feat: add ML models

[arunvasudevan]

ML Models schema is inspired from Google's Model cards Paper- https://arxiv.org/pdf/1810.03993.pdf

Checklist

• The PR conforms to DataHub's Contributing Guideline (particularly Commit Message Format)
• Links to related issues (if applicable)
• Tests for the changes have been added/updated (if applicable)
• Docs related to the changes have been added/updated (if applicable)

[hshahoss]

Thanks Arun for the extensive PR. Left a few overall comments on the models.

[jywadhwani]

Thanks for the initiative @arunvasudevan

[arunvasudevan]

@hshahoss @jywadhwani Thanks for your detailed review. I have addressed your comments.
Kindly take a look at the same.

[hshahoss]

One option might be to define an enum for type of Cost called something like CostType. It could be an enum to identify the different type of costs that can be attached. To start with, we can have ORG_COST_CODE as one of the types. Then, we can define cost value as a object which can be a union of primitive/complex cost objects types. Something similar to a model like below:

/*
* Cost Details for an Entity
*/
record Cost {

    costType: CostType

    cost: CostValue
}

enum CostType { 

  /** 
   * Org Cost Code to which the Cost of this entity should be attributed to
   */ 
  ORG_COST_CODE 

// additional cost types here
} 

typeref CostValue = union[ 
  numericalCost: double
  stringCost: string
  customCost: customCostModel
  ....
]

Referencing the above model, we can now have an array of Cost aspect associated with the Model entity. The above model will provide an option for others to expand it to include their own cost definitions.

I'd encourage for us to think of all common models in this way so that they are extensible to other use cases. Let me know what you think.

[zimengy]

Does this Fabric refer to data centers? Just for my understanding, would models be trained per fabric?

[arunvasudevan]

FabricType refers to the origin or environment.
In most cases i don't think someone would train in every environment but they can attach the trained datasets (DatasetUrn) per platform, Model and Environment.

[arunvasudevan]

@hshahoss @zimengy @jywadhwani Thanks a lot for your review comments.
Based on the last round of review it made sense for me to make MlFeature a top level entity.
In addition to that i have also made other changes as per review.
Kindly take a look and let me know what you think.
Thanks!

[arunvasudevan]

@jywadhwani Updated all model and feature references to MLModel and MLFeature.

[hshahoss]

Overall, Looks good to me. Let's try to close on the open questions and get this in?

[jywadhwani]

While working on the RFC, came across some questions that I have added in my review comments.

[arunvasudevan]

@jywadhwani @hshahoss Appreciate your time in the review process.
Kindly take a look and let me know your comments.

[jywadhwani]

Thanks for the hard work! Excited to see this come to fruition.

[mars-lan]

Great work @arunvasudevan! Thanks for the unyielding will to see this through.