB": 1000**3,
+                "TB": 1000**4,
                 "K": 1000,
                 "M": 1000**2,
                 "G": 1000**3,

flowyml/core/retry_policy.py

@@ -0,0 +1,80 @@
+"""Retry policies for orchestrators."""
+
+from dataclasses import dataclass
+from typing import Optional, TYPE_CHECKING
+from flowyml.core.error_handling import RetryConfig, ExponentialBackoff, execute_with_retry
+
+if TYPE_CHECKING:
+    from flowyml.core.pipeline import Pipeline
+
+
+@dataclass
+class OrchestratorRetryPolicy:
+    """Retry policy for orchestrator-level failures.
+
+    This handles retries at the orchestrator level (entire pipeline runs),
+    distinct from step-level retries.
+    """
+
+    max_attempts: int = 3
+    """Maximum number of pipeline retry attempts"""
+
+    initial_delay: float = 60.0
+    """Initial delay between retries in seconds"""
+
+    max_delay: float = 600.0
+    """Maximum delay between retries in seconds"""
+
+    multiplier: float = 2.0
+    """Backoff multiplier for exponential backoff"""
+
+    retry_on_status: list[str] = None
+    """Retry on specific execution statuses (e.g., ['FAILED', 'STOPPED'])"""
+
+    def __post_init__(self):
+        if self.retry_on_status is None:
+            self.retry_on_status = ["FAILED"]
+
+    def to_retry_config(self) -> RetryConfig:
+        """Convert to RetryConfig for execute_with_retry."""
+        backoff = ExponentialBackoff(
+            initial=self.initial_delay,
+            max_delay=self.max_delay,
+            multiplier=self.multiplier,
+            jitter=True,
+        )
+
+        return RetryConfig(
+            max_attempts=self.max_attempts,
+            backoff=backoff,
+            retry_on=[Exception],  # Catch all exceptions
+            not_retry_on=[KeyboardInterrupt],  # Don't retry on manual interruption
+        )
+
+
+def with_retry(orchestrator_method):
+    """Decorator to add retry logic to orchestrator methods.
+
+    Usage:
+        @with_retry
+        def run_pipeline(self, pipeline, ...):
+            ...
+    """
+
+    def wrapper(self, pipeline: "Pipeline", *args, retry_policy: Optional[OrchestratorRetryPolicy] = None, **kwargs):
+        if retry_policy is None:
+            # No retry policy, execute normally
+            return orchestrator_method(self, pipeline, *args, **kwargs)
+
+        # Execute with retry
+        retry_config = retry_policy.to_retry_config()
+        return execute_with_retry(
+            orchestrator_method,
+            retry_config,
+            self,
+            pipeline,
+            *args,
+            **kwargs,
+        )
+
+    return wrapper

flowyml/core/step.py

@@ -1,5 +1,6 @@
 """Step Decorator - Define pipeline steps with automatic context injection."""
 
+import contextlib
 import hashlib
 import inspect
 import json
@@ -9,9 +10,10 @@ from dataclasses import dataclass, field
 
 # Import resource types
 try:
-    from flowyml.core.resources import ResourceRequirements
+    from flowyml.core.resources import ResourceRequirements, GPUConfig
 except ImportError:
     ResourceRequirements = None  # Type: ignore
+    GPUConfig = None  # Type: ignore
 
 
 @dataclass
@@ -62,6 +64,21 @@ class Step:
 
         # Store resources (accept both dict for backward compatibility and ResourceRequirements)
         self.resources = resources
+        if self.resources and ResourceRequirements and not isinstance(self.resources, ResourceRequirements):
+            if isinstance(self.resources, dict):
+                resource_kwargs = dict(self.resources)
+                gpu_value = resource_kwargs.get("gpu")
+                if GPUConfig and gpu_value is not None:
+                    if isinstance(gpu_value, dict):
+                        resource_kwargs["gpu"] = GPUConfig(
+                            gpu_type=gpu_value.get("gpu_type") or gpu_value.get("type") or "generic",
+                            count=int(gpu_value.get("count", 1)),
+                            memory=gpu_value.get("memory"),
+                        )
+                    elif isinstance(gpu_value, (int, float)):
+                        resource_kwargs["gpu"] = GPUConfig(gpu_type="generic", count=int(gpu_value))
+                with contextlib.suppress(TypeError):
+                    self.resources = ResourceRequirements(**resource_kwargs)
 
