Just curious, can I use cloud bigtable as a feature store instead of using vertex AI feature store?

[chiayi]

I am trying to migrate my features table stored in bigquery to a feature store with lower latency. I'm choosing whether I should make use of vertex AI feature store or just cloud bigtable.

My features tables are <10MB, and it is used for real time prediction hence a database with low latency is sufficient.

Im just wondering aside from pricing, and ease of exporting data (bigtable requires more steps than vertex ai feature store), what is the difference between the 2 options?

Also, what type of database (eg: bigtable or redis?) is vertex AI feature store behind the scenes, when I am creating the feature store using the web UI? 

[comaro]

Hi, Chiayi,

The main difference between these 2, is that Vertex AI Feature Store is not considered a database as such, it is more like a product that provides a centralized repository for organizing, storing, and serving ML features [1]. In the other hand, Cloud Bigtable is a scalable NoSQL database service for large analytical and operational workloads [2]. More about [1][2].

[1] https://cloud.google.com/vertex-ai/docs/featurestore/overview
[2] https://cloud.google.com/bigtable/docs/overview

[ee07kkr]

Thanks @comaro 

I have a followup question on this as i am also in the phase of deciding between the two.

`it is more like a product that provides a centralized repository for organizing, storing, and serving ML features` - but the idea of organizing, storing and serving can also be done by BT. 

So, how does one decide or what factors impact the decision-making between the two, when picking one ? I guess the criteria would be super helpful. 

In my case, I have a BigTable instance, that is currently being used for other purposes but the volumes is not large enough currently (it will be in a while) but since its run in production, i have to deploy it in a fault tolerant and scalable way, which doesn't make it cost-effective right now.

So, i am keen to understand, whether BT will make sense for my fraud-detection model for computing features on the fly, or would it be best to use the feature store. I am using Vertex AI for my ML model architecture.

My model will need to compute a few features dynamically and input into the predictor.