We wanted to use the Jina embeddings v2 model for an Elixir project but it’s not supported out of the box in Bumblebee.
Here, I describe how to implement and verify the Jina model so we can run it in Elixir.
Spoiler alert: it’s probably less work than you expect.
implementing more code from ‘Advances in Financial Machine Learning’ in Elixir. The labeling step, using the Triple Barrier Method:
The Triple Barrier Method: Labeling Financial Time Series for ML in Elixir
I had an issue with esbuild subprocess staying around as “zombie” and I finally found a solution I wanted to share with you:
At first I tried to exit when stdin was closed but it was not working, I found a workaround which uses an interval to send a SIGNULL signal to the parent to check if it is alive.
You should check for the closure of stdin to detect the parent closing the port. E.g. that’s how the bun package does it: bun/priv/wrapper.js at main · crbelaus/bun · GitHub
const cleanup = async () => {
// Dispose the watch context properly
console.log("--- dispose");
await ctx.dispose();
process.exit(0);
};
process.stdin.resume();
process.stdin.on("close", () => cleanup());
await ctx.watch();
This seems to work just fine for me.
This works on my local machine but not on our shared dev server. With this I still find some zombie processes after a few days of people working on the project.
I recently made some optimizations at work that saved more than 40% CPU at peak time. The post explains how I identified and fixed the bottleneck
Blog - ETS vs Redis
I wrote an article about how we’re using Elixir’s Mix tasks, Sourceror, and Igniter to streamline project and dependency setup on client projects:
I have been trying to exercise my writing muscles, getting past the barrier of finishing it up and posting it online. Therefore I chose a small subject for it, namely how default function arguments work in Elixir.