Step Orchestration

Step Orchestration is a powerful workflow pattern that enables you to break down complex operations into discrete, manageable steps. Each step can have its own execution logic, progress weight, skip conditions, and rollback behavior.

Why Use Step Orchestration?

Benefits

Benefit	Description
Separation of Concerns	Each step handles one specific task, making code easier to understand and maintain
Reusability	Steps can be shared across different workflows and actions
Testability	Individual steps can be unit tested in isolation
Progress Tracking	Weighted progress calculation provides accurate progress reporting
Error Recovery	Automatic rollback on failure cleans up partial operations
Flexibility	Steps can be inserted, removed, or reordered dynamically
Observability	Built-in logging and timing utilities for debugging

When to Use

Step orchestration is ideal for:

• Multi-phase operations: Upload workflows (init -> validate -> upload -> cleanup)
• Long-running tasks: Training pipelines (load data -> train -> save model)
• Operations requiring cleanup: File processing with temp file cleanup on failure
• Composable workflows: Building workflows from reusable step components

When NOT to Use

Step orchestration adds overhead. For simple operations, implement execute() directly in your action class instead of using setup_steps():

• Single-phase operations (e.g., simple data fetch)
• Operations without cleanup requirements
• Workflows with fewer than 3 logical phases

Core Concepts

Step

A step is a discrete unit of work in a workflow. Each step:

• Has a unique name for identification
• Specifies a progress weight (0.0 to 1.0) for progress calculation
• Implements execute() to perform the actual work
• Can optionally implement can_skip() and rollback()

from dataclasses import dataclass, field
from synapse_sdk.plugins.steps import BaseStep, BaseStepContext, StepResult

# Define a custom context with the fields your workflow needs
@dataclass
class UploadContext(BaseStepContext):
    params: dict = field(default_factory=dict)
    files: list[dict] = field(default_factory=list)  # Files to validate/upload

class ValidateFilesStep(BaseStep[UploadContext]):
    @property
    def name(self) -> str:
        return 'validate_files'

    @property
    def progress_weight(self) -> float:
        return 0.1  # 10% of total workflow progress

    def execute(self, context: UploadContext) -> StepResult:
        invalid_files = []
        for file in context.files:
            if not self._is_valid(file):
                invalid_files.append(file)

        if invalid_files:
            return StepResult(
                success=False,
                error=f'Invalid files: {invalid_files}'
            )

        return StepResult(success=True, data={'validated': len(context.files)})

    def can_skip(self, context: UploadContext) -> bool:
        # Skip validation if explicitly disabled
        return context.params.get('skip_validation', False)

    def rollback(self, context: UploadContext, result: StepResult) -> None:
        # Nothing to rollback for validation
        pass

    def _is_valid(self, file: dict) -> bool:
        # Validation logic
        return file.get('size', 0) > 0

StepResult

Every step returns a StepResult containing:

Field	Type	Description
success	bool	Whether the step completed successfully
data	dict[str, Any]	Output data from the step
error	`str	None`
rollback_data	dict[str, Any]	Data needed for rollback
skipped	bool	Whether the step was skipped
timestamp	datetime	When the step completed

# Success result
return StepResult(success=True, data={'files_processed': 10})

# Failure result
return StepResult(success=False, error='Connection timeout')

# Result with rollback data
return StepResult(
    success=True,
    data={'uploaded_ids': [1, 2, 3]},
    rollback_data={'uploaded_ids': [1, 2, 3]}  # For cleanup on failure
)

Context

Context is a shared state object passed between all steps. It:

• Extends BaseStepContext with workflow-specific fields
• Provides access to RuntimeContext for logging/progress
• Accumulates data as steps execute
• Tracks step results and errors

from dataclasses import dataclass, field
from synapse_sdk.plugins.steps import BaseStepContext

# Custom context example - extend BaseStepContext with your workflow-specific fields
@dataclass
class MyUploadContext(BaseStepContext):
    """Shared context for upload workflow."""
    # Workflow parameters
    params: dict = field(default_factory=dict)

    # Accumulated state
    files_to_upload: list[str] = field(default_factory=list)
    uploaded_files: list[dict] = field(default_factory=list)
    total_bytes: int = 0

    # Helper property to access backend client
    @property
    def client(self):
        return self.runtime_ctx.client

팁

The SDK provides pre-built contexts like UploadContext, TrainContext, and ExportContext with common fields. You can use these directly or create custom contexts as shown above.

