Using the H2O Model Evaluator Operator

Introduction

The TIBCO StreamBase® Operator For H2O Model Evaluator enables StreamBase applications to execute numerical models generated with the H2O Fast Scalable Machine Learning package. See http://www.h2o.ai/ and this external documentation.

H2O is distributed model training and evaluation software. With an H2O cluster it is possible to process much larger datasets than with traditional single process solutions. H2O is frequently used in Big Data solutions to process the data and build predictive models. An additional advantage of H2O is the runtime-oriented model representation (POJO/GenModel). With the H2O operator, it is possible to deploy an arbitrary number of models to be executed against incoming events. The EventFlow designer's responsibility is conversion of the incoming data into the models' expected features. It frequently requires event enrichment, such as cross-referencing, attribute lookup, or prior events related to the same context.

The operator processes input data given as a tuple or a list of tuples. The tuple schema corresponds to the input parameters of the models. For each model, the operator generates output data that matches the defined output schema. Depending on the input data, the output can be a single tuple or a list of tuples.

Dynamic model definitions allow you to provide additional metadata to the deployed models. The metadata is attached to the model result, allowing the EventFlow to take action based on the model attributes. Examples of attributes include: champion/challenger flag, category for propensity scoring, and so on.

The H2O model evaluator is implemented using the H2O library.

Model operators support an arbitrary number of models simultaneously, as well as scoring single samples and data frames.

Operator Properties

This section describes the properties you can set for this operator, using the various tabs of the Properties view in StreamBase Studio.

General Tab

Name: Use this required field to specify or change the name of this instance of this component, which must be unique in the current EventFlow module. The name must contain only alphabetic characters, numbers, and underscores, and no hyphens or other special characters. The first character must be alphabetic or an underscore.

Operator: A read-only field that shows the formal name of the operator.

Class name: Shows the fully qualified class name that implements the functionality of this operator. If you need to reference this class name elsewhere in your application, you can right-click this field and select Copy from the context menu to place the full class name in the system clipboard.

Start options: This field provides a link to the Cluster Aware tab, where you configure the conditions under which this operator starts.

Enable Error Output Port: Select this check box to add an Error Port to this component. In the EventFlow canvas, the Error Port shows as a red output port, always the last port for the component. See Using Error Ports to learn about Error Ports.

Description: Optionally enter text to briefly describe the component's purpose and function. In the EventFlow Editor canvas, you can see the description by pressing Ctrl while the component's tooltip is displayed.

Operator Properties Tab

Property Type Description
Control Port check box Enables dynamic reconfiguration of the model list. The control port also enables the control output port which reports status of the model loading request. The control port supports all-or-nothing semantics. That is, either the full list successfully loads and replaces the currently deployed models, or it reports failure.
Status Port check box Enables failure notifications. If the scoring fails, the failure is emitted to the status port, including the original input tuple.
Timing Info check box Fine granular timing information. It collects the effective times of input conversion, model evaluation, and output conversion. Time is in nanoseconds.
Log Level INFO Controls the level of verbosity the adapter uses to issue informational traces to the console. This setting is independent of the containing application's overall log level. Available values, in increasing order of verbosity, are: OFF, ERROR, WARN, INFO, DEBUG, TRACE.

Models Tab

Property Type Description
Model type enumeration Model representation type.
  • Compiled - precompiled POJO class statically available on the classpath.

  • POJO - POJO source code exported from H2O model.

  • Native - native serialized model saved with H2O.

Note that native serialization formats require H2O libraries in the correct version available to the StreamBase.
Model URLs name/value pairs List of design-time specified models. The models consist of name and URL pointing to the model definition. Models can also be loaded from HDFS.

Schemas

Property Type Description
Output schema type enumeration Type of result representation.
  • Raw - list of doubles directly returned from the model.

  • Generic - generic schema supporting regression and binomial and multinomial classification.

  • Custom - subset of Generic fields to remove clutter.

Result Data Schema schema Anticipated schema for model output. Only fields defined in the schema are used in the output tuple.

The Custom schema contains subset of Generic schema fields. The fields filled by operator are:

  • prediction - double value representing regression result

  • class - string value representing result class label

  • score - score of the result class

  • probability - probability of the result class

  • probabilities - list of scores for each class

  • probabilities.class - class label

  • probabilities.probability - class probability

Note

that in H2O the resulting interpretation depends on the model type.

  • regression - prediction field is filled

  • binomial classification - class and score are filled. 1/TRUE class result is passed as the score. In H2O the default discrimination threshold maximizes the F1 score.

  • multinomial classification - class and probability are filled. For each response, the class and entry in probabilities are provided. The winner class is the one with the highest probability.

AMS Tab

Use the AMS tab to specify which artifacts should be pulled from a running TIBCO Artifact Management Server, which is a separately installed product.