         self.tags = tags or {}
         self.condition = condition

flowyml/core/submission_result.py

@@ -0,0 +1,53 @@
+"""Submission result for async pipeline execution."""
+
+from typing import Any, Optional
+from collections.abc import Callable
+
+
+class SubmissionResult:
+    """Result of submitting a pipeline run to an orchestrator.
+
+    This class enables async execution patterns where the orchestrator
+    submits the pipeline and returns immediately, optionally providing
+    a way to wait for completion.
+    """
+
+    def __init__(
+        self,
+        job_id: str,
+        wait_for_completion: Optional[Callable[[], None]] = None,
+        metadata: Optional[dict[str, Any]] = None,
+    ):
+        """Initialize a submission result.
+
+        Args:
+            job_id: The remote job/run identifier.
+            wait_for_completion: Optional function to block until pipeline completes.
+            metadata: Optional metadata about the submission.
+        """
+        self.job_id = job_id
+        self.wait_for_completion = wait_for_completion
+        self.metadata = metadata or {}
+
+    def wait(self, timeout: Optional[int] = None) -> None:
+        """Wait for the pipeline run to complete.
+
+        Args:
+            timeout: Optional timeout in seconds. If None, waits indefinitely.
+
+        Raises:
+            RuntimeError: If no wait_for_completion function was provided.
+            TimeoutError: If timeout is exceeded.
+        """
+        if not self.wait_for_completion:
+            raise RuntimeError(
+                f"Cannot wait for job {self.job_id}: no wait function provided",
+            )
+
+        # TODO: Add timeout support
+        if timeout:
+            import warnings
+
+            warnings.warn("Timeout parameter not yet implemented", UserWarning, stacklevel=2)
+
+        self.wait_for_completion()

flowyml/core/versioning.py

@@ -3,7 +3,7 @@
 import json
 import hashlib
 from pathlib import Path
-from typing import Any, Never
+from typing import Any, NoReturn
 from datetime import datetime
 from dataclasses import dataclass, asdict
 
@@ -204,7 +204,7 @@ class VersionedPipeline:
             if changes["modified"]:
                 pass
 
-    def rollback(self, version: str) -> Never:
+    def rollback(self, version: str) -> NoReturn:
         """Rollback to a previous version (not implemented - would need to reconstruct pipeline)."""
         raise NotImplementedError("Rollback requires pipeline reconstruction from saved state")
 

flowyml/integrations/keras.py

@@ -2,6 +2,7 @@
 
 from pathlib import Path
 from datetime import datetime
+import uuid
 
 try:
     from tensorflow import keras
@@ -16,28 +17,38 @@ from flowyml.storage.metadata import SQLiteMetadataStore
 
 
 class FlowymlKerasCallback(keras.callbacks.Callback if keras else object):
-    """Keras callback for flowyml tracking.
+    """Keras callback for flowyml tracking with automatic training history collection.
 
     Automatically logs:
-    - Training metrics (loss, accuracy, etc.)
-    - Model checkpoints (optional)
+    - Training metrics (loss, accuracy, etc.) per epoch
+    - Complete training history for visualization
+    - Model checkpoints with training history attached
     - Model architecture
     - Training parameters
+
+    Example:
+        >>> from flowyml.integrations.keras import FlowymlKerasCallback
+        >>> callback = FlowymlKerasCallback(experiment_name="my-experiment", project="my-project", auto_log_history=True)
+        >>> model.fit(x_train, y_train, epochs=50, callbacks=[callback])
     """
 
