Skip to content

Azure Serverless

Scope

Azure serverless compute and event-driven architecture. Covers Azure Functions (Consumption, Flex Consumption, Premium, Dedicated plans), Durable Functions, Logic Apps, Event Grid, Service Bus, Event Hubs, and idempotency patterns across at-least-once delivery services.

Checklist

  • [Critical] Is the Azure Functions hosting plan selected based on workload characteristics -- Consumption (pay-per-execution, cold starts) vs Flex Consumption (per-function scaling, VNet integration, fast scaling with always-ready instances) vs Premium (pre-warmed instances, full VNet integration) vs Dedicated (App Service plan, predictable cost)?
  • [Recommended] Are Durable Functions used for stateful orchestration patterns -- function chaining, fan-out/fan-in, async HTTP APIs, monitoring, and human interaction workflows?
  • [Recommended] Are Azure Functions bindings and triggers configured to minimize boilerplate -- input/output bindings for Blob Storage, Cosmos DB, Service Bus, and Event Hubs instead of SDK calls?
  • [Recommended] Is Azure Logic Apps (Standard or Consumption) deployed for low-code workflow automation with 400+ managed connectors for SaaS, on-premises, and Azure service integration?
  • [Recommended] Is Azure Event Grid configured as the event backbone with system topics for Azure resource events and custom topics for application-domain events?
  • [Recommended] Is Azure Service Bus deployed for enterprise messaging with queues (point-to-point) and topics/subscriptions (pub/sub) with dead-letter queues, sessions, and duplicate detection enabled?
  • [Critical] Is Azure Event Hubs configured for high-throughput streaming ingestion with appropriate partition count (can be increased up to 1024 in Premium/Dedicated tiers but cannot be decreased) and consumer groups per downstream processor?
  • [Recommended] Is Event Hubs Capture enabled to automatically archive streaming events to Blob Storage or Data Lake Gen2 in Avro format for long-term retention and batch processing?
  • [Critical] Are retry policies and dead-letter destinations configured for all messaging components -- Event Grid (retry + dead-letter to storage), Service Bus (max delivery count + dead-letter queue), Functions (retry policy)?
  • [Recommended] Is VNet integration enabled for Premium or Flex Consumption Azure Functions and Logic Apps Standard to access private endpoints and on-premises resources through VNet?
  • [Recommended] Are function app slots configured for production deployments with slot-based deployment swaps and pre-warm verification?
  • [Optional] Are Event Grid event subscriptions filtered with subject and advanced filters to minimize unnecessary function invocations and reduce cost?
  • [Recommended] Is Service Bus configured with appropriate tier -- Basic (queues only) vs Standard (queues + topics, 256KB messages) vs Premium (dedicated capacity, 100MB messages, VNet integration)?
  • [Critical] Are idempotency patterns implemented across all event-driven functions, handling at-least-once delivery guarantees from Event Grid, Service Bus, and Event Hubs?

Why This Matters

Azure's serverless ecosystem is more fragmented than AWS Lambda + EventBridge, requiring architects to combine Azure Functions, Logic Apps, Event Grid, Service Bus, and Event Hubs into a cohesive event-driven architecture. Azure Functions Consumption plan has cold starts of 1-10 seconds (worse for Java and .NET), making the Premium plan necessary for latency-sensitive workloads. Event Grid and Service Bus serve different patterns -- Event Grid is optimized for reactive event distribution (high fan-out, low latency) while Service Bus provides transactional messaging guarantees (sessions, ordering, dead-letter). Event Hubs is purpose-built for streaming ingestion at millions of events per second but requires careful partition planning since partition count is immutable after creation.

Common Decisions (ADR Triggers)

  • Functions Consumption vs Flex Consumption vs Premium vs Dedicated -- zero-cost at idle with cold starts vs Flex Consumption (per-function scaling, VNet support, always-ready instances for reduced cold starts, pay-per-execution with provisioned baseline) vs Premium with pre-warmed instances ($0.173/vCPU/hr) vs App Service plan sharing with web apps
  • Event Grid vs Service Bus vs Event Hubs -- reactive event distribution (fire-and-forget) vs transactional messaging (guaranteed delivery, ordering) vs streaming ingestion (high-throughput, partitioned)
  • Durable Functions vs Logic Apps -- code-first orchestration in C#/Python/JS with unit testing vs low-code visual designer with managed connectors and no-code transforms
  • Logic Apps Consumption vs Standard -- pay-per-action serverless vs single-tenant with VNet support, local development, and workflow hosting on App Service Environment
  • Service Bus sessions vs partitions -- ordered message processing per session ID (FIFO per group) vs throughput scaling across partitions (no ordering guarantee)
  • Event Hubs partition count -- more partitions enable more parallel consumers but increase cost; 4-32 for most workloads; Premium/Dedicated tiers allow increasing partition count up to 1024 after creation but partitions cannot be decreased (Kafka-compatible API available)
  • Function language runtime -- .NET (fastest cold start, best integration) vs Python/Node.js (faster development) vs Java (slowest cold start on Consumption, use Premium)

Reference Architectures

See Also

  • general/compute.md -- General compute patterns including serverless selection criteria
  • providers/azure/messaging.md -- Detailed coverage of Service Bus, Event Hubs, and Event Grid
  • providers/azure/containers.md -- Azure Container Apps as a serverless container alternative
  • providers/azure/networking.md -- VNet integration for Premium and Flex Consumption Functions