/workflow-orchestration-patterns

workflow-orchestration-patterns — detailed patterns and worked examples

Critical Design Decision: Workflows vs Activities

The Fundamental Rule (Source: temporal.io/blog/workflow-engine-principles):

• Workflows = Orchestration logic and decision-making
• Activities = External interactions (APIs, databases, network calls)

Workflows (Orchestration)

Characteristics:

• Contain business logic and coordination
• MUST be deterministic (same inputs → same outputs)
• Cannot perform direct external calls
• State automatically preserved across failures
• Can run for years despite infrastructure failures

Example workflow tasks:

• Decide which steps to execute
• Handle compensation logic
• Manage timeouts and retries
• Coordinate child workflows

Activities (External Interactions)

Characteristics:

• Handle all external system interactions
• Can be non-deterministic (API calls, DB writes)
• Include built-in timeouts and retry logic
• Must be idempotent (calling N times = calling once)
• Short-lived (seconds to minutes typically)

Example activity tasks:

• Call payment gateway API
• Write to database
• Send emails or notifications
• Query external services

Design Decision Framework

Does it touch external systems? → Activity
Is it orchestration/decision logic? → Workflow

Core Workflow Patterns

1. Saga Pattern with Compensation

Purpose: Implement distributed transactions with rollback capability

Pattern (Source: temporal.io/blog/compensating-actions-part-of-a-complete-breakfast-with-sagas):

For each step:
  1. Register compensation BEFORE executing
  2. Execute the step (via activity)
  3. On failure, run all compensations in reverse order (LIFO)

Example: Payment Workflow

1. Reserve inventory (compensation: release inventory)
2. Charge payment (compensation: refund payment)
3. Fulfill order (compensation: cancel fulfillment)

Critical Requirements:

• Compensations must be idempotent
• Register compensation BEFORE executing step
• Run compensations in reverse order
• Handle partial failures gracefully

2. Entity Workflows (Actor Model)

Purpose: Long-lived workflow representing single entity instance

Pattern (Source: docs.temporal.io/evaluate/use-cases-design-patterns):

• One workflow execution = one entity (cart, account, inventory item)
• Workflow persists for entity lifetime
• Receives signals for state changes
• Supports queries for current state

Example Use Cases:

• Shopping cart (add items, checkout, expiration)
• Bank account (deposits, withdrawals, balance checks)
• Product inventory (stock updates, reservations)

Benefits:

• Encapsulates entity behavior
• Guarantees consistency per entity
• Natural event sourcing

3. Fan-Out/Fan-In (Parallel Execution)

Purpose: Execute multiple tasks in parallel, aggregate results

Pattern:

• Spawn child workflows or parallel activities
• Wait for all to complete
• Aggregate results
• Handle partial failures

Scaling Rule (Source: temporal.io/blog/workflow-engine-principles):

• Don't scale individual workflows
• For 1M tasks: spawn 1K child workflows × 1K tasks each
• Keep each workflow bounded

4. Async Callback Pattern

Purpose: Wait for external event or human approval

Pattern:

• Workflow sends request and waits for signal
• External system processes asynchronously
• Sends signal to resume workflow
• Workflow continues with response

Use Cases:

• Human approval workflows
• Webhook callbacks
• Long-running external processes

State Management and Determinism

Automatic State Preservation

How Temporal Works (Source: docs.temporal.io/workflows):

• Complete program state preserved automatically
• Event History records every command and event
• Seamless recovery from crashes
• Applications restore pre-failure state

Determinism Constraints

Workflows Execute as State Machines:

• Replay behavior must be consistent
• Same inputs → identical outputs every time

Prohibited in Workflows (Source: docs.temporal.io/workflows):

• ❌ Threading, locks, synchronization primitives
• ❌ Random number generation (random())
• ❌ Global state or static variables
• ❌ System time (datetime.now())
• ❌ Direct file I/O or network calls
• ❌ Non-deterministic libraries

Allowed in Workflows:

• ✅ workflow.now() (deterministic time)
• ✅ workflow.random() (deterministic random)
• ✅ Pure functions and calculations
• ✅ Calling activities (non-deterministic operations)

Versioning Strategies

Challenge: Changing workflow code while old executions still running

Solutions:

1. Versioning API: Use workflow.get_version() for safe changes
2. New Workflow Type: Create new workflow, route new executions to it
3. Backward Compatibility: Ensure old events replay correctly

Resilience and Error Handling

Retry Policies

Default Behavior: Temporal retries activities forever

Configure Retry:

• Initial retry interval
• Backoff coefficient (exponential backoff)
• Maximum interval (cap retry delay)
• Maximum attempts (eventually fail)

Non-Retryable Errors:

• Invalid input (validation failures)
• Business rule violations
• Permanent failures (resource not found)

Idempotency Requirements

Why Critical (Source: docs.temporal.io/activities):

• Activities may execute multiple times
• Network failures trigger retries
• Duplicate execution must be safe

Implementation Strategies:

• Idempotency keys (deduplication)
• Check-then-act with unique constraints
• Upsert operations instead of insert
• Track processed request IDs

Activity Heartbeats

Purpose: Detect stalled long-running activities

Pattern:

• Activity sends periodic heartbeat
• Includes progress information
• Timeout if no heartbeat received
• Enables progress-based retry