     def __init__(
         self,
         experiment_name: str,
         run_name: str | None = None,
+        project: str | None = None,
         log_model: bool = True,
         log_every_epoch: bool = True,
+        auto_log_history: bool = True,
         metadata_store: SQLiteMetadataStore | None = None,
     ):
         """Args:
         experiment_name: Name of the experiment
         run_name: Optional run name (defaults to timestamp)
+        project: Project name for organizing runs
         log_model: Whether to save the model as an artifact
         log_every_epoch: Whether to log metrics every epoch
+        auto_log_history: Whether to automatically collect training history
         metadata_store: Optional metadata store override.
         """
         if keras is None:
@@ -46,8 +57,10 @@ class FlowymlKerasCallback(keras.callbacks.Callback if keras else object):
         super().__init__()
         self.experiment_name = experiment_name
         self.run_name = run_name or datetime.now().strftime("run_%Y%m%d_%H%M%S")
+        self.project = project
         self.log_model = log_model
         self.log_every_epoch = log_every_epoch
+        self.auto_log_history = auto_log_history
 
         self.metadata_store = metadata_store or SQLiteMetadataStore()
 
@@ -57,6 +70,16 @@ class FlowymlKerasCallback(keras.callbacks.Callback if keras else object):
         # Track params
         self.params_logged = False
 
+        # Training history accumulator
+        self.training_history = {
+            "epochs": [],
+            "train_loss": [],
+            "train_accuracy": [],
+            "val_loss": [],
+            "val_accuracy": [],
+        }
+        self.custom_metrics = set()
+
     def on_train_begin(self, logs=None) -> None:
         """Log initial parameters."""
         if not self.params_logged:
@@ -85,6 +108,7 @@ class FlowymlKerasCallback(keras.callbacks.Callback if keras else object):
                     "name": "model_architecture",
                     "type": "json",
                     "run_id": self.run_name,
+                    "project": self.project,
                     "value": model_json,
                     "created_at": datetime.now().isoformat(),
                 },
@@ -93,28 +117,52 @@ class FlowymlKerasCallback(keras.callbacks.Callback if keras else object):
             self.params_logged = True
 
     def on_epoch_end(self, epoch, logs=None) -> None:
-        """Log metrics at the end of each epoch."""
-        if self.log_every_epoch and logs:
-            # Log metrics to DB
-            for k, v in logs.items():
-                self.metadata_store.save_metric(
+        """Log metrics at the end of each epoch and accumulate training history."""
+        if logs:
+            # Log metrics to DB (existing behavior)
+            if self.log_every_epoch:
+                for k, v in logs.items():
+                    self.metadata_store.save_metric(
+                        run_id=self.run_name,
+                        name=k,
+                        value=float(v),
+                        step=epoch,
+                    )
+
+                # Update experiment run
+                self.metadata_store.log_experiment_run(
+                    experiment_id=self.experiment_name,
                     run_id=self.run_name,
-                    name=k,
-                    value=float(v),
-                    step=epoch,
+                    metrics=logs,
                 )
 
-            # Update experiment run
-            self.metadata_store.log_experiment_run(
-                experiment_id=self.experiment_name,
-                run_id=self.run_name,
-                metrics=logs,
-            )
+            # Accumulate training history (NEW)
+            if self.auto_log_history:
+                self.training_history["epochs"].append(epoch + 1)  # 1-indexed
+
+                # Standard metrics
+                if "loss" in logs:
+                    self.training_history["train_loss"].append(float(logs["loss"]))
+                if "accuracy" in logs or "acc" in logs:
+                    acc_key = "accuracy" if "accuracy" in logs else "acc"
+                    self.training_history["train_accuracy"].append(float(logs[acc_key]))
+                if "val_loss" in logs:
+                    self.training_history["val_loss"].append(float(logs["val_loss"]))
+                if "val_accuracy" in logs or "val_acc" in logs:
+                    val_acc_key = "val_accuracy" if "val_accuracy" in logs else "val_acc"
+                    self.training_history["val_accuracy"].append(float(logs[val_acc_key]))
+
+                # Custom metrics
+                for metric_name, value in logs.items():
+                    if metric_name not in ["loss", "accuracy", "acc", "val_loss", "val_accuracy", "val_acc"]:
+                        if metric_name not in self.custom_metrics:
+                            self.custom_metrics.add(metric_name)
+                            self.training_history[metric_name] = []
+                        self.training_history[metric_name].append(float(value))
 
