mlservicewrapper-core 0.3.8a0

Configure a Python service for repeated execution

Installing

• Install using pip

  pip install mlservicewrapper-core
  

• Install directly from the latest GitHub Release.
• Install from source

  git clone https://github.com/MaJaHa95/ml-service-wrapper.git
  cd ml-service-wrapper
  pip install .
  

Implementing an ML service

Write a class that matches the interface defined by Service:

import mlservicewrapper
import mlservicewrapper.core.services

class SampleService(mlservicewrapper.core.services.Service):
    async def load(self, ctx: mlservicewrapper.core.contexts.ServiceContext):
        pass

    async def process(self, ctx: mlservicewrapper.core.contexts.ProcessContext):
        pass

    def dispose(self):
        pass

The three functions describe the three phases of service lifetime:

1. load is called when the service is first initializing, and should load models into memory and do necessary pre-work. It makes sense to parse and store ServiceContext parameters in this function, as they won't be accessible later.
2. process will be called many times for each load, and is where data should actually be handled. See ProcessContext for how to process data.
3. dispose may or may not be called, and is optional, but should be used for cleanup tasks during service shutdown, e.g. disposal of handles or temporary files.

Contexts

ServiceContext

A ServiceContext object will be passed to the load function when the service is first initialized. It exposes a single function:

• get_parameter_value(name: str, required: bool = True, default: str = None)
  • Used to get a parameter from the environment. These parameters may be sourced from:
    • A configuration file (using the parameters property)
    • Environment variables
    • Other, environment-specific key-value stores
  • Note that all parameter values are either type str or None. It is the implementation's responsibility to parse string input and handle missing values, potentially with use of the default parameter. Numbers will not be parsed.
  • Service parameters are considered required unless required is specified as False, or a non-None value is passed as a default.

ProcessContext

A ProcessContext object is passed to the process function, and exposes key details about a particular execution. It has more functions than a ServiceContext:

• get_input_dataframe(name: str, required: bool = True)
  • Returns a Pandas DataFrame object containing the named input dataset.
  • Note that an optional parameter required may be set to False in rare cases when an input dataset is optional.
• set_output_dataframe(self, name: str, df: pd.DataFrame)
  • Set the named output dataset using an existing Pandas DataFrame
• get_parameter_value(name: str, required: bool = True, default: str = None) -> str
  • Returns execution-specific parameters, not including those defined in the ServiceContext. To use service-level parameters, store them on the service instance.
  • Process parameters are considered required unless required is specified as False, or a non-None value is passed as a default.
  • Heads up: most implementations will not use execution parameters. Consider using ServiceContext parameters instead. It's also advisable to provide sensible default values, either in-code or through ServiceContext parameters.

Depending on the deployment environment, input and output datasets may be sourced from:

• Local CSV files,
• SQL tables or queries,
• JSON documents or request bodies, or
• Other sources...

Validation

Validating input and raising appropriate errors helps callers understand usage of the ML service. Some built-in errors may have special behaviors in supporting environments. Use the one that most specifically describes the problem.

As best practice, work to validate input datasets and parameters as early as possible. For example, test that all required categorical fields are present before performing work to preprocess text ones.

Parameters

• MissingParameterError(name: str, message: str = None)
  • Used internally when a parameter is requested via the get_parameter_value function, but cannot be found on the request. Similarly to the MissingDatasetError, logic is likely best left to the required parameter on that function.
• BadParameterError(name: str, message: str = None)
  • Raise for all other parameter validation errors, e.g. when a string is not parsable.

Datasets

• MissingDatasetFieldError(dataset_name: str, field_name: str, message: str = None)
  • Used when a required field is missing from an input dataset. For example:
• DatasetFieldError(dataset_name: str, field_name: str, message: str = None)
  • Used when a dataset field is present, but is otherwise invalid. Use is implementation-specific, but could describe an unparsable number field, a duplicate value in an expected-unique field, or other like input inconsistencies.
• MissingDatasetError(dataset_name: str, message: str = None)
  • Thrown internally when a call to get_input_dataframe is made when no dataset exists by the name. It is unlikely implementations will benefit from calling this error directly, and should defer to the ProcessContext using the required parameter on get_input_dataframe.
• BadDatasetError(dataset_name: str, message: str = None)
  • Base class for other errors, callable when a dataset does not match the agreed-upon contract.

Configuration files

Each service is accompanied by a simple JSON configuration file, that tells the wrapper some basic details about the service.

• modulePath
  • The path, relative to the configuration file, where the Service inheritor is defined.
• className
  • The name of the Service inheritor itself. Note that this class must be initializable with no parameters; those should be saved for the load function.
• serviceInstanceName
  • In cases when you choose to instantiate the Service yourself, the name of the instance exposed by the module.
  • Only used if className is omitted.
• parameters
  • An optional dictionary of configuration-specific key-value pairs, which should be passed via the ServiceContext parameters. This is useful when multiple configuration files can be used for the same Service.
• meta
  • Application-level metadata, not passed to the Service. Useful for managing configurations internally.

Note that modulePath is required, along with either className or serviceInstanceName.

Naming Conventions

By nature of its configurability, external type-checking is not possible. Validation can help to identify errors, but only at runtime. As consequence, a consistent naming schema is important to reducing unnecessary cycles.

Follow simple rules when naming parameters and datasets:

• Use Pascal Case (e.g. MaxValue)
• Use only letters and numbers
• Be concise, but descriptive
• Be consistent with casing

Debugging a service

Examples below use the sample service, and consequently require cloning of this repository to run.

Note that for all debug executions, datasets are discovered in the input directory based on name without extension. For example, await ctx.get_input_dataframe("Input") will match $input_dir/Input.csv. Name collisions (i.e. when two file names match the desired dataset but they have different extensions) will result in an error.

Write your own debug script

See ./sample/1_simple/src/debug.py for an example.

Test end-to-end using configuration file

Call the debug module directly:

python -m mlservicewrapper.core.debug \
    --config "./sample/1_simple/src/config.json" \
    --input-dir "./sample/1_simple/data/input" \
    --load-params ModBy=3

Performance evaluation

Call the debug module with --split-dataset-for-perf <dataset name> to perform timing. The named dataset, sourced from the input directory, will be split into individual rows, and passed independently to the process context. Timing information will be printed after execution completes.

This is meant to simulate a non-batching HTTP client.

python -m mlservicewrapper.core.debug \
    --config "./sample/1_simple/src/config.json" \
    --input-dir "./sample/1_simple/data/input" \
    --load-params ModBy=3
    --split-dataset-for-perf Data

Accuracy Evaluation

Any form of model evaluation is welcome, but a helper is exposed for assessing accuracy for labeling tasks.

Call the debug module with --assess-accuracy <input dataset name>.<input dataset field>=<output dataset name>.<output dataset field> to calculate the percentage of time the output dataset field matches the input dataset.

This only works, of course, when the input dataset contains a truth label field.