Skip to content

Numaflow Core Concepts Overview

Numaflow is a Kubernetes-native platform for event processing at scale, purpose-built for creating event-driven applications, real-time stream processing pipelines, and serving systems. It provides a set of modular and composable building blocks to help developers process events and serve efficiently in cloud-native environments.

The following sections introduce the core concepts of Numaflow.

MonoVertex

MonoVertex is a lightweight way to develop event-driven applications in Numaflow. Each MonoVertex contains source, transformer, and sink within a single unit, making it ideal for simple event processing tasks. The entire unit scales as one, making it a lightweight option for handling simple event processing patterns.

Use Cases

  • Stateless event transformation
  • Filtering, mapping, or enriching event data
  • Microservice-style event handling (e.g., call external APIs per event)
  • Ingesting from or writing to external systems within a single unit

Pipeline

Pipeline is designed for developing real-time stream processing pipelines. It allows you to connect multiple vertices — each representing a processing step — to handle data. Pipelines can include multiple sources and sinks, enabling integration with diverse systems at both the input and output stages.

Each step in a pipeline can perform operations like transformation, routing, reduction, or aggregation, and is independently scalable. This makes Pipelines ideal for building robust, modular stream processing applications.

Use Cases

  • Aggregation and windowing (e.g., sum, count, reduce over time windows)
  • Routing events based on payload or metadata
  • Multi-step data processing and enrichment
  • Scalable data pipelines for analytics and monitoring

Serving

Serving enables request/response interaction with streaming systems, allowing external clients to send data and receive processed results in real time through interfaces like REST or Server-Sent Events (SSE).

Use Cases

  • Machine learning inference (e.g., send features, receive predictions)
  • Event-driven APIs backed by stream processing logic
  • Asynchronous request/response workflows over SSE or HTTP