Registry

The StepRegistry manages an ordered list of steps:

from synapse_sdk.plugins.steps import StepRegistry

registry = StepRegistry[UploadContext]()

# Register steps in order
registry.register(InitializeStep())
registry.register(ValidateStep())
registry.register(UploadStep())
registry.register(CleanupStep())

# Dynamic step manipulation
registry.insert_before('upload', CompressionStep())  # Add compression before upload
registry.insert_after('validate', SanitizeStep())    # Add sanitization after validate
registry.unregister('cleanup')                        # Remove cleanup step

# Get step count and total weight
print(f"Steps: {len(registry)}")
print(f"Total weight: {registry.total_weight}")

Orchestrator

The Orchestrator executes steps and handles:

• Sequential step execution
• Weighted progress tracking
• Automatic rollback on failure
• Skip condition evaluation

from synapse_sdk.plugins.steps import Orchestrator

orchestrator = Orchestrator(
    registry=registry,
    context=context,
    progress_callback=lambda current, total: print(f'{current}/{total}%')
)

try:
    result = orchestrator.execute()
    # {'success': True, 'steps_executed': 4, 'steps_total': 4}
except RuntimeError as e:
    # Step failed, rollback was performed
    print(f"Workflow failed: {e}")

Progress Tracking

Progress is calculated based on step weights:

class Step1(BaseStep[MyContext]):
    @property
    def progress_weight(self) -> float:
        return 0.2  # 20%

class Step2(BaseStep[MyContext]):
    @property
    def progress_weight(self) -> float:
        return 0.6  # 60%

class Step3(BaseStep[MyContext]):
    @property
    def progress_weight(self) -> float:
        return 0.2  # 20%

# Progress updates:
# After Step1: 20%
# After Step2: 80%
# After Step3: 100%

The orchestrator normalizes weights, so they don't need to sum exactly to 1.0.

Rollback Behavior

When a step fails, the orchestrator:

1. Stops execution immediately
2. Calls rollback() on all previously executed steps in reverse order
3. Raises RuntimeError with the failure details

# Using the custom UploadContext defined earlier (with files field)
class UploadFilesStep(BaseStep[UploadContext]):
    def execute(self, context: UploadContext) -> StepResult:
        uploaded_ids = []
        for file in context.files:  # context.files from custom UploadContext
            file_id = self._upload(file)
            uploaded_ids.append(file_id)

        return StepResult(
            success=True,
            rollback_data={'uploaded_ids': uploaded_ids}
        )

    def rollback(self, context: UploadContext, result: StepResult) -> None:
        # Clean up uploaded files
        for file_id in result.rollback_data.get('uploaded_ids', []):
            try:
                self._delete(file_id)
            except Exception:
                context.errors.append(f'Failed to rollback file {file_id}')

Utility Steps

The SDK provides utility step wrappers for common patterns:

LoggingStep

Wraps a step with start/end logging. The wrapped step's name is prefixed with logged_:

from synapse_sdk.plugins.steps import LoggingStep

# Wrap any step with logging
logged_step = LoggingStep(UploadFilesStep())
registry.register(logged_step)
# Note: Step name becomes 'logged_upload_files'

# Logs:
# step_start {'step': 'upload_files'}
# step_end {'step': 'upload_files', 'elapsed': 1.234, 'success': True, 'skipped': False}

TimingStep

Measures step execution duration. The wrapped step's name is prefixed with timed_:

from synapse_sdk.plugins.steps import TimingStep

timed_step = TimingStep(ProcessDataStep())
registry.register(timed_step)
# Note: Step name becomes 'timed_process_data'

# Result includes duration:
# result.data['duration_seconds'] = 1.234567

ValidationStep

Validates context state before proceeding:

from synapse_sdk.plugins.steps import ValidationStep

# Validator function receives the context and returns (is_valid, error_message)
def check_files_exist(context: MyUploadContext) -> tuple[bool, str | None]:
    if not context.files_to_upload:
        return False, 'No files to upload'
    return True, None

registry.register(ValidationStep(
    validator=check_files_exist,
    name='validate_files_exist',
    progress_weight=0.05
))