Note

If you deploy an artifact from the AMS system, it will first check your list of artifacts to match the path and if matched will use the model name given. If the path is not matched, then the artifact's filename is used without the file extension as the model name. Example sample/audit.rds would resolve to a model name of audit.

Property Data Type Description
Required On Startup check box When enabled, the artifacts listed are requested from AMS at initialization and the system waits until all artifacts are loaded.
Artifacts list (string, string) List of artifacts to load from AMS. The first value of the path is the project name followed by the full path to the artifact. Use a / separator with an optional @version at the end. If @version is not specified, then the latest version is assumed.

For example: project/path1/path2/artifact@1

Cluster Aware Tab

Use the settings in this tab to allow this operator or adapter to start and stop based on conditions that occur at runtime in a cluster with more than one node. During initial development of the fragment that contains this operator or adapter, and for maximum compatibility with TIBCO Streaming releases before 10.5.0, leave the Cluster start policy control in its default setting, Start with module.

Cluster awareness is an advanced topic that requires an understanding of StreamBase Runtime architecture features, including clusters, quorums, availability zones, and partitions. See Cluster Awareness Tab Settings on the Using Cluster Awareness page for instructions on configuring this tab.

Concurrency Tab

Use the Concurrency tab to specify parallel regions for this instance of this component, or multiplicity options, or both. The Concurrency tab settings are described in Concurrency Options, and dispatch styles are described in Dispatch Styles.

Caution

Concurrency settings are not suitable for every application, and using these settings requires a thorough analysis of your application. For details, see Execution Order and Concurrency, which includes important guidelines for using the concurrency options.

Data Input Port

The data port is the default input port for the model operator. It is always enabled. Use data port to execute the model scoring.

The default schema for the data input port is:

  • frame, tuple or list(tuple). Samples to be scored by the deployed models.

    The tuple structure contains primitive fields (int, long, double, string or boolean) with names corresponding to model input fields.

    Sparse dictionaries. A sparse dictionary is a list of tuples with name (string) or idx (int) and value (any supported primitive type) fields. Sparse dictionaries are used when 1) the fields are not known at the design time, 2) StreamBase does not support the field names, 3) the number of fields is large. The input tuple may contain any number of sparse dictionaries. For example, to provide categorical and continuous values or fields from various domains.

  • modelName (optional), string. If this field exists and is not null it will be used to specify which model the input tuple will score against. If this field is missing or null then all models are scored.

  • * arbitrary pass through parameters

Unrecognized fields are transparently passed. The frame field is not propagated; the scores field is not allowed.

Scores Output Port

The scores port provides a list of model evaluation results.

The schema for the scores output port is:

  • scores, list(tuple). List of record for each currently deployed model.

  • scores.modelName, string. Name of the model defined in the Model URLs or provided by the control port.

  • scores.modelUrl, string. URL defining the model configured in the Model URLs or provided by the control port.

  • scores.modelData, blob. Binary defining the model if used to load the model.

  • scores.score, tuple or list(tuple). The type depends on the type of frame input. List of scores in the same order as the input list. The schema is defined as Result Data Schema property.

  • scores.*. Arbitrary parameters provided during model redeployment on the control port.

  • * parameters other than frame.

The scores port transparently replicates unrecognized fields. The frame field is not propagated.

Control Input Port

The control port enables runtime redeployment of models. The models are deployed in all-or-nothing semantics. That means if all the provided models are successfully loaded, they fully replace the current set.

The schema for the control input port is:

  • models, list(tuple). List of record for each model to be deployed.

  • models.modelName, string. Logical name of the model.

  • models.modelUrl, string. URL defining the model.

  • models.modelData, blob. The binary data of the model. This field can be used to load a model directly from any source. This field is only used if models.modelUrl is null or empty.

  • models.*. Arbitrary parameters describing the model. They are later provided in the score.

  • *. Arbitrary parameters provided during model redeployment on the control port.

The status port transparently replicates unrecognized fields; do not use the status or message fields on the input port.

Status Output Port

The status port provides responses for runtime model deployment. The tuples are emitted only as responses to the control port tuples.

The schema for the status output port is:

  • status, string. Status of deployment. Either success or failure.

  • message, string. Descriptive status message.

  • models, list(tuple). List of record for each model to be deployed.

  • models.status, string. Status of the model loading. Either success or failure.

  • models.message, string. Descriptive model status message.

  • models.modelName, string. Logical name of the model.

  • models.modelUrl, string. URL defining the model.

  • models.modelData, blob. Binary defining the model if used to load the model.

  • models.*. Arbitrary parameters describing the model. They are later provided in the score.

  • * parameters other than models.

The status port transparently replicates unrecognized fields from the control port.