Data Universe

Although Polyforms are meant to be used asynchronously, where data is first submitted on a reactive basis - as the data changes, or as it is needed it is drawn in.

However, for the sake of a simple example, you could call a Polyform and introduce all of the data up-front (more of the classic way of doing AI). Using this example, we have a Polyform which is set up as:

• Accepts 10 attributes (consider it information around a person and their credit)
• The model is already trained and living in the Universe
• When invoked, it loads the model from the universe but uses the data provided as inputs.
• On average, it runs in 500ms from start to finish.
• Because of the application, it is, on an average, called 220 times a day, and returns a score for risk on the customer.
• Because it doesn't return any data from the data universe (just a score), there are no storage costs.
• The Polyform is shared from the market, with a Per-Duration cost of $0.001.
• We also show how much the owner of the Polyform would receive as market revenue.

Knowing this, the forecast costs are:

Average Runtime (ms)	GB Memory	Calls per day	Polyform Duration*GB Cost	Market Cost	Monthly Expected Total	Market Owner Revenue
500	1	2,000	$0.61	$30.50	$30.56	$27.45

In this example, we use the asynchronous model where data is introduced in a reactive manner, as it changes or is needed. We have four forms:

1. PersonData - Receive's information about a person into the universe, but does not do any data science. It has no additional market multipliers.
2. ScorePerson - This is the primary science call, and it assumes the data has already been put into the universe with PersonData, or that it will be included from EnrichWithApiQuery (to see how this works, look into the Polyform definition in the Innovation Center
3. EnrichWithApiQuery - This makes an API call to an external system, as needed, and comes with a Per-Request market cost because of the external vendor.
4. RetrainModel - With this model, inputs accepted are used to recreate the model that is used during ScorePerson

With this in place, assume the following:

• PersonData is called as changes happen in the source system, in a reactive event manner.
• ScorePerson is called as it is needed (perhaps when somebody requests a loan).
• EnrichWithApiQuery is called on behalf of the Polyform, indirectly, as part of the expectation defined on the intersect with ScorePerson.
• RetrainModel is called only when we need to retrain. For this example, we'll only include a single call.

Knowing this, the forecast costs are:

Form	Req / day	GB Mem	Avg Runtime (ms)	Duration*GB Cost	Market Duration Cost	Market Request Cost	Expected Monthly Cost
PersonData	150	1	300	$0.03	$0.00	$0.00	$0.03
ScorePerson	250	2	500	$0.16	$7.63	$0.00	$7.79
EnrichWithApiQuery	250	0.5	200	$0.02	$0.00	$0.76	$0.79
Total							$8.62

The second example doesn't include storage costs, which are a part of this equation, so now let's add those in. If we assume 20GB of user storage is used per month, the cost of that is $3.00. But you add to that the IOPs cost, which are shown per Polyform:

Form	k IOPs	Expected Monthly Cost
PersonData	9,150	$2.01
ScorePerson	18,300	$4.03
EnrichWithApiQuery	0	$0.00
Total		$6.04