Description

franciscojavierarceo

opened on Aug 27, 2024

Issue body actions

Is your feature request related to a problem? Please describe.
We have outlined some options for feature serving and model inference in this documentation: https://docs.feast.dev/v/master/getting-started/architecture/model-inference

In it, we briefly discuss communication patterns and their tradeoffs (elaborated further here: https://docs.feast.dev/v/master/getting-started/architecture/write-patterns).

We don't have an architecture diagram to support these pieces.

Describe the solution you'd like
For a Fintech article I'm writing, I ended up drafting this diagram. I think we could potentially adjust this to use in our documentation.

Describe alternatives you've considered
N/A

Additional context
Good diagrams provide clarity and what not, so I thought this would be useful.

Activity

1. franciscojavierarceo

   added

   kind/featureNew feature or requestNew feature or request

   on Aug 27, 2024
2. franciscojavierarceo

   assigned

   tokoko

   ,

   haoxu0

   and

   shuchu

   on Aug 27, 2024
3. franciscojavierarceo commented on Sep 3, 2024

   

   franciscojavierarceo

   on Sep 3, 2024

   MemberAuthor

   More actions

   @tokoko any thoughts here? Otherwise I'll cut a PR with this to add to the documentation. I think it's missing the registry, so I could add that as well.
4. tokoko commented on Sep 4, 2024

   

   tokoko

   on Sep 4, 2024

   Collaborator

   More actions

   a couple of remarks

   • should feast box contain training datasets? Once they are generated, they are effectively outside of feast at that point.
   • What does a line coming from an online store marked Logs/CDC/ELT refer to?
5. franciscojavierarceo commented on Sep 4, 2024

   

   franciscojavierarceo

   on Sep 4, 2024

   MemberAuthor

   More actions

   That's a good call out!

   The logs/CDC/ELT refers to logs that should be sent from the feature server during transformations. This is the suggestion I would make to most folks to ensure feature and model reproducibility/replay.

   It could be events to SQS or Kafka/Kinesis or through a sidecar. We don't support this in Feast right now but it could be a valuable thing. It's somewhat complicated though.
6. franciscojavierarceo commented on Sep 24, 2024

   

   franciscojavierarceo

   on Sep 24, 2024

   MemberAuthor

   More actions

   @tokoko I added the training datasets and a registry, lmk what you think.

   
7. franciscojavierarceo

   closed this as completedon Sep 27, 2024

Issue actions