REST web service for scoring PMML models.
- Full support for PMML specification versions 3.0 through 4.3. The evaluation is handled by the JPMML-Evaluator library.
- Simple and powerful REST API:
- Model deployment and undeployment.
- Model evaluation in single prediction, batch prediction and CSV prediction modes.
- Model metrics.
- High performance and high throughput:
- Sub-millisecond response times.
- Request and response compression using
gzipanddeflateencodings. - Thread safe.
- Open, extensible architecture for easy integration with proprietary systems and services:
- User authentication and authorization.
- Metrics dashboards.
The project requires Java 1.8 or newer to run.
Enter the project root directory and build using Apache Maven:
mvn clean install
The example PMML file DecisionTreeIris.pmml along with example JSON and CSV files can be found in the openscoring-service/src/etc directory.
The build produces an executable uber-JAR file openscoring-server/target/server-executable-1.4-SNAPSHOT.jar. Change the working directory to openscoring-server and execute the following command:
java -jar target/server-executable-1.4-SNAPSHOT.jar
By default, the REST web service is started at http://localhost:8080/openscoring. The main class org.openscoring.server.Main accepts a number of configuration options for URI customization and other purposes. Please specify --help for more information.
The working directory contains a sample Java logging configuration file logging.properties.sample that should be copied over to a new file logging.properties and customized to current needs. A Java logging configuration file can be imposed on the JVM by defining the java.util.logging.config.file system property:
java -Djava.util.logging.config.file=logging.properties -jar target/server-executable-1.4-SNAPSHOT.jar
Additionally, the working directory contains a sample Typesafe's Config configuration file application.conf.sample that should be copied over to a new file application.conf and customized to current needs. This local configuration file can be imposed on the JVM by defining the config.file system property:
java -Dconfig.file=application.conf -jar target/server-executable-1.4-SNAPSHOT.jar
The local configuration file overrides the default configuration that is defined in the reference REST web service configuration file openscoring-service/src/main/reference.conf. For example, the following configuration file selectively overrides the list-valued modelRegistry.visitorClasses property:
modelRegistry {
visitorClasses = [
"org.jpmml.model.visitors.LocatorNullifier" // Erases SAX Locator information from the PMML class model object, which will considerably reduce the memory consumption of deployed models
]
}
The build produces a WAR file openscoring-webapp/target/openscoring-webapp-1.4-SNAPSHOT.war. This WAR file can be deployed using any Java web container.
The web application can be launced using Jetty Maven Plugin. Change the working directory to openscoring-webapp and execute the following command:
mvn jetty:run-war
The build produces an executable uber-JAR file openscoring-client/target/client-executable-1.4-SNAPSHOT.jar. Change the working directory to openscoring-client and replay the life cycle of a sample DecisionTreeIris model (in "REST API", see below) by executing the following sequence of commands:
java -cp target/client-executable-1.4-SNAPSHOT.jar org.openscoring.client.Deployer --model http://localhost:8080/openscoring/model/DecisionTreeIris --file DecisionTreeIris.pmml
java -cp target/client-executable-1.4-SNAPSHOT.jar org.openscoring.client.Evaluator --model http://localhost:8080/openscoring/model/DecisionTreeIris -XSepal_Length=5.1 -XSepal_Width=3.5 -XPetal_Length=1.4 -XPetal_Width=0.2
java -cp target/client-executable-1.4-SNAPSHOT.jar org.openscoring.client.CsvEvaluator --model http://localhost:8080/openscoring/model/DecisionTreeIris --input input.csv --output output.csv
java -cp target/client-executable-1.4-SNAPSHOT.jar org.openscoring.client.Undeployer --model http://localhost:8080/openscoring/model/DecisionTreeIris
Additionally, this JAR file contains an application class org.openscoring.client.DirectoryDeployer, which monitors the specified directory for PMML file addition and removal events:
java -cp target/client-executable-1.4-SNAPSHOT.jar org.openscoring.client.DirectoryDeployer --model-collection http://localhost:8080/openscoring/model --dir pmml
Model REST API endpoints:
| HTTP method | Endpoint | Required role(s) | Description |
|---|---|---|---|
| GET | /model | - | Get the summaries of all models |
| PUT | /model/${id} | admin | Deploy a model |
| GET | /model/${id} | - | Get the summary of a model |
| GET | /model/${id}/pmml | admin | Download a model as a PMML document |
| POST | /model/${id} | - | Evaluate data in "single prediction" mode |
| POST | /model/${id}/batch | - | Evaluate data in "batch prediction" mode |
| POST | /model/${id}/csv | - | Evaluate data in "CSV prediction" mode |
| DELETE | /model/${id} | admin | Undeploy a model |
Metric REST API endpoints:
| HTTP method | Endpoint | Required role(s) | Description |
|---|---|---|---|
| GET | /metric/model | admin | Get the metric sets of all models |
| GET | /metric/model/${id} | admin | Get the metric set of a model |
By default, the "admin" role is granted to all HTTP requests that originate from the local network address.
