#workflow #deterministic #component #wasi #engine #wasm-component #model

app obeli-sk

Deterministic workflow engine

17 releases

new 0.3.0 Sep 13, 2024
0.2.2 Aug 24, 2024
0.1.15 Aug 10, 2024
0.1.10 Jul 23, 2024
0.1.7 Jun 13, 2024

#5 in #deterministic

Download history 430/week @ 2024-06-07 136/week @ 2024-06-14 20/week @ 2024-07-05 2/week @ 2024-07-12 109/week @ 2024-07-19 17/week @ 2024-07-26 149/week @ 2024-08-02 635/week @ 2024-08-09 41/week @ 2024-08-16 137/week @ 2024-08-23 4/week @ 2024-08-30

965 downloads per month

AGPL-3.0-only

700KB
16K SLoC

Obelisk

Deterministic workflow engine built on top of WASI Component Model

Project status / Disclaimer

This is a pre-release.

This repo contains backend code for local development and testing. The software doesn't have backward compatibility guarantees for CLI nor database format. Please exercise caution if attempting to use it for production.

Supported platforms

  • Linux x64

Core principles

  • Schema first, using WIT as the interface between workflows and activities.
  • Backend developer's delight
    • Single process for running the executor, workflows and activities, with an escape hatch for external activities (planned).
    • Automatic retries on errors, timeouts, workflow executions continuing after a server crash.
    • Observability (planned) - parameters and results together with function hierarchy must be preserved.
    • Composability - nesting workflows, calling activities written in any supported language
    • Replay and fork existing workflows(planned). Fix problems and continue.
    • Time traveling debugger for workflows (planned)

Concepts and features

  • Activities that must be idempotent, so that they can be stopped and retried at any moment. This contract must be fulfilled by the activity itself.

    • WASI activities are executed in a WASM sandbox
      • Able to contact HTTP servers using the WASI 0.2 HTTP client.
      • Able to read/write to the filesystem (planned).
    • Max execution duration support, after which the execution is suspended into intermittent timeout.
    • Retries on errors - on WASM traps (panics), or when returning an Error result.
    • Retries on timeouts with exponential backoff.
    • Execution result is persisted.
    • Performance option to keep the parent workflow execution hot or unload and replay the event history.
  • Deterministic workflows

    • Running in a WASM sandbox
    • Isolated from the environment
    • Able to spawn child workflows or activities, either blocking or awaiting the result eventually
    • Execution is persisted at every state change, so that it can be replayed after an interrupt or an error.
    • Ability to replay workflows with added log messages and other changes that do not alter the determinism of the execution (planned)
  • HTTP triggers (planned)

    • Mounted as a URL path, serving HTTP traffic.
    • Able to spawn child workflows or activities.
  • Work stealing executor

    • Periodically locking a batch of currently pending executions, starts/continues their execution
    • Cleaning up old hanging executions with expired locks. Executions that have the budget will be retried (planned).

Installation

Docker

CONTAINER_ID=$(docker run -d getobelisk/obelisk)
docker logs --follow $CONTAINER_ID | grep "Serving gRPC requests"
docker exec $CONTAINER_ID obelisk client component list
# See Usage for more details

Pre-built binary

Download latest release from the GitHub Release page.

Using latest version from crates.io

Download using cargo-binstall

cargo binstall obeli-sk

or build

cargo install --locked obeli-sk

Nix flakes

nix --extra-experimental-features nix-command --extra-experimental-features flakes run github:obeli-sk/obelisk

Usage

Starting the server

obelisk server run

Getting the list of loaded functions

obelisk client component list

wasm_workflow   fibo_workflow   wasm_workflow:sha256:885d60e8d9b09fadecae99d6143ec65ad819e1991275cba78e2431619381da75
Exports:
        testing:fibo-workflow/workflow.fiboa : func(n: u8, iterations: u32) -> u64
...

Submitting a function to execute (either workflow or activity)

# Call fibonacci(10) activity from the workflow 500 times in series.
obelisk client execution submit testing:fibo-workflow/workflow.fiboa '[10, 500]' --follow

Milestones

Milestone 1: Release the binary - done

  • Getting the obelisk application up and running as a Linux binary
  • Scheduling of workflows and wasm activities, retries on timeouts and failures
  • Persistence using sqlite
  • Launching child workflows/activities concurrently using join sets
  • Basic CLI for wasm component configuration and scheduling
  • Github release, docker image, publish to crates.io, support cargo-binstall

Milestone 2: Allow remote interaction via CLI - done

  • Move component and general configuration into a TOML file
  • Pull components -from an OCI registry
  • Publish the obelisk image to the Docker Hub (ubuntu, alpine)
  • obelisk client component push
  • gRPC API for execution management
  • Track the topmost parent
  • Params typecheck on creation, introspection of types of all functions in the system
  • Logging and tracing configuration, sending events to an OTLP collector

Milestone 3: Triggers, external activities, Web UI - started

  • HTTP webhook triggers able to start new executions (workflows and activities), able to wait for result before sending the response.
  • Expose filesystem for WASI activities with directory mapping, store stdout and stderr (configurable)
  • External activities - running external processes
  • HTML based UI for showing executions, event history and relations
  • Add examples with C#, Go, JS, Python
  • WASM to WIT exporter with extensions
  • Print the component's WIT, filter by export/import.

Planned features

  • Interactive CLI for execution management
  • OpenAPI activity generator
  • Limits on insertion of pending tasks or an eviction strategy like killing the oldest pending tasks.
  • Multi process executors
  • Labels restricting workflows/activities to executors
  • Periodic scheduling
  • Deadline propagation
  • Cancellation propagation
  • Queue capacity setting, adding backpressure to execution submission
  • Ability to simulate behaviour of the system with injected failures
  • Notify activities. When called, thir return value must be supplied via an API endpoint.
  • An API for listing executions with their open notify activities.
  • Read only query function that can be called during an await point or after execution finishes.
  • Workflow snapshots for faster replay
  • Optional stdout,stderr persistence / forwarding
  • Smart dependency routing from a caller via an interface import to one of many components that export it.
  • Smart retries - Retry budget, disabling retries when the activity is failing certain % of requests
  • Add jitter to retries

Building from source

Set up the development dependencies using nix flakes:

nix develop
# or `direnv allow`, after simlinking .envrc-example -> .envrc

Or manually download all dependencies, see dev-deps.txt and Ubuntu based verification Dockerfile Run the program

cargo run --release

Running Tests

./scripts/test.sh

Deterministic tests using the madsim simulator

./scripts/test-madsim.sh

Contributing

This project has a roadmap and features are added and tested in a certain order. If you would like to contribute a feature, please discuss the feature on GitHub. In order for us to accept patches and other contributions, you need to adopt our Contributor License Agreement (the "CLA"). The current version of the CLA can be found here.

Dependencies

~105MB
~2M SLoC