HW SKU cost optimization

[siddharthteotia]

Our work on a Multi SKU recommender / cost-to-serve optimizer recently got accepted in USENIX and we will be presenting the work in March '25. We have built the entire machinery in-house and have a ton of results / data to share. The reason we built in-house first is because this ran longer than usual with an elaborate POC first to demonstrate value followed by actually doing it in production. Also, we first wanted to nail the integrations within LinkedIn on a fast path to solve our problem.

I feel the core algorithm that takes in production metrics, SKU profiles, processes high cardinality profiling data and recommends an optimal SKU can potentially be open-sourced.

Starting this thread to see if there is sufficient interest in exploring such a tool by Pinot users for their respective prod environments on-prem or in cloud. If so, we can think about putting it in open source. Not sure of LOE / timelines etc. Right now just trying to gauge interest

We had originally built this for Azure but then repurposed it to on-prem successfully without changing the core pieces of algorithm.

[hpvd]

yeah sounds pretty interesting! Looking forward to your presentation and hope for sharing with community!