Orcher Features
Orcher is a durable workflow engine with SDKs for Rust, Python and TypeScript. You write a workflow as ordinary async code; the engine journals each step in PostgreSQL, so a run that loses its worker resumes from its last completed step.
Durable Execution
Every step of a workflow is journaled before the workflow moves on. If a worker crashes mid-run, another worker replays the workflow from the journal and continues from the last completed step instead of starting over. The same holds across engine restarts and deploys.
- Automatic Journaling — Task results, timers and events are recorded without any checkpointing code
- Crash Recovery — Runs resume from the last completed step after an engine restart or a worker crash
- No Repeated Steps — Completed tasks return their journaled results on replay instead of running again; use idempotency keys to keep external side effects safe under retries
Dispatch
The engine is written in Rust. Workers are your own processes: they connect to the engine over gRPC, poll for tasks and report each result back.
- Eager Dispatch — A worker that completes a step can be handed its next task in the same response, saving a poll round trip
- One Dependency — The engine needs PostgreSQL and nothing else: no message broker and no separate scheduler
Typed SDKs
Each SDK works in its own language's idioms, so your editor gives you autocomplete and inline documentation for workflow inputs, task arguments and return values.
- Rust —
#[workflow]and#[task]macros with typed task references; the compiler checks task names, and inputs and outputs are checked when the task runs - Python —
@workflowand@taskdecorators on plain async functions, with type hints your editor and type checker read - TypeScript —
@Workflowand@Taskdecorators with typed task references
Code-First Approach
Workflows are regular functions marked with decorators or macros: no YAML, no DSL, no separate configuration files. They live alongside your application code, are versioned in Git, and go through your team's existing code review and CI.
- Version Control — Workflows are just code; track changes, review diffs and roll back with Git
- Testing — Test tasks and workflows with your language's usual test framework
- Reusability — Extract common steps into shared libraries and compose workflows from them
Security
The engine encrypts its traffic and authenticates the clients and workers that call it. See Security for the configuration.
- TLS Encryption — Encrypt gRPC traffic between workers, clients and the engine
- Authentication — API keys for programmatic access, and bearer tokens from an OIDC provider
Patterns
The building blocks a workflow uses are journaled like task results, so they survive crashes too.
- Parent-Child Workflows — Split a large process into child workflows, each with its own error handling
- Saga Pattern — Run compensating steps to undo completed work when a later step fails
- Fan-Out/Fan-In — Run tasks in parallel and collect their results, handling the ones that fail
- Events — Pause a run until an event arrives, such as a human approval, with an optional timeout
- Durable Timers — Sleep for minutes or months; the engine holds the timer, so no worker has to stay busy while the run waits
- Actors — Long-lived, addressable state, such as a shopping cart, that you invoke by key
Observability
The engine reports what it is doing in formats your existing tools read.
- Metrics — Prometheus-compatible metrics on
/metricsfor workflow and task execution - Logging — Structured logs, as JSON with
LOG_FORMAT=json - Tracing — The Python SDK can open an OpenTelemetry span around each workflow and task execution on the worker
Deployment
The engine ships as a Docker image, ghcr.io/orcher-io/orcher, and needs only PostgreSQL. Workers deploy like the rest of your code: a Rust binary, or a Python or Node.js application.
- Docker — Run the engine image with Docker or Docker Compose
- Kubernetes — Run the engine and your workers as standard Kubernetes deployments
- Horizontal Scaling — Add workers to handle more load; any worker registered for a workflow can pick up its next step