In case of an error (ie. response status codes 4XX or 5XX), the response body is a JSON serialized form of an org.openscoring.common.SimpleResponse (source) object.
Java clients may use the following idiom to check if an operation succeeded or failed:
ModelResponse response = ...;
// The error condition is encoded by initializing the "message" field and leaving all other fields uninitialized
String message = response.getMessage();
if(message != null){
throw new RuntimeException(message);
}
// Proceed as usualCreates or updates a model.
The request body is a PMML document (indicated by content-type header text/xml or application/xml).
The response body is a JSON serialized form of an org.openscoring.common.ModelResponse (source) object.
Response status codes:
- 200 OK. The model was updated.
- 201 Created. A new model was created.
- 400 Bad Request. The deployment failed permanently. The request body is not a valid and/or supported PMML document.
- 403 Forbidden. The acting user does not have an "admin" role.
- 500 Internal Server Error. The deployment failed temporarily.
Sample cURL invocation:
curl -X PUT --data-binary @DecisionTreeIris.pmml -H "Content-type: text/xml" http://localhost:8080/openscoring/model/DecisionTreeIris
The same, using the gzip encoding:
curl -X PUT --data-binary @DecisionTreeIris.pmml.gz -H "Content-encoding: gzip" -H "Content-type: text/xml" http://localhost:8080/openscoring/model/DecisionTreeIris
Gets the summaries of all models.
The response body is a JSON serialized form of an org.openscoring.common.BatchModelResponse (source) object.
Response status codes:
- 200 OK. The model collection was queried.
Sample cURL invocation:
curl -X GET http://localhost:8080/openscoring/model
Gets the summary of a model.
The response body is a JSON serialized form of an org.openscoring.common.ModelResponse (source) object.
Response status codes:
- 200 OK. The model was queried.
- 404 Not Found. The requested model was not found.
Sample cURL invocation:
curl -X GET http://localhost:8080/openscoring/model/DecisionTreeIris
Sample response:
{
"id" : "DecisionTreeIris",
"miningFunction" : "classification",
"summary" : "Tree model",
"properties" : {
"created.timestamp" : "2015-03-17T12:41:35.933+0000",
"accessed.timestamp" : "2015-03-21T09:35:58.582+0000",
"file.size" : 4306,
"file.md5sum" : "2d4698076ed807308c5ae40563b70345"
},
"schema" : {
"inputFields" : [
{
"id" : "Sepal_Length",
"name" : "Sepal length in cm",
"dataType" : "double",
"opType" : "continuous",
"values" : [ "[4.3, 7.9]" ]
},
{
"id" : "Sepal_Width",
"name" : "Sepal width in cm",
"dataType" : "double",
"opType" : "continuous",
"values" : [ "[2.0, 4.4]" ]
},
{
"id" : "Petal_Length",
"name" : "Petal length in cm",
"dataType" : "double",
"opType" : "continuous",
"values" : [ "[1.0, 6.9]" ]
},
{
"id" : "Petal_Width",
"name" : "Petal width in cm",
"dataType" : "double",
"opType" : "continuous",
"values" : [ "[0.1, 2.5]" ]
}
],
"groupFields" : [],
"targetFields" : [
{
"id" : "Species",
"dataType" : "string",
"opType" : "categorical",
"values" : [ "setosa", "versicolor", "virginica" ]
}
],
"outputFields" : [
{
"id" : "Probability_setosa",
"dataType" : "double",
"opType" : "continuous"
},
{
"id" : "Probability_versicolor",
"dataType" : "double",
"opType" : "continuous"
},
{
"id" : "Probability_virginica",
"dataType" : "double",
"opType" : "continuous"
},
{
"id" : "Node_Id",
"dataType" : "string",
"opType" : "categorical"
}
]
}
}Field definitions are retrieved from the MiningSchema and Output elements of the PMML document. The input and group-by fields relate to the arguments attribute of the evaluation request, whereas the target and output fields relate to the result attribute of the evaluation response (see below).
Downloads a model.
The response body is a PMML document.
Response status codes:
- 200 OK. The model was downloaded.
- 403 Forbidden. The acting user does not have an "admin" role.
- 404 Not Found. The requested model was not found.
Sample cURL invocation:
curl -X GET http://localhost:8080/openscoring/model/DecisionTreeIris/pmml
Evaluates data in "single prediction" mode.
