Serverless Workflows

Concepts

• Formalised processes
• Connection of two or more steps

Connection patterns

• Sequence / chain
• Parallelism
• Conditions
• Split and join
• Complex state machine

Description language

• BPMN
• BPEL

Primitives (explicit)

• ordering
• control flow
• selectors
• data transformation

Features (implicit)

• DAGs (validation, optimisation, visualisation, editing)
• checkpointing
• monitoring, notification
• connectors, integration

Models

Traditional model: Engine -> Queue -> Worker VMs -> Storage

Queue model: Engine -> Queue -> Worker Cloud Functions -> Storage

Direct executor model: Engine -> Worker Cloud Functions -> Storage

Bridge model: Engine -> Queue -> Worker VMs + Bridge to Worker Cloud Functions -> Storage

Decentralized model: Cloud Functions -> Storage

Systems and languages

• declarative vs imperative
• JSON vs YAML vs JavaScript
• identification of functions
• handling of errors, I/O

Metrics	Amazon Step Functions	IBM Composer	Azure Durable Functions
ST-safe (composition as functions)	No	Yes	Yes
Programming model	DSL(JSON)	Composition library (JS)	async/await (C#)
Reflective API	Yes(limited)	No	Yes
Parallel execution support	Yes(limited)	No	Yes(limited)
Software packaging and repositories	Yes	Yes	Yes(no repo)
Billing model	$0.025 per 1000 state transitions	Orchestrator function execution	Orchestrator function execution + storage cost
Architecture	Synchronous client scheduler	Reactive scheduler	Reactive scheduler

Optimisation

• User (workflow author) optimisation: Code
• System (workflow engine) optimisation: Processing

Workflow preparation/rewriting: Achieve series-parallel structure for polynomial execution time from directed acyclic graphs

Function execution

• Resource re-use

  • Keep function alive (full function reuse) vs spawn new functions (full isolation)
  • Long keep-alive -> more warm executions / short keep-alives -> less idle resources

• Runtime pooling

  • large pool / handle high concurrency -> increases resource overhead / better performance
  • minimal pool / fast pool exhaustion -> minimize pool; less idle resources / minimize cost

Workflow execution-specific

• Prefetching: Fetch function source upfront

  • Higher latency -> less storage cost
  • Lower latency -> more storage cost

• Prewarming: Predictive function execution (via runtime analysis and function stats)

  • Optimistic prewarning / low treshold -> more performance
  • Pessimistic prewarning / high threshold -> less cost
  • Other domains: CPU branch predictor, proactive autoscalers, predictive caches