     def on_train_end(self, logs=None) -> None:
-        """Save model at the end of training."""
+        """Save model at the end of training with complete training history."""
         if self.log_model:
-            # Create artifacts directory
             # Create artifacts directory
             artifact_dir = Path(f".flowyml/artifacts/{self.run_name}")
             artifact_dir.mkdir(parents=True, exist_ok=True)
@@ -122,13 +170,38 @@ class FlowymlKerasCallback(keras.callbacks.Callback if keras else object):
             model_path = artifact_dir / "model.keras"
             self.model.save(model_path)
 
+            # Clean up empty history lists
+            cleaned_history = {
+                k: v
+                for k, v in self.training_history.items()
+                if v  # Only include non-empty lists
+            }
+
+            # Calculate final metrics
+            final_metrics = {}
+            if "train_loss" in cleaned_history and cleaned_history["train_loss"]:
+                final_metrics["loss"] = cleaned_history["train_loss"][-1]
+            if "train_accuracy" in cleaned_history and cleaned_history["train_accuracy"]:
+                final_metrics["accuracy"] = cleaned_history["train_accuracy"][-1]
+
+            # Save model artifact with training history
+            artifact_id = str(uuid.uuid4())
             self.metadata_store.save_artifact(
-                artifact_id=f"{self.run_name}_model",
+                artifact_id=artifact_id,
                 metadata={
-                    "name": "trained_model",
-                    "type": "keras_model",
+                    "artifact_id": artifact_id,
+                    "name": f"model-{self.run_name}",
+                    "type": "model",
                     "run_id": self.run_name,
+                    "project": self.project,
                     "path": str(model_path.resolve()),
+                    "properties": {
+                        "framework": "keras",
+                        "epochs_trained": len(cleaned_history.get("epochs", [])),
+                        "optimizer": str(self.model.optimizer.__class__.__name__),
+                        **final_metrics,
+                    },
+                    "training_history": cleaned_history,  # NEW: UI will display this!
                     "created_at": datetime.now().isoformat(),
                 },
             )

flowyml/monitoring/alerts.py

@@ -1,6 +1,6 @@
 from dataclasses import dataclass, field
 from enum import Enum
-from typing import Any, Never
+from typing import Any, NoReturn
 from datetime import datetime
 import logging
 
@@ -24,7 +24,7 @@ class Alert:
 
 
 class AlertHandler:
-    def handle(self, alert: Alert) -> Never:
+    def handle(self, alert: Alert) -> NoReturn:
         raise NotImplementedError
 
 

flowyml/stacks/__init__.py

@@ -2,6 +2,9 @@
 
 from flowyml.stacks.base import Stack, StackConfig
 from flowyml.stacks.local import LocalStack
+from flowyml.stacks.gcp import GCPStack, VertexAIOrchestrator, GCSArtifactStore, GCRContainerRegistry
+from flowyml.stacks.aws import AWSStack, AWSBatchOrchestrator, S3ArtifactStore, ECRContainerRegistry
+from flowyml.stacks.azure import AzureMLStack, AzureMLOrchestrator, AzureBlobArtifactStore, ACRContainerRegistry
 from flowyml.stacks.components import (
     ResourceConfig,
     DockerConfig,
@@ -15,6 +18,18 @@ __all__ = [
     "Stack",
     "StackConfig",
     "LocalStack",
+    "GCPStack",
+    "AWSStack",
+    "AzureMLStack",
+    "VertexAIOrchestrator",
+    "AWSBatchOrchestrator",
+    "AzureMLOrchestrator",
+    "GCSArtifactStore",
+    "S3ArtifactStore",
+    "AzureBlobArtifactStore",
+    "GCRContainerRegistry",
+    "ECRContainerRegistry",
+    "ACRContainerRegistry",
     "ResourceConfig",
     "DockerConfig",
     "Orchestrator",