durable

Lightweight workflow durability for Python. Make any async workflow resumable after crashes with just a decorator.

Backed by SQLite out of the box; swap in Redis or any Store subclass for production.

Install

pip install python-durable

# With Redis support
pip install python-durable[redis]

# With Pydantic AI integration
pip install python-durable[pydantic-ai]

Quick start

from durable import Workflow
from durable.backoff import exponential

wf = Workflow("my-app")

@wf.task(retries=3, backoff=exponential(base=2, max=60))
async def fetch_data(url: str) -> dict:
    async with httpx.AsyncClient() as client:
        return (await client.get(url)).json()

@wf.task
async def save_result(data: dict) -> None:
    await db.insert(data)

@wf.workflow(id="pipeline-{source}")
async def run_pipeline(source: str) -> None:
    data = await fetch_data(f"https://api.example.com/{source}")
    await save_result(data)

# First call: runs all steps and checkpoints each one.
# If it crashes and you call it again with the same args,
# completed steps are replayed from SQLite instantly.
await run_pipeline(source="users")

How it works

1. @wf.task wraps an async function with checkpoint + retry logic. When called inside a workflow, results are persisted to the store. On re-run, completed steps return their cached result without re-executing.

2. @wf.workflow marks the entry point of a durable run. It manages a RunContext (via ContextVar) so tasks automatically know which run they belong to. The id parameter is a template string resolved from function arguments at call time.

3. Store is the persistence backend. SQLiteStore is the default (zero config, backed by aiosqlite). RedisStore is available for distributed setups. Subclass Store to use Postgres or anything else.

Features

• Crash recovery — completed steps are never re-executed after a restart
• Automatic retries — configurable per-task with exponential, linear, or constant backoff
• Signals — durably wait for external input (approvals, webhooks, human-in-the-loop)
• Loop support — use step_id to checkpoint each iteration independently
• Zero magic outside workflows — tasks work as plain async functions when called without a workflow context
• Pluggable storage — SQLite by default, Redis built-in, or bring your own Store

Signals

Workflows can pause and wait for external input using signals:

@wf.workflow(id="order-{order_id}")
async def process_order(order_id: str) -> None:
    await prepare_order(order_id)
    approval = await wf.signal("manager-approval")  # pauses here
    if approval["approved"]:
        await ship_order(order_id)

# From the outside (e.g. a web handler):
await wf.complete("order-42", "manager-approval", {"approved": True})

Signals are durable — if the workflow crashes and restarts, a previously delivered signal replays instantly.

Redis store

For distributed or multi-process setups, use RedisStore instead of the default SQLite:

from durable import Workflow, RedisStore

store = RedisStore(url="redis://localhost:6379/0", ttl=86400)
wf = Workflow("my-app", db=store)

Keys auto-expire based on ttl (default: 24 hours).

Backoff strategies

from durable.backoff import exponential, linear, constant

@wf.task(retries=5, backoff=exponential(base=2, max=60))  # 2s, 4s, 8s, 16s, 32s
async def exp_task(): ...

@wf.task(retries=3, backoff=linear(start=2, step=3))      # 2s, 5s, 8s
async def linear_task(): ...

@wf.task(retries=3, backoff=constant(5))                   # 5s, 5s, 5s
async def const_task(): ...

Loops with step_id

When calling the same task in a loop, pass step_id so each iteration is checkpointed independently:

@wf.workflow(id="batch-{batch_id}")
async def process_batch(batch_id: str) -> None:
    for i, item in enumerate(items):
        await process_item(item, step_id=f"item-{i}")

If the workflow crashes mid-loop, only the remaining items are processed on restart.

Pydantic AI integration

Make any pydantic-ai agent durable with zero infrastructure — no Temporal server, no Prefect cloud, no Postgres. Just decorators and a SQLite file.

Pydantic AI natively supports three durable execution backends: Temporal, DBOS, and Prefect. All three require external infrastructure. python-durable is a fourth option that trades scale for simplicity:

Best for: prototyping, CLI tools, single-process services, SME deployments, and any situation where you want durable agents without ops overhead.

DurableAgent

from pydantic_ai import Agent
from durable import Workflow
from durable.pydantic_ai import DurableAgent, TaskConfig
from durable.backoff import exponential

wf = Workflow("my-app")
agent = Agent("openai:gpt-5.2", instructions="Be helpful.", name="assistant")

durable_agent = DurableAgent(agent, wf)

result = await durable_agent.run("What is the capital of France?")
print(result.output)  # Paris

# Same run_id after crash → replayed from SQLite, no LLM call
result = await durable_agent.run("What is the capital of France?", run_id="same-id")

With custom retry config:

durable_agent = DurableAgent(
    agent,
    wf,
    model_task_config=TaskConfig(retries=5, backoff=exponential(base=2, max=120)),
    tool_task_config=TaskConfig(retries=3),
)

@durable_tool

Make individual tool functions durable:

from durable.pydantic_ai import durable_tool

@durable_tool(wf, retries=3, backoff=exponential(base=2, max=60))
async def web_search(query: str) -> str:
    async with httpx.AsyncClient() as client:
        return (await client.get(f"https://api.search.com?q={query}")).text

@durable_pipeline

Multi-agent workflows with per-step checkpointing. On crash, completed steps replay from the store and only remaining work executes:

from durable.pydantic_ai import durable_pipeline

@durable_pipeline(wf, id="research-{topic_id}")
async def research(topic_id: str, topic: str) -> str:
    plan = await plan_research(topic)
    findings = []
    for i, query in enumerate(plan["queries"]):
        r = await search(query, step_id=f"q-{i}")
        findings.append(r)
    return await summarize(findings)

Comparison with Temporal

# Temporal — requires server + worker + plugin
from temporalio import workflow
from pydantic_ai.durable_exec.temporal import TemporalAgent

temporal_agent = TemporalAgent(agent)

@workflow.defn
class MyWorkflow:
    @workflow.run
    async def run(self, prompt: str):
        return await temporal_agent.run(prompt)

# python-durable
from durable import Workflow
from durable.pydantic_ai import DurableAgent

wf = Workflow("my-app")
durable_agent = DurableAgent(agent, wf)
result = await durable_agent.run("Hello")

Caveats

• Tool functions registered on the pydantic-ai agent are NOT automatically wrapped. If they perform I/O, decorate them with @durable_tool(wf) or @wf.task.
• Streaming (agent.run_stream()) is not supported in durable mode (same limitation as DBOS). Use agent.run().
• Single process — unlike Temporal/DBOS, python-durable runs in-process. For distributed workloads, use the Redis store.

See examples/pydantic_ai_example.py for five complete patterns.

Important: JSON serialization

Task return values must be JSON-serializable (dicts, lists, strings, numbers, booleans, None). The store uses json.dumps internally.

For Pydantic models, return .model_dump() from tasks and reconstruct with .model_validate() downstream:

@wf.task
async def validate_invoice(draft: InvoiceDraft) -> dict:
    validated = ValidatedInvoice(...)
    return validated.model_dump()

@wf.task
async def book_invoice(data: dict) -> dict:
    invoice = ValidatedInvoice.model_validate(data)
    ...

License

MIT