The configuration of a maki nage application is done via a yaml configuration file. This configuration is organized into dedicated sections for each element of the application.
This configuration file is the only parameter used by the Maki Nage CLIs (makinage and makinage-serve), as the config parameter.
In the following documentation, each field is documented with a type. The meaning of each type is:
object: A yaml mapping
list: A yaml collection
string: A string
module: A string describing a python module. The syntax follows the python “import” notation, i.e. modules separated by dots.
function: A string describing a python function. The syntax is a module name followed by a function name. Both are separated by “:”.
This section contains information on the application.
name: The name of the application. This name is used as the kafka consumer group.
This section contains the kafka configuration. It contains the following fields:
endpoint: [string] The kafka server endpoint. e.g. “localhost”
This sections describes the different topics that can be used by the operators. Each entry contains the following fields:
name: [string] values
encoder: [module, optional] The encoder used for kafka records. default: “makinage.encoding.string”
partition_selector: [function, optional]. default: int(random.random() * 1000)
start_from: [string, optional]. Defines how records are consumed on service reload. possible values are [end|beginning|last]. default: “end”
This section contains the list of operators in the application. Each operator field is a named object of this section. Each operator object contains the following fields:
factory: [function] The name of the operator factory.
sources: [list, optional] The list of source observables for this operator.
sinks: [list, optional] The list of sink observables for this operator.
This section is a placeholder for application specific configuration. This is typically where one can put the feature engineering parameters, or anything that can be configured without code change.
It contains some reserved sub-section keywords:
This sub-section contains the configuration for model serving. The following fields can be used to configure the inference behavior:
pre_transform: [function] A factory function called with the configuration as a parameter. It must return a function that takes utterances as parameter and returns data in a format suitable for inference.
post_transform: [function] A factory function called with the configuration as a parameter. It must return a function that takes utterances and predicions as parameters and returns data in a format suitable for emission on the sink topic.
predict: [function] A factory function called with the model and configuration as a parameters. It must return a function that takes pre-transformed utterances as parameter and return the model prediction.
See the Serving chapter for more information on these fields.
If the configuration file contains a redirect section the the actual content of the configuration is retrived from the location described in this section.
It contains the following fields:
connector: [string] The type of redirection to do. Only consul is supported for now.
endpoint: [string] The consul endpoint url. e.g. “http://localhost:8500”
key: [string] The name of the KV Store key to read. e.g. “myservice”
When consul is used for the configuration, then changes are monitored and exposed in real-time on the config argument of the operators. This allows to dynamically change the configuration of a running service.