Integration with Actions

All base action classes support optional step-based execution via setup_steps():

Upload Action

from synapse_sdk.plugins import BaseUploadAction
from synapse_sdk.plugins.actions.upload import UploadContext

class MyUploadAction(BaseUploadAction[UploadParams]):
    def setup_steps(self, registry: StepRegistry[UploadContext]) -> None:
        registry.register(InitStorageStep())
        registry.register(OrganizeFilesStep())
        registry.register(UploadFilesStep())
        registry.register(GenerateMetadataStep())
        registry.register(CleanupStep())

Train Action

from synapse_sdk.plugins import BaseTrainAction
from synapse_sdk.plugins.actions.train import TrainContext

class MyTrainAction(BaseTrainAction[TrainParams]):
    def setup_steps(self, registry: StepRegistry[TrainContext]) -> None:
        registry.register(LoadDatasetStep())     # 20%
        registry.register(TrainModelStep())       # 60%
        registry.register(UploadModelStep())      # 20%

    # If setup_steps() is not overridden or registers no steps,
    # the action uses simple execute() mode instead

Export Action

from synapse_sdk.plugins import BaseExportAction
from synapse_sdk.plugins.actions.export import ExportContext

class MyExportAction(BaseExportAction[ExportParams]):
    def setup_steps(self, registry: StepRegistry[ExportContext]) -> None:
        registry.register(FetchResultsStep())
        registry.register(ProcessBatchStep())
        registry.register(WriteOutputStep())

Complete Example

Here's a complete example of a file upload workflow:

from dataclasses import dataclass, field
from pathlib import Path
from pydantic import BaseModel
from synapse_sdk.plugins import BaseUploadAction
from synapse_sdk.plugins.steps import (
    BaseStep, StepResult, StepRegistry, BaseStepContext, LoggingStep
)

# Define params model
class FileUploadParams(BaseModel):
    source_path: str

# Define context
@dataclass
class FileUploadContext(BaseStepContext):
    source_path: Path | None = None
    files: list[Path] = field(default_factory=list)
    uploaded_ids: list[int] = field(default_factory=list)

    @property
    def client(self):
        """Access backend client from runtime context."""
        return self.runtime_ctx.client

# Define steps
class DiscoverFilesStep(BaseStep[FileUploadContext]):
    @property
    def name(self) -> str:
        return 'discover_files'

    @property
    def progress_weight(self) -> float:
        return 0.1

    def execute(self, context: FileUploadContext) -> StepResult:
        if not context.source_path or not context.source_path.exists():
            return StepResult(success=False, error='Source path not found')

        context.files = list(context.source_path.glob('**/*'))
        context.files = [f for f in context.files if f.is_file()]

        if not context.files:
            return StepResult(success=False, error='No files found')

        return StepResult(success=True, data={'file_count': len(context.files)})

class UploadFilesStep(BaseStep[FileUploadContext]):
    @property
    def name(self) -> str:
        return 'upload_files'

    @property
    def progress_weight(self) -> float:
        return 0.8

    def execute(self, context: FileUploadContext) -> StepResult:
        for i, file in enumerate(context.files):
            # Upload each file
            file_id = context.client.upload_file(file)
            context.uploaded_ids.append(file_id)

            # Update progress within step
            progress = (i + 1) / len(context.files)
            context.set_progress(int(progress * 100), 100, 'upload')

        return StepResult(
            success=True,
            rollback_data={'uploaded_ids': context.uploaded_ids.copy()}
        )

    def rollback(self, context: FileUploadContext, result: StepResult) -> None:
        for file_id in result.rollback_data.get('uploaded_ids', []):
            try:
                context.client.delete_file(file_id)
            except Exception:
                context.errors.append(f'Failed to delete file {file_id}')

class FinalizeStep(BaseStep[FileUploadContext]):
    @property
    def name(self) -> str:
        return 'finalize'

