harel

Durable, distributed statecharts for Python.

A hierarchical statechart engine for Python — with the durability and distribution of a workflow engine, but the right model underneath.

You author a machine in a small textual DSL (.stm), it compiles to an immutable Definition, and a pure, effects-based engine runs it against a serializable Execution that can be persisted, distributed across workers, and survive crashes.

Why this exists

In ten years across different companies I kept hitting the same problem: a thing that was really a hierarchical state machine — an order, a subscription, a device, a claim, a deployment — got called a "pipeline", a "workflow", or just "a graph". So it got built with the tool that matched the name instead of the shape: a DAG runner, a task queue, a pile of if status == ... branches.

The result was always the same: states implied but never named, transitions scattered across the codebase, illegal states reachable, and "creative" retry/cancel logic that was hell to debug, follow, or extend.

A statechart is the correct model for that shape — hierarchy, orthogonal regions, guarded transitions, explicit terminal verdicts — and it has been since Harel, 1987. What was missing in Python was a statechart engine that is also durable and distributed:

• sismic models statecharts well but is in-memory only.
• Temporal / DBOS give you durable execution, but model work as imperative code, not as a declarative statechart.
• XState is the gold standard for statecharts — but it's JavaScript.

harel sits in that gap: the statechart as the model, durability and distribution as the runtime. See the comparison below.

Install

pip install harel        # core: DSL + engine + in-memory/sqlite durability
# optional backends, pick what you need:
pip install "harel[redis]"     # Redis store + transport
pip install "harel[postgres]"  # Postgres store + transport
pip install "harel[mongo]"     # MongoDB store + transport
pip install "harel[libsql]"    # libSQL store + transport — EXPERIMENTAL (local file tested; Turso/sqld path needs a Turso account to validate)
pip install "harel[dynamodb]"  # DynamoDB store (pairs with sqs for an all-AWS stack)
pip install "harel[sqs]"       # AWS SQS FIFO transport
pip install "harel[lsp]"       # DSL language server (editor tooling)

Requires Python 3.11+.

Quickstart

Scaffold a starter machine that validates and runs out of the box — zero to a working state machine in one command:

harel new approval.stm
harel run approval.stm -e Submit -e Approve   # (start) -> Draft -> Review -> Approved

Or author one yourself — events are declared up front, transitions live inside the state they leave, and terminals declare their verdict:

event PlaceOrder {}
event PaymentAuthorized {}
event CancelOrder {}
event Delivered {}

machine order {
  initial Cart
  state Cart {}
  state AwaitingPayment {}
  state Paid {}
  final Delivered success
  final Cancelled cancelled

  from Cart to AwaitingPayment on PlaceOrder
  from AwaitingPayment to Paid on PaymentAuthorized
  from AwaitingPayment to Cancelled on CancelOrder
  from Paid to Delivered on Delivered
}

Run it through the headless durable runner (here over an in-memory store):

from harel import definition_from_dsl, DurableRunner, DictStore, Event

SOURCE = """
event PlaceOrder {}
event PaymentAuthorized {}
event CancelOrder {}
event Delivered {}

machine order {
  initial Cart
  state Cart {}
  state AwaitingPayment {}
  state Paid {}
  final Delivered success
  final Cancelled cancelled

  from Cart to AwaitingPayment on PlaceOrder
  from AwaitingPayment to Paid on PaymentAuthorized
  from AwaitingPayment to Cancelled on CancelOrder
  from Paid to Delivered on Delivered
}
"""

defn = definition_from_dsl(SOURCE, "order")          # compile the DSL
runner = DurableRunner(DictStore(), {defn.id: defn})  # swap DictStore for Sqlite/Redis/Postgres

exe = runner.create(defn.id)
for kind in ["PlaceOrder", "PaymentAuthorized", "Delivered"]:
    exe = runner.process(exe.id, Event(kind=kind))
    print(f"{kind} -> {exe.active_path}")

print(exe.status.name, exe.outcome)

PlaceOrder -> AwaitingPayment
PaymentAuthorized -> Paid
Delivered -> Delivered
DONE success

A complete, runnable example (nested states, a selector-driven retry, actions) lives in examples/place_order/ — run it with uv run python -m examples.place_order.run.

