🙏

very good

Thanks bro 😁

Hi Arun, even if data is coming from another prediction model then also it has to pass through the feature store. In that way it will be ensured that new data is in sync with training data else you have to write separate functions to validate it. So the thing is there are various ways to achieve things depends what fits to us. Also feature store serve data with low latency so if your old prediction model is deployed in other server and target model is deployed elsewhere then latency will be high and in that case you can utilise feature store online serving functionality. I am not saying it is a must but yes if you have a feature store then it comes with a lot of advantages which I have mentioned however you can read it in their official documents itself. Hope it is clear. If not please don't hesitate to ask again. Thanks

Thanks Ashutosh. i have one more question. For example while serving a model, we receive an inference request with some details (an id ) which is not available in the feature store, how will it be handled. So feature store can be only used for usecases where the features are available for both training and serving.

@@arunr2265 so in that case feature store will return empty record now it is upto us like how we handle this. We can apply a no record found check just before serving api. Ideally the ID is created only after record is generated and then it will be pushed to feature store and then served to api. So no point of ID doesnt exist. But I agree due to human error or other reasons if record does not exist then we have write this check explicitly as for as I know.

Thanks Ashutosh, so there might be scenarios where feature store can be used only for training?

@@arunr2265 it does not make much sense to use it only for training. Because one of the main reasons to use a feature store is to have sync between training and serving data. Just read the advantages of the feature store then it will be clear.

We already talked regarding this.

Just run the command `feast feature-views list` and see if feature view exist and in feature definition py file check if you have created the entity df or not. This seems error in your feature def py file.

If you are looking for any specific one, then please mention I will try to cover that in upcoming videos. Thanks

@@AshutoshTripathi_AI Credit card fraud and customer lifetime values cluster

It requires a good amount of work from some devops engineer. Because along with integrating a feast with kubeflow you have to setup backend offline and online databases. You can refer kubeflow documents where they have mentioned how to integrate third party tools like feast. But if you just want to test out certain things you simply open notebook from kubeflow and do pip install feast. It will install within notebook env. Then run feast init. This will initialize a repository with default configuration which will have file as a offline db. And there only you can configure connection string if you want to use the redis as online store. After that you can follow my video to use feast as feature store.

@@AshutoshTripathi_AI I have sent you an email via your website can we please get in touch. Any sort of help from you would be greatly appreciated. I'm actually trying to implement this for my company

@@erenyeager41 replied to your email.

Could you please share your feature def python file. It seems some problem there. You can share directly in LinkedIn msg.

@@AshutoshTripathi_AI i am not able to reply through linkdin as i dont have premium subscription

@@AshutoshTripathi_AI Do you have any alternate contact(email id) through which I can send

Ok. Will create one.

What is the size of data you are trying to load using feature store?

@@AshutoshTripathi_AI The size of the data is less than 5 mb, but I resolved it. It did not pass the entity_df correctly.

But now I have another problem the dataframe which is retrieved using the .to_df() method is empty. Any support on this.

@@user-sg6bx3sc2l is this happening after materialize. It means there is no data present in ur online store. Just increase the TTL entry or increase the duration of start and end date during materialize

@@AshutoshTripathi_AI Is it compulsory to materialize even if I am using a offline feature store?

For that you need to utilise feast api for data serving. For example get data from online store and during this time your online store should be active means deployed pod should be up.