    @property
    def progress_weight(self) -> float:
        return 0.1

    def execute(self, context: FileUploadContext) -> StepResult:
        context.log('upload_complete', {
            'file_count': len(context.uploaded_ids),
            'file_ids': context.uploaded_ids
        })
        return StepResult(success=True)

# Use in action
class FileUploadAction(BaseUploadAction[FileUploadParams]):
    def setup_steps(self, registry: StepRegistry[FileUploadContext]) -> None:
        # Wrap steps with logging for debugging
        registry.register(LoggingStep(DiscoverFilesStep()))
        registry.register(LoggingStep(UploadFilesStep()))
        registry.register(LoggingStep(FinalizeStep()))

    def create_context(self) -> FileUploadContext:
        return FileUploadContext(
            runtime_ctx=self.ctx,
            source_path=Path(self.params.source_path)
        )

Best Practices

1. Keep Steps Focused

Each step should do one thing well:

# Good: Focused steps
class ValidateFilesStep(BaseStep): ...
class CompressFilesStep(BaseStep): ...
class UploadFilesStep(BaseStep): ...

# Bad: Monolithic step
class ProcessEverythingStep(BaseStep): ...  # Does validation, compression, and upload

2. Use Meaningful Progress Weights

Assign weights based on actual time/complexity:

# Good: Weights reflect actual time distribution
LoadDataStep:    0.1   # Quick file read
TrainModelStep:  0.8   # Long training loop
SaveModelStep:   0.1   # Quick save

# Bad: Equal weights don't reflect reality
LoadDataStep:    0.33
TrainModelStep:  0.33  # Training takes 10x longer!
SaveModelStep:   0.33

3. Implement Rollback for Destructive Steps

Any step that creates resources should clean them up on failure:

class CreateResourcesStep(BaseStep):
    def execute(self, context) -> StepResult:
        resource_id = create_resource()
        return StepResult(
            success=True,
            rollback_data={'resource_id': resource_id}
        )

    def rollback(self, context, result) -> None:
        resource_id = result.rollback_data.get('resource_id')
        if resource_id:
            delete_resource(resource_id)

4. Use can_skip() for Conditional Steps

class CompressionStep(BaseStep[FileUploadContext]):
    def can_skip(self, context: FileUploadContext) -> bool:
        # Skip if files are already compressed
        return all(f.suffix == '.gz' for f in context.files)

5. Log Important Events

Use context logging for debugging:

def execute(self, context) -> StepResult:
    context.log('step_progress', {'phase': 'starting', 'item_count': 100})
    # ... work ...
    context.log('step_progress', {'phase': 'complete', 'processed': 100})
    return StepResult(success=True)

API Reference

BaseStep[C]

Method/Property	Description
name: str	Unique step identifier (abstract property)
progress_weight: float	Relative progress weight 0.0-1.0 (abstract property)
execute(context: C) -> StepResult	Execute the step (abstract method)
can_skip(context: C) -> bool	Check if step can be skipped (default: False)
rollback(context: C, result: StepResult) -> None	Cleanup on failure (default: no-op)

StepResult

Field	Type	Default
success	bool	True
data	dict[str, Any]	{}
error	`str	None`
rollback_data	dict[str, Any]	{}
skipped	bool	False
timestamp	datetime	datetime.now()

StepRegistry[C]

Method	Description
register(step)	Add step to end of workflow
unregister(name)	Remove step by name
insert_before(name, step)	Insert step before another
insert_after(name, step)	Insert step after another
get_steps()	Get ordered list of steps
total_weight	Sum of all step weights
__len__()	Return number of registered steps (use len(registry))

Orchestrator[C]

Method	Description
__init__(registry, context, progress_callback=None)	Create orchestrator
execute() -> dict	Execute all steps with rollback

BaseStepContext

Field/Method	Description
runtime_ctx: RuntimeContext	Parent runtime context
step_results: list[StepResult]	Results from executed steps
errors: list[str]	Accumulated error messages
current_step: str | None	Name of currently executing step (set by Orchestrator)
current_step_order: int | None	0-based order of currently executing step (set by Orchestrator)
log(event, data, file=None)	Log via runtime context
set_progress(current, total, step=None)	Update progress (uses current_step if step is None)
set_metrics(value, step=None)	Set metrics (uses current_step if step is None)