Documentation

Read the docs online at acasadom.github.io/harel — a step-by-step tutorial (14 stages that grow one example from a turnstile to durable, distributed submachines) plus operations and reference guides. The source lives under docs/; build the HTML locally with make docs. Every code example in the docs is executed in CI, so it stays in sync with the engine.

• Tutorial — start here; the model, one step at a time.
• Architecture — how harel works inside: the pure engine, the effect protocol, the single atomic checkpoint, in-memory vs. distributed (with diagrams).
• Reference: DSL · CLI · Public API · Visualization.
• Operations: durability · distribution · control plane · monitor TUI · remote actions / FaaS.

Where it fits

Use a statechart when the domain is a machine of named states with hierarchy, guarded transitions and explicit terminal verdicts. If your domain is genuinely "run these steps in order with retries", a workflow engine is the better fit — this is not trying to replace one.

What's in the box

• A textual DSL (.stm, parsed with lark) that reads like a spec: nested composite states, orthogonal regions, guarded transitions, named guards, computed select branches, parametrized fragments, imports, and black-box invoke of sub-machines (including data-parallel fan-out).
• A pure, effects-based engine: start/process are generators that describe effects (run an action, emit an event, spawn regions) and mutate a serializable Execution. No IO, no user code inside the engine — which makes runs deterministic and trivially testable.
• Durable & distributed execution: optimistic-concurrency (CAS) checkpointing, a transactional outbox, event dedupe, durable timers, and a control plane (cancel/suspend/resume/terminate). Stores: in-memory, SQLite, Redis, Postgres, rqlite, MongoDB, DynamoDB (+ libSQL/Turso, experimental). Transports: in-memory, SQLite, Redis, Postgres, rqlite, MongoDB, SQS (+ libSQL/Turso, experimental) — mix freely.
• Sync or async, your choice of execution model: async runners and worker for asyncio code, and sync runners that run either on a shared background event loop or inline in the caller's own thread (execution="inline") — on the caller's database connection and inside its transaction. The same logic and the same commit under each.
• Your own backend: a store or transport written outside harel builds on harel.engine.store.base / harel.engine.transport.base and checks itself against the same contract tests harel runs on its own (harel.testing).
• Static validation (validate): unreachable states, non-deterministic transitions, unresolved selector targets, missing terminal verdicts, timeout shape — surface-independent, run it before you execute.
• Editor tooling: a DSL language server (diagnostics, hover, go-to-definition, completion across imports) and a VSCode extension with a live Mermaid statechart preview.
• Visualization: render any machine to PlantUML or Mermaid.

Design principle: it's a statechart, not a job engine

The engine schedules; the model decides. Retry/backoff is a composite with a timeout and a selector — not an engine feature. Cancellation is a modelled on Cancel transition when the machine wants to own its cleanup. There is no hidden "default to success/failed" — a validation rule forces you to declare the terminal verdict where it's consumed, instead of the engine guessing. Policy stays in the model; the engine stays small.

Status

Beta. The core engine, DSL, durability/distribution layers, the monitor TUI, remote action execution on FaaS (AWS Lambda / HTTP functions), and editor tooling are all in place and covered by an extensive test suite (run uv run pytest); every code example in the docs is executed in CI.

Roadmap: validating the libSQL/Turso path against a real account (today experimental); fault-injection testing of the distributed path under crash/lease-expiry; opt-in bounded retry for transient action errors.

Development

uv sync                 # create/refresh the venv (Python 3.13)
uv run pytest           # run the suite
make lint               # ruff + .stm formatter check
make type-check         # mypy

The name

Named after David Harel, who introduced statecharts in his 1987 paper Statecharts: A Visual Formalism for Complex Systems (PDF · DOI), and later recounted how it came to be in Statecharts in the Making: A Personal Account (2007). This engine makes that formalism durable and distributed, in his honour.

License

Apache License 2.0 © Alberto Casado