waymark is a library to let you build durable background tasks that withstand server restarts, task crashes, and long-running jobs. It's built on Postgres, with no other infrastructure to deploy. Python is the first supported language; more are coming.

Documentation: waymark.sh

Getting started:

Running Waymark:

Python:

We ship all client and server binaries in one Python package. Install it via your package manager of choice:

Let's say you need to send welcome emails to a batch of users, but only the active ones. You want to fetch them all, filter out inactive accounts, then fan out emails in parallel. Actions are the distributed work - plain async functions, sent to your workers:

from typing import Annotated

from waymark import Depends, action

@action
async def fetch_users(
    user_ids: list[str],
    db: Annotated[Database, Depends(get_db)],
) -> list[User]:
    return await db.get_many(User, user_ids)

@action
async def send_email(
    to: str,
    subject: str,
    emailer: Annotated[EmailClient, Depends(get_email_client)],
) -> EmailResult:
    return await emailer.send(to=to, subject=subject)

The workflow is the durable control flow that orchestrates them:

import asyncio

from waymark import Workflow, workflow

@workflow
class WelcomeEmailWorkflow(Workflow):
    async def run(self, user_ids: list[str]) -> list[EmailResult]:
        users = await fetch_users(user_ids)
        active_users = [user for user in users if user.active]

        results = await asyncio.gather(
            *[
                send_email(to=user.email, subject="Welcome")
                for user in active_users
            ],
            return_exceptions=False,
        )

        return results

Run the node - the process that executes workflows and their actions - against your database:

export WAYMARK_DATABASE_URL=postgresql://postgres:postgres@localhost:5432/waymark
export WAYMARK_HTTP_ENABLED=true  # the webapp, on http://localhost:24119
uv run waymark-executor

Then kick off the workflow from any Python process, and wait for its result:

async def welcome_users(user_ids: list[str]):
    await WelcomeEmailWorkflow().run(user_ids)

None of this executes inline in your webserver: the run is queued to Postgres and executed by the node.

Actions are the distributed work: network calls, database queries, anything that can fail and should be retried independently.

Workflows are the control flow: loops, conditionals, parallel branches. They orchestrate actions but don't do heavy lifting themselves.

Workflows are plain async Python. A few of the things they can do:

1. Retries and timeouts, per call. By default an action runs once, like a regular function call. Wrap a call in self.run_action(...) to give it a retry policy and a timeout - see Retries & Timeouts.

   from datetime import timedelta
   
   from waymark import RetryPolicy
   
   async def run(self, order_id: str) -> Receipt:
       return await self.run_action(
           charge_card(order_id),
           retry=RetryPolicy(attempts=5, backoff_seconds=10),
           timeout=timedelta(minutes=2),
       )

2. Branches and loops. if/elif/else, for and while compile into the workflow program and are executed by the runtime, just like your actions.

   async def run(self, order_ids: list[str]) -> list[str]:
       shipped = []
       for order_id in order_ids:
           status = await fetch_status(order_id)
           if status == "shipped":
               shipped.append(order_id)
       return shipped

3. Parallel fan-out and durable sleep. asyncio.gather runs actions in parallel; asyncio.sleep pauses the workflow durably, surviving restarts - for a second or for a day.

   import asyncio
   
   async def run(self, user_id: str) -> str:
       profile, history = await asyncio.gather(
           fetch_profile(user_id),
           fetch_purchase_history(user_id),
           return_exceptions=False,
       )
   
       # wait a day before following up
       await asyncio.sleep(24 * 60 * 60)
       return await send_recommendations(profile, history)

4. Schedules. Run a workflow on a cron expression or a fixed interval, with no extra infrastructure - see Scheduled Workflows.

   await schedule_workflow(DataSyncWorkflow, schedule_name="hourly", schedule="0 * * * *")

5. A built-in webapp. waymark-executor serves a webapp showing every workflow instance, each one's timeline of calls, and the load on your nodes.

Background jobs in webapps are so frequently used that they should really be a primitive of your fullstack library: database, backend, frontend, and background jobs. Otherwise you're stuck in a situation where users either have to always make blocking requests to an API or you spin up ephemeral tasks that will be killed during re-deployments or an accidental docker crash.

After trying most of the ecosystem in the last 3 years, I believe background jobs should provide a few key features:

• Easy to write control flow as ordinary code in the language you already use
• Should be both very simple to test locally and very simple to deploy remotely
• Reasonable default configurations to scale to a reasonable request volume without performance tuning

On the point of control flow, we shouldn't be forced into a DAG definition (decorators, custom syntax). It should be regular control flow just distinguished because the flows are durable and because some portions of the parallelism can be run across machines.

Nothing on the market provides this balance - waymark aims to try. We don't expect ourselves to reach best in class functionality for load performance. Instead we intend for this to scale most applications well past product market fit.

Waymark takes a different approach from replay-based workflow engines like Temporal or Vercel Workflow.

The first time a workflow is run or scheduled, it is compiled to an intermediate representation (IR); in Python, the SDK parses its run() method's AST. This IR captures your control flow - loops, conditionals, parallel branches - and is lowered to bytecode for a durable virtual machine. The Rust runtime executes the bytecode and snapshots the VM's state to Postgres as it goes, so a workflow resumes from where it was after a crash. Your original workflow definition is never re-executed during workflow recovery.

This is convenient in practice because it means that if your workflow compiles, your workflow will run as advertised. There's no need to hack around stdlib functions that are non-deterministic (like time/uuid/etc) because you'll get an error on compilation to switch these into an explicit action. The few patterns that currently slip past the compiler are listed under known issues and tracked in #772.

When should you use Waymark?

• You're already using Postgres and a supported language (Python today), with any framework
• You have a lot of async heavy logic that needs to be durable and can be retried if it fails (common with 3rd party API calls, db jobs, etc)
• You want something that works the same locally as when deployed remotely
• You want background job code to plug and play with your existing unit test & static analysis stack
• You are focused on getting to product market fit versus scale

Performance is a top priority of waymark. That's why it's written with a Rust core and runs continuous benchmarks on CI. But it's not the only priority. After all there's only so much we can do with Postgres as an ACID backing store. Once you start to tax Postgres' capabilities you're probably at the scale where you should switch to a more complicated architecture.

When shouldn't you?

• You have particularly latency sensitive background jobs, where you need <100ms acknowledgement and handling of each task.
• You have a huge scale of concurrent background jobs, order of magnitude >10k actions being coordinated concurrently.
• You have tried some existing task coordinators and need to scale your solution to the next 10x worth of traffic.

There is no shortage of robust background queues, including ones like Temporal.io/RabbitMQ that scale to millions of requests a second.

Almost all of these require a dedicated task broker that you host alongside your app. This usually isn't a huge deal during POCs but can get complex as you need to performance tune it for production. Cloud hosting of most of these are billed per-event and can get very expensive depending on how you orchestrate your jobs. They also typically force you to migrate your logic to fit the conventions of the framework.

Open source solutions like RabbitMQ have been battle tested over decades & large companies like Temporal are able to throw a lot of resources towards optimization. Both of these solutions are great choices - just intended to solve for different scopes. Expect an associated higher amount of setup and management complexity.

If you have a particular workflow that you think should be working but isn't yet compiling correctly, please file an issue.

The webapp is compiled into waymark-executor, so building from source requires Node.js (version in .node-version) and npm:

make js-deps
cargo build --bin waymark-executor

If you want to contribute, check out the contributing guidelines.