Prefect

Prefect uses Python decorators (@flow, @task) to transform standard functions into orchestrated units with built-in state management. The execution engine wraps decorated functions to automatically track execution state (Pending, Running, Completed, Failed, Cached) through a state machine, persisting state transitions to the backend database. This enables resumability, retry logic, and observability without requiring explicit state handling in user code.

Prefect provides built-in retry logic via task decorators with exponential backoff, jitter, and max retry limits. Task-level caching uses a content-addressable key (based on task name, version, and input parameters) to skip re-execution of identical tasks within a configurable time window. Both features are configured declaratively in decorator arguments and enforced by the execution engine without requiring try-catch blocks in user code.

The Prefect Client is a Python library that provides programmatic access to the Prefect server API, enabling custom integrations and automation. The client supports operations like creating deployments, triggering flow runs, querying run history, and managing blocks. It uses async/await patterns for non-blocking I/O and supports both Prefect Cloud and self-hosted servers. The client is used internally by the CLI and can be imported directly into user code for custom workflows.

Prefect provides a web-based dashboard (React UI v2) for monitoring flow and task execution in real-time. The dashboard displays flow run status, task execution timelines, logs, and state transitions. It supports filtering and searching by flow name, deployment, run status, and time range. The dashboard connects to the Prefect server via WebSocket for real-time updates, eliminating the need to refresh the page to see new runs or status changes.

Prefect supports deploying the same flow to multiple environments (dev, staging, prod) with environment-specific configuration. Deployments can be parameterized with environment variables, work pool assignments, and schedule overrides. The prefect.yaml configuration file supports variable substitution and environment-specific profiles, enabling a single flow definition to be deployed to multiple environments without code changes. The system also supports deployment of flow code from version control (GitHub, GitLab) with automatic updates when code is pushed.

Prefect supports concurrency limits at multiple levels: global (server-wide), per-work-pool, and per-task. Concurrency limits are enforced by the execution engine, which queues task runs and releases them as capacity becomes available. Task-level concurrency limits can be set via the @task decorator, preventing a specific task from running more than N times concurrently. Work pool concurrency limits control the total number of concurrent tasks across all flows using that pool. The system uses a token-bucket algorithm to enforce limits fairly.

Prefect decouples task scheduling from execution through a Worker/Work Pool abstraction. The server enqueues task runs to named Work Pools; distributed Workers poll their assigned pool and execute tasks in isolated environments (Docker containers, Kubernetes pods, or local processes). Workers report execution status back to the server, enabling horizontal scaling and multi-cloud deployments without modifying pipeline code. The architecture uses a pull-based model (workers pull work) rather than push (server pushes work), reducing firewall complexity.

Prefect's Events system enables workflows to react to external events (deployment status changes, task failures, custom events) via Automations. Automations are trigger-action rules defined in the UI or API that listen for events matching a filter (e.g., 'task.failed') and execute actions (pause flow, trigger deployment, send notification). Events are emitted by the execution engine and can be published by external systems via the Events API, creating a reactive orchestration model where workflows respond to runtime conditions rather than following a static schedule.

Prefect's Deployment system packages flow code, dependencies, and configuration into versioned artifacts that can be deployed to remote environments. The prefect deploy CLI command builds a deployment from a flow definition (Python file or module), captures the flow code and metadata, and registers it with the server. Deployments support multiple versions, allowing A/B testing or gradual rollouts. The system uses a 'flow definition' (Python code) + 'deployment configuration' (YAML or Python) pattern, separating code from infrastructure concerns.

Prefect supports declarative scheduling via cron expressions or interval-based triggers (e.g., every 5 minutes) defined in deployment configuration. The server maintains a schedule state machine that determines when to enqueue flow runs based on the schedule definition. Schedules support timezone awareness, catchup behavior (whether to run missed schedules if the server was down), and anchor times for interval-based schedules. The scheduling engine is decoupled from execution; the server enqueues runs to work pools, and workers execute them independently.

Prefect's logging system automatically captures logs from task and flow execution with contextual metadata (flow name, task name, run ID, worker name). Logs are structured as JSON and stored in the backend database, enabling rich querying and filtering via the UI or API. The logging system uses Python's standard logging module under the hood but injects Prefect context (via thread-local storage) to automatically tag logs with execution metadata without requiring explicit log formatting in user code.

Prefect's Block system provides a declarative way to manage credentials, API keys, and configuration values as reusable, encrypted objects stored in the server. Blocks are defined as Python classes (inheriting from Block base class) and can be instantiated via the UI or API. Tasks and flows reference blocks by name, and the execution engine injects the block's values at runtime. Blocks support encryption at rest and can be scoped to specific work pools or teams, enabling secure multi-tenant credential management without embedding secrets in code.

Prefect's task mapping feature enables a single task to be executed multiple times with different parameters, with results automatically aggregated. Mapping is declared using the .map() method on a task, which accepts an iterable of parameters and returns a list of futures. The execution engine creates a separate task run for each item in the iterable, executing them in parallel (subject to concurrency limits). Results are automatically collected and can be passed to downstream tasks, enabling fan-out/fan-in patterns without explicit loop logic.

Prefect supports conditional task execution via the if/else pattern on task results. Tasks return State objects that can be inspected to determine whether to execute downstream tasks. The system also supports dynamic DAG construction where the set of tasks to execute is determined at runtime based on upstream results. This is achieved through Python's native control flow (if/else, for loops) within flow functions, allowing the DAG to be constructed dynamically rather than statically.