The request body is a JSON serialized form of an org.openscoring.common.EvaluationRequest (source) object.
The response body is a JSON serialized form of an org.openscoring.common.EvaluationResponse (source) object.
Response status codes:
- 200 OK. The evaluation was successful.
- 400 Bad Request. The evaluation failed permanently due to missing or invalid input data.
- 404 Not Found. The requested model was not found.
- 500 Internal Server Error. The evaluation failed temporarily.
Sample cURL invocation:
curl -X POST --data-binary @EvaluationRequest.json -H "Content-type: application/json" http://localhost:8080/openscoring/model/DecisionTreeIris
Sample request:
{
"id" : "record-001",
"arguments" : {
"Sepal_Length" : 5.1,
"Sepal_Width" : 3.5,
"Petal_Length" : 1.4,
"Petal_Width" : 0.2
}
}Sample response:
{
"id" : "record-001",
"result" : {
"Species" : "setosa",
"Probability_setosa" : 1.0,
"Probability_versicolor" : 0.0,
"Probability_virginica" : 0.0,
"Node_Id" : "2"
}
}Evaluates data in "batch prediction" mode.
The request body is a JSON serialized form of an org.openscoring.common.BatchEvaluationRequest (source) object.
The response body is a JSON serialized form of an org.openscoring.common.BatchEvaluationResponse (source) object.
Response status codes:
- 200 OK. The evaluation was successful.
- 400 Bad Request. The evaluation failed permanently due to missing or invalid input data.
- 404 Not Found. The requested model was not found.
- 500 Internal Server Error. The evaluation failed temporarily.
Sample cURL invocation:
curl -X POST --data-binary @BatchEvaluationRequest.json -H "Content-type: application/json" http://localhost:8080/openscoring/model/DecisionTreeIris/batch
The evaluation is performed at "record" isolation level. If the evaluation of some org.openscoring.common.EvaluationRequest object fails, then the corresponding org.openscoring.common.EvaluationResponse object encodes the error condition (see above).
Evaluates data in "CSV prediction" mode.
The request body is a CSV document (indicated by content-type header text/plain). The data table must contain a data column for every input and group-by field. The ordering of data columns is not significant, because they are mapped to fields by name.
The CSV reader component detects the CSV dialect by probing ,, ; and \t as CSV delimiter characters. This detection functionality can be suppressed by supplying the value of the CSV delimiter character using the delimiterChar query parameter.
The response body is a CSV document. The data table contains a data column for every target and output field.
The first data column can be employed for row identification purposes. It will be copied over from the request data table to the response data table if its name equals to "Id" (the comparison is case insensitive) and the number of rows did not change during the evaluation.
Response status codes:
- 200 OK. The evaluation was successful.
- 400 Bad request. The evaluation failed permanently. The request body is not a valid and/or supported CSV document, or it contains cells with missing or invalid input data.
- 404 Not Found. The requested model was not found.
- 500 Internal Server Error. The evaluation failed temporarily.
Sample cURL invocation:
curl -X POST --data-binary @input.csv -H "Content-type: text/plain; charset=UTF-8" http://localhost:8080/openscoring/model/DecisionTreeIris/csv > output.csv
The same, using the gzip encoding:
curl -X POST --data-binary @input.csv.gz -H "Content-encoding: gzip" -H "Content-type: text/plain; charset=UTF-8" -H "Accept-encoding: gzip" http://localhost:8080/openscoring/model/DecisionTreeIris/csv > output.csv.gz
Sample request:
Id,Sepal_Length,Sepal_Width,Petal_Length,Petal_Width
record-001,5.1,3.5,1.4,0.2
record-002,7,3.2,4.7,1.4
record-003,6.3,3.3,6,2.5
Sample response:
Id,Species,Probability_setosa,Probability_versicolor,Probability_virginica,Node_Id
record-001,setosa,1.0,0.0,0.0,2
record-002,versicolor,0.0,0.9074074074074074,0.09259259259259259,6
record-003,virginica,0.0,0.021739130434782608,0.9782608695652174,7
The evaluation is performed at "all-records-or-nothing" isolation level. If the evaluation of some row fails, then the whole CSV document fails.
Deletes a model.
The response body is a JSON serialized form of an org.openscoring.common.SimpleResponse (source) object.
Response status codes:
- 200 OK. The model was deleted.
- 403 Forbidden. The acting user does not have an "admin" role.
- 404 Not Found. The requested model was not found.
- 500 Internal Server Error. The undeployment failed temporarily.
Sample cURL invocation:
curl -X DELETE http://localhost:8080/openscoring/model/DecisionTreeIris
An HTTP PUT or DELETE method can be masked as an HTTP POST method by using the