Hacker News

Subscribe to Hacker News feed
Hacker News RSS
Updated: 37 min 19 sec ago

Show HN: Lightweight Durable Workflows Built on Postgres

3 hours 46 min ago

Hi HN! This is Qian here with Peter (KraftyOne) and Jeremy (jedberg). We’re building DBOS, an open-source Python library for lightweight, durable workflows and queues. It’s comparable to systems like Airflow and Celery, but is designed to be more lightweight and embedded directly in your application process.

Our core idea is to make reliability and failure recovery easier without introducing additional infrastructure. You don’t need a separate workflow server: the DBOS library just connects to a Postgres database to persist state and recover from failures.

You define workflows by annotating normal Python functions, and DBOS handles checkpointing each step output to Postgres. If your app crashes or restarts, workflows automatically resume from the last completed step. Queues are backed by Postgres too, with deduplication, retries, and flow control.

You add durable workflows to your existing program by annotating ordinary functions as workflows and steps:

from dbos import DBOS @DBOS.step() def step_one(): ... @DBOS.step() def step_two(): ... @DBOS.workflow() def workflow(): step_one() step_two() The workflow is just an ordinary Python function. You can call it any way you like–from a FastAPI handler, in response to events, wherever you’d normally call a function.

We’ve just released DBOS Python 1.0. This enhances workflows with many powerful features we’ve been building over the last few months, including:

- Durable queues. Postgres-backed queues with all the queuing features of BullMQ/Celery (concurrency limits, rate limits, timeouts, priority, deduplication, etc.). Plus, they integrate with durable workflows, so you can write a workflow that enqueues 1K tasks, waits for and processes their results, and automatically recovers from any interruption.

- Programmatic workflow management. Your workflows are stored as rows in a Postgres table, so you have full programmatic control over them. Write scripts to query workflow executions, batch pause or resume workflows, or even restart failed workflows from a specific step. This makes it much easier to diagnose and recover from bugs and failures that affect thousands of workflows.

- Full support for both sync and async Python–write your workflows and steps as code either synchronously or asynchronously, it all works out of the box.

- Improved tooling, including dashboards, workflow graph visualization, workflow management via web UI, and more.

We’d love to hear your feedback and hope you can try DBOS out!

Comments URL: https://news.ycombinator.com/item?id=44202930

Points: 1

# Comments: 0

Categories: Hacker News

Medroid: AI Doctor, Always on Call

3 hours 48 min ago

Article URL: https://app.medroid.ai/join-founders

Comments URL: https://news.ycombinator.com/item?id=44202912

Points: 1

# Comments: 1

Categories: Hacker News

Show HN: Poseidon, an Oncall AI agent (open source alt to resolve.ai)

4 hours 7 min ago

Oncall devs, say goodbye to 2 AM pages on a Saturday.

Poseidon v0.1 analyzes past incidents, logs, metrics, and ops history to suggest root causes (demo in README).

In the next version, Poseidon will hook up to more tools in your org with MCP, and along with the summary, will send a "stop loss" button in your Slack message.

This stop loss button can: 1. revert a faulty config 2. revert a faulty release 3. turn off a feature flag 4. purge queues, etc

Integrating with netflix/dispatch will make this more powerful.

Looking for feedback on the concept, code, and how to make this more universal!

Comments URL: https://news.ycombinator.com/item?id=44202717

Points: 1

# Comments: 0

Categories: Hacker News

Ask HN: Will something replace transformers and GPUs in the next 10 years?

4 hours 11 min ago

I don't know why but I have the feeling that a huge amount of money (NVIDIA worth trillions, billions dumped into LLM-based companies) is supported by a single idea that might eventually be replaced by a completely different paradigm or by a new approach to LLMs that could make all of this worth nothing in a matter of months.

To be clear, I’m not saying this isn’t a super impactful idea and I actually think it's a revolution. Since Attention Is All You Need came out in 2017, everything has changed fast but it’s just that I also feel that another breakthrough from research could appear in a similar way.

Comments URL: https://news.ycombinator.com/item?id=44202671

Points: 2

# Comments: 1

Categories: Hacker News

Pages