Exposing a Python prediction model

In addition to standard models trained using the DSS Machine Learning component, the API node can also expose custom models written in Python by the user.

To write a “custom Python prediction” endpoint in an API node service, you must write a Python class that implements a predict method.

The custom model can optionally use a DSS managed folder. This managed folder is typically used to store the serialized version of the model. The code for the custom model is written in the “API service” part DSS.

Creating the Python prediction endpoint

To create a custom prediction endpoint, start by creating a service. (See Your first API service for more information). Then, create an endpoint of type “Custom prediction (Python)”.

You will need to indicate whether you want to create a Regression (predicting a continous value) or a Classification (predicting a discrete value) model.

DSS prefills the Code part with a sample depending on the selected model type.

Using a managed folder

A custom model can optionally (but most of the time) use a DSS managed folder. When you package your service, the contents of the folder is bundled with the package, and your custom code receives the path to the managed folder content.

A typical usage is when you have a custom train recipe that dumps the serialized model into a folder. Your custom prediction code then uses this managed folder.

../_images/custom_model_folder.png

Structure of the code

To create a custom model, you need to write a single Python class. This class must extend dataiku.apinode.predict.predictor.ClassificationPredictor or dataiku.apinode.predict.predictor.RegressionPredictor. The name of the class does not matter. DSS will automatically find your class.

The constructor of the class receives the path to the managed folder, if any.

Regression

A regression predictor must implement a single method: def predict(self, features_df)

This method receives a Pandas Dataframe of the input features. It must output one of the following forms:

  • prediction_series
  • (prediction_series, custom_keys_list)

Answer details:

  • prediction_series (mandatory): a Pandas Series of the predicted values. The output series must have the same number of rows as the input dataframe. If the model does not predict a row, it can leave numpy.nan in the output series.
  • custom_keys_list (optional, may be None): a Python list of dictionaries for each input row. Each list entry contains a dict of the customKeys that will be sent in the output (freely usable.)

The predict method must be able to predict multiple rows.

class dataiku.apinode.predict.predictor.RegressionPredictor(data_folder)

The base interface for a classification Custom API node predictor

predict(df)

The main prediction method.

Param:df: a dataframe of 1 or several records to predict
Returns:Either: prediction_series or (prediction_series, custom_keys_list)

prediction_series must be a Pandas Series of decisions

custom_keys_list is optional and must contain one entry per input row. Each entry of custom_keys_list must be a Python dictionary. These custom keys will be sent in the output result

prediction_series and custom_keys_list must have the same number of rows than df. It is legal to refuse to score a record. Leave a NA in prediction_series

Classification

A classification predictor must implement a single method: def predict(self, features_df)

This method receives a Panda Dataframe of the input features.

It must output one of the following forms:

  • prediction_series
  • (prediction_series, probas_df)
  • (prediction_series, probas_df, custom_keys_list)

Answer details:

  • prediction_series (mandatory): a Pandas Series of the predicted values. The output series must have the same number of rows as the input dataframe. If the model does not predict a row, it can leave None in the output series.
  • probas_df (optional, may be None): a Pandas Dataframe of the predicted probas. Must have one column per class, and the same number of rows as the input dataframe. If the model does not predict a row, it must leave numpy.nan in the probas dataframe.
  • custom_keys_list (optional, may be None): a Python list of dictionaries for each input row. Each list entry contains a dict of the customKeys that will be sent in the output (freely usable.)

The predict method must be able to predict multiple rows.

class dataiku.apinode.predict.predictor.ClassificationPredictor(data_folder)

The base interface for a classification Custom API node predictor

predict(df)

The main prediction method.

Param:df: a dataframe of 1 or several records to predict
Returns:Either: decision_series or (decision_series, proba_df) or (decision_series, proba_df, custom_keys_list)

decision_series must be a Pandas Series of decisions

proba_df is optional and must contain one column per class

custom_keys_list is optional and must contain one entry per input row. Each entry of custom_keys_list must be a Python dictionary. These custom keys will be sent in the output result

decision_series, proba_df and custom_keys_list must have the same number of rows than df. It is legal to refuse to score a record. Leave a NA in decision_series

Testing your code

Developing a custom model implies testing often. To ease this process, a “Development server” is integrated in the DSS UI.

To test your code, click on the “Deploy to Dev Server” button. The dev server starts and load your model. You are redirected to the Test tab where you can see whether your model loads.

You can then define Test queries, i.e. JSON objects akin to the ones that you would pass to the API node user API. When you click on the “Play test queries” button, the test queries are sent to the dev server, and the result is printed.

Python packages and versions

We strongly recommend that you use code environments for deploying custom model packages if these packages use any external (not bundled with DSS) library

Available APIs in a custom model code

Note that, while the dataiku.* libraries are accessible, most of the APIs that you use in Python recipes will not work: the code is not running with the DSS Design node, so datasets cannot be read by this API. If you need to enrich your features with data from your datasets, see /apinode/data_enrich. If you need to access a Folder, see Using a managed folder above.

Using your own libraries

You will sometimes need to write custom library functions (for example, shared between your custom training recipe and your custom model).

You can place these custom Python files in the project’s “libraries” folder, or globally in the lib/python folder of the DSS installation. Both recipes and custom models can import modules defined there.

When you package a service, the whole content of the lib/python folders (both project and instance) are bundled in the package. Note that this means that it is possible to have several generations of the service running at the same time, using different versions of the custom code from lib/python.

Server-side tuning

It is possible to tune the behavior of Python prediction endpoints on the API node side. You can tune how many concurrent requests your API node can handle. This depends mainly on your model (its speed and in-memory size) and the available resources on the server running the API node.

You can configure the parallelism parameters for an endpoint by creating a JSON file in the config/services folder in the API node’s data directory.

mkdir -p config/services/<SERVICE_ID>

Then create or edit the config/services/<SERVICE_ID>/<ENDPOINT_ID>.json file

This file must have the following structure and be valid JSON:

{
    "pool" : {
        "floor" : 1,
        "ceil" : 8,
        "cruise": 2,
        "queue" : 16,
        "timeout" : 10000
    }

}

This configuration allows you to control the number of allocated pipelines.

One allocated pipeline means one Python process running your code, preloaded with your initialization code, and ready to serve a prediction request. If you have 2 allocated pipelines (meaning 2 Python processes), 2 requests can be handled simultaneously, other requests will be queued until one of the pipelines is freed (or the request times out). When the queue is full, additional requests are rejected.

Those parameters are all positive integers:

  • floor (default: 1): Minimum number of pipelines. Those are allocated as soon as the endpoint is loaded.
  • ceil (default: 8): Maximum number of allocated pipelines at any given time. Additional requests will be queued. ceil floor
  • cruise (default: 2): The “nominal” number of allocated pipelines. When more requests come in, more pipelines may be allocated up to ceil. But when all pending requests have been completed, the number of pipeline may go down to cruise. floor cruise ceil
  • queue (default: 16): The number of requests that will be queued when ceil pipelines are already allocated and busy. The queue is fair: first received request will be handled first.
  • timeout (default: 10000): Time, in milliseconds, that a request may spend in queue wating for a free pipeline before being rejected.

Each Python process will only serve a single request at a time.

It is important to set “cruise”:

  • At a high-enough value to serve your expected reasonable peak traffic. If you set cruise too low, DSS will kill excedental Python processes, and will need to recreate a new one just afterwards.
  • But also at a not-too-high value, because each pipeline implies a running Python process consuming the memory required by the model.

You can also deploy your service on multiple servers, see High availability and scalability.