Introduction 

Machine learning and model building requires some tooling around them like model management and deployment tools. A typical machine learning pipeline looks as follows: 

Usually, data scientists spend most of the time in non machine learning tasks, like data cleansing, sampling, managing models and stats for different iterations etc. We can solve this problem by providing Machine learning model management framework, Athena.

Goals

Build Machine learning model management framework to provide better experience to data scientists.

User Stories 

• As a data scientist, I should be able to provide training and test data or split the data into training and test datasets.
• As a data scientist, I should be able to prepare data to create machine learning model.
• As a data scientist, I should be able to create experiments, apply directives and tune machine learning model parameters. I should be able to do this repetitively in order to get accurate model.
• As a data scientist, I should be able to view, list different experiments, feature stats and model stats
• As a data scientist, I should be able to deploy and use finalized model for verification and prediction.

Discussions

The flow of the machine learning framework in cdap will be as below:

1. Choose Data(training/test data)
2. Decide splitting method (Random and others?)
3. Create an experiment out of split data. This means each experiment is associated with a dataset. After this point, the dataset can not be changed.
4. Prepare data using data prep and apply directives. These directives will be applied to whole training and test datasets.
5. Users can see stats about features.
6. Build model with some parameters. 
7. Evaluate and repeat steps 4 to 7 till accurate model is created.
8. Once accurate model is created, deploy the model.
9. After model is deployed, it will be stored in partitioned fileset and data scientists can use it in pipelines using Predictor plugins.

Design - wip

The idea is to use a spark program along with Data Prep application to expose different rest end points to calculate stats and manage models. When we choose and request split for dataset, the rest endpoint exposed by spark program will be called. Usually dataset for model building as small (couple of MBs). This method can either be synchronous or asynchronous. It will basically split the data and store them under <basepath>/training and <basepath>/test directories on HDFS. Once the experiment is created, ui will deploy and run a pipeline along with model parameters. We are implementing Trainer plugin to execute and build machine learning model. This model for that iteration will be persisted on HDFS along with model metadata. Once model is evaluated and finalized, it will be deployed which later can be used by Predictor plugin to predict values through hydrator pipeline.

Implementation

Model Lifecycle:

Wrangler Changes:

• Verify if emitting stats for all the features/columns takes long time in spark program. We need to assess this because we are relying on spark's map and reducebykey methods to get stats for each column in data. We need to make sure it does not take much time to evaluate. Test performance with 100k records.
• Spark program should also be able to execute directives and UDDs. 
• Define Rest Endpoints for list models, list experiments, get stats etc
• Design schema for datasets used to store stats and models.

Platform Changes:

• Support Rest Server in Spark Program
• Dynamic Plugin Loading. Similar to servic

Approach

Approach #1

Approach #2

API changes

New Programmatic APIs

New Java APIs introduced (both user facing and internal)

Deprecated Programmatic APIs

New REST APIs

srcpath="/path", 

offset=0,

limit=10,

[
  {
    "name": "",
    "description": "",
    "srcpath": "/path/to/files",
    "outcome": ""
  },
  ...
]

{
  "name": "",
  "description": "",
  "srcpath": "/path/to/files",
  "outcome": "price"
}

{
  "description": "",
  "srcpath": "/path/to/files",
  "outcome": "price",
  "outcomeType": "double",
  "directives": [ ... ]
}

{
  "directives": [ ... ],
  "schema": {...},
  "type": "random" | "first",
  "parameters": {
    "percent": "80" 
  },
  "description": "",
  "trainingPath": "",
  "testPath": "",
  "models": [ ... ]
}

{
  "directives": [ ... ],
  "schema": {...},
  "type": "random" | "first",
  "parameters": {
    "percent": "80" 
  },
  "description": ""
}

{
  "id": "123456"
}

{
  "id": "splitid",
  "directives": [ ... ],
  "schema": {...},
  "type": "random" | "first",
  "parameters": {
    "percent": "80" 
  },
  "description": "",
  "trainingPath": "",
  "testPath": "",
  "stats": [
    {
      "field": "numeric-field",
      "numTotal": {
        "train": ...,
        "test": ...,
      },
      "numNull": {
        "train": ...,
        "test": ...,
      },
      "mean": {
        "train": ...,
        "test": ...,
      },
      "min": {
        "train": ...,
        "test": ...,
      },
      "max": {
        "train": ...,
        "test": ...,
      },
      "stddev": {
        "train": ...,
        "test": ...,
      },
      "numZero": {
        "train": ...,
        "test": ...,
      },
      "numNegative": {
        "train": ...,
        "test": ...,
      },
      "numPositive": {
        "train": ...,
        "test": ...,
      },
      "histo": [
         {
           "bin": "[0,100)",
           "count": {
             "train": ...,
             "test": ...,
           },
         },
         ...
      ],
      "similarity": 0.88
    },
    {
      "field": "categorical-field",
      "numTotal": {
        "train": ...,
        "test": ...,
      },
      "numNull": {
        "train": ...,
        "test": ...,
      },
      "numEmpty": {
        "train": ...,
        "test": ...,
      },
      "unique": {
        "train": ...,
        "test": ...,
      },
      "histo": [
         {
           "bin": "cat1",
           "count": {
             "train": ...,
             "test": ...,
           },
         },
         ...
      ],
      "similarity": 0.88
    },
  ],
  "models": [ ... ]
}

{
  "name": "",
  "description": "",
  "directives": [ ... ],
  "split": id // optional,
}

{
  "id": "woeifn"
}

[
  {
    "id": "",
    "name": "",
    "description": "",
    "split": "id123456",
    "algorithm": "decision tree",
    "hyperparameters": {
      "maxBins": "3"
    },
    "evaluationMetrics": {
      "precision": 0.8,
      "recall": 0.5,
      "f1": 0.7,
      "rmse": ,
      "r2": ,
      "evariance": ,
      "mae": ,
    },
    "directives": [
      "parse-as-csv ...",
      ...
    ],
    "features": [ "f1", "f2", "f3" ],
    "deploytime": timestamp, // missing if not deployed
    "traintime": timestamp, // missing if not trained
    "createtime": timestamp
  },
  ...
]

{
  "schema": {...},
  "type": "random" | "first",
  "parameters": {
    "percent": "80" 
  },
  "description": "",
  // optional, if not given, model directives will be used
  "directives": [ ... ]
}

{
  "directives": [ ... ]
}

{
    "algorithm": "decision tree",
    "hyperparameters": {
      "maxBins": "3"
    }
}

Deprecated REST API

CLI Impact or Changes

• Impact #1
• Impact #2
• Impact #3

UI Impact or Changes

• Impact #1
• Impact #2
• Impact #3

Security Impact 

What's the impact on Authorization and how does the design take care of this aspect

Impact on Infrastructure Outages 

System behavior (if applicable - document impact on downstream [ YARN, HBase etc ] component failures) and how does the design take care of these aspect

Test Scenarios

Releases

Release X.Y.Z

Release X.Y.Z

Related Work

• Work #1
• Work #2
• Work #3

Future work