Define the pipeline sources and destinations

A deployment is a way to run a Machine Learning pipeline in a repeatable and automated way. First, you have to choose one of the 2 deployments methods.

Then, you need to connect the pipeline inputs and outputs with the desired sources and destinations.

Sources : This is the origin of the data that you want to connect to the pipeline inputs. The data can come from the data store, from environment variables, from constants or from the endpoint (if this deployment method has been chosen).

Destinations : This is the data drop point that you want to connect to the pipeline outputs. The data can go to the data store or to the endpoint (if this deployment method has been chosen).

Note

On this page, we will mainly focus on the platform's SDK interface.

However, it is possible to deploy a pipeline directly from the web interface by going to the Pipelines page, selecting a pipeline and then clicking the Deploy button.

Objects name	Constructor	Return type	Description
InputSource	InputSource(pipeline_input_name, required=False, default=None)	InputSource Object	Create a mapping source object for deployment.
OutputDestination	OutputDestination(pipeline_output_name, endpoint_output_name, required=False, default=None)	OutputDestination Object	Create a mapping destination object for deployment.

General function of the I/O mapping

Mapping rules

When you start a new deployment, the data flow is configured with a mapping. You can create this mapping in two ways: auto mapping (only available with endpoint trigger) or manual mapping in the SDK or UI.

Auto mapping automatically maps all inputs to endpoint variables for sources. If you need a different mapping or another trigger, you must map your inputs and outputs manually.

Auto mapping Example

For each input / output, you had defined a data type in pipeline. This data type will be the same as the mapped pipeline input / output. You need to map the good source and destination with the good data type :

Endpoint (input and output) → Variable (string, number, ...) and file
Data store → Only file
Constant → Only variable (string, number, ...)
Environment variable → Only variable (string, number, ...)
None → Variable (string, number, ...) and file

Limitations

The inputs (as outputs) of a pipeline mapped through the endpoint (default mapping) can consist of only a single file or multiple variables. Thus, if you use deployment by triggering an endpoint, you cannot have multiple files as inputs or outputs due to a technical limitation of the API calls.

To have multiple files as the inputs or outputs of an endpoint, you can compress the files into a single file (for example, with the [tar]{.title-ref} command) to be sent in the API call.

Examples: You can do

Warning

The data store is not yet available for the input/output mapping, but it's coming soon, so stay tuned.

Examples: You can't do

Creation input mapping

Import Dependencies

Before creating a mapping between the input/output of a pipeline and the sources/destinations of an endpoint, you need to import the InputSource and OutputDestination objects from the SDK.

from craft_ai_sdk.io import InputSource, OutputDestination

Function Definition

For each input of your pipeline, in manual mapping, you need to create a mapping object that will be given to the create_deployment() function.

boat_endpoint_input = InputSource(
    pipeline_input_name="apply_model",

    ## Choose from one of these parameters
    endpoint_input_name="boat", ## For mapping with an endpoint
    environment_variable_name=None, ## For mapping with an environment variable
    constant_value=None, ## For mapping with a constant value
    is_null=True ## For mapping with None

    is_required=True,
    default_value="empty",
)

Parameters

pipeline_input_name (str): name of the input which the deployment input will be linked to.

Different possible sources:

endpoint_input_name (str, optional): name of the input at the endpoint level to which the pipeline input will be linked to.
environment_variable_name (str, optional): name of the environment variable to which the pipeline input will be linked to.
constant_value (Any, optional): a constant value to which the pipeline input will be linked to.

Other parameters:

is_null (True, optional): if specified, the input will not take any value at execution time.
default_value (Any, optional): this parameter can only be specified if the deployment is an endpoint. In this case, if nothing is passed at the endpoint level, the pipeline input will take the default_value.
is_required (bool, optional): this parameter can only be specified if the deployment is an endpoint.If set to True, the corresponding endpoint input should be provided at execution time.

Return

An InputSource object that can be used in the deployment creation in the dictionary format.

Additional parameters for source definition

By default, the source is configured as an endpoint parameter, but you can configure a different source for your mapping. For each source, you have to add 1 parameter to :

Define the type of source with the name of parameter added
Precise element about the source

We will list all parameters you can have in your input mapping.

Warning

You can only add 1 parameter of source definition. By default, it's always endpoint source that is configured.

Endpoint source

Parameter name : endpoint_input_name

Source from : Outside through the endpoint

Value to put in parameter : name of the input received by the endpoint in the body of the HTTP call

Example :
```
endpoint_input = InputSource(
    pipeline_input_name="seagull",
    endpoint_input_name="seagull",
    required=False, #optional 
    default="Eureka", #optional
)
```

Data store source

Parameter name : datastore_path

Source from : file content from the data store

Value to put in parameter : path to a data store file

Example :

data_store_input = InputSource(
    pipeline_input_name="trainingData",
    datastore_path="path/to/trainingData.csv",
)

Example pipeline code to read file :

def pipelineFileIO (trainingData) :
   print (trainingData)

   with open(trainingData["path"]) as f:
      contents = f.readlines()
      print (contents)

Constant source

Parameter name : constant_value

Source from : static value

Value to put in parameter : direct value

Example :
```
constant_input = InputSource(
  pipeline_input_name="salt",
    constant_value=3,
)
```
Environment variable

Parameter name : environment_variable_name

Source from : the variables set at the level of an Environment in the platform

Value to put in parameter : name of the environment variable

Example :
```
env_var_input = InputSource(
  pipeline_input_name="fish",
    environment_variable_name="nameOfEnvVar",
)
```
None value

Parameter name : no_value

Destination to : void

Value to put in parameter : True

Example :
```
null_input = InputSource(
  pipeline_input_name="fish",
    is_null=True
)
```

Create output mapping

Import dependency

Before creating mapping between input / output of pipeline and sources / destination of endpoint, you have to import InputSource and OutputDestination objects from SDK.

from craft_ai_sdk.io import InputSource, OutputDestination

Function definition

For each output of your pipeline, in manual mapping, you have to create an object, that will be given to create_deployment() function.

endpoint_output = OutputDestination(
    pipeline_output_name="pred_0",

    ## Choose from one of these parameters
    endpoint_output_name="pred_0", ## For mapping with an endpoint
    is_null=True ## For mapping with None
)

Parameters

pipeline_output_name (str) - the specific output in the pipeline.
endpoint_output_name (str, optional) -- Name of the endpoint output to which the output is mapped.
is_null (True, optional) -- If specified, the output is not exposed as a deployment output.

Return

An OutputSource object who can be used in the deployment creation.

Additional parameter for destination definition

By default, the destination is configured as an endpoint parameter, but can configure a different source for your mapping. For each source, you have to add 1 parameter to :

Define the type of destination with the name of parameter added
Precise element about the destination

We will list all parameters you can have in your output mapping.

Warning

You have to add just 1 parameter of destination definition. If a parameter destination is missing, the function will generate an error (as opposed to input mapping).

Endpoint destination

Parameter name : endpoint_output_name

Destination to : Outside through the endpoint

Value to put in parameter : name of the output received by the endpoint in the body of the HTTP call

Example :
```
prediction_deployment_ouput = OutputDestination(
    pipeline_output_name="prediction",
    endpoint_output_name="beautiful_prediction",
)
```

Data store destination

Parameter name : datastore_path

Destination to : write a file into the data store

Value to put in parameter : path to a data store folder

Example :

prediction_deployment_ouput = OutputDestination(
    pipeline_output_name="history_prediction",
    datastore_path="path/to/history/folder.csv",
)

Example pipeline code to send file :

def pipelineFileIO () :
    text_file = open('history_prediction.txt', 'wb')  # Open the file in binary mode
    text_file.write("Result of pipeline send in file output :) ".encode('utf-8'))  # Encode the string to bytes
    text_file.close()
    fileOjb = {"history_prediction" : {"path": "history_prediction.txt"}}

   return fileOjb

Dynamic path :

You can also specify a dynamic path for the file to be uploaded by using one of the following patterns in your datastore path:

{execution_id}: The execution id of the deployment.
{date}: The date of the execution in truncated ISO 8601 (YYYYMMDD) format.
{date_time}: The date of the execution in ISO 8601 (YYYYM-MDD_hhmmss) format.

Example with a dynamic path :

prediction_deployment_ouput = OutputDestination(
    pipeline_output_name="history_prediction",
    datastore_path="path/to/history/exec_{execution_id}.csv",
)

Void destination

Parameter name : no_destination

Destination to : void

Value to put in parameter : True

Example :

prediction_deployment_ouput = OutputDestination(
    pipeline_output_name="history_prediction",
    is_null=True,
)

4. Generate new endpoint token

If you need to alter the endpoint token for an endpoint, you can generate a new one with the following SDK function.

sdk.generate_new_endpoint_token(endpoint_name="*your-endpoint-name*")

Warning

This will permanently deactivate the previous token.