_1722956449156001fz1R.png "Anyscale Logo")
Ray is very popular among AI practitioners for its rich ecosystem and great integration. However, its core distributed memory model and python-first design makes it very appealing to be applied in research infra in quantitative trading. I would like to share from my experiences building a platform for quant research and trading with Ray on Kubernetes that supports up to 100k concurrent CPU cores and easily extensible to GPUs.
I would like to focus on points below:
1. Think about integrate Ray into a quantitative research workflow, unifying large offline research jobs and interactive "small" intraday jobs.
2. Kuberay learnings and challenges faced running Ray on Kubernetes that only use spot instance and the attemped solution which is a customized Ray controller with different retry policies to build a reliable layer on unreliable infra.
3. Experiences pioneering Ray within a firm from getting buy-in to educating users.
I would like to focus on points below:
1. Think about integrate Ray into a quantitative research workflow, unifying large offline research jobs and interactive "small" intraday jobs.
2. Kuberay learnings and challenges faced running Ray on Kubernetes that only use spot instance and the attemped solution which is a customized Ray controller with different retry policies to build a reliable layer on unreliable infra.
3. Experiences pioneering Ray within a firm from getting buy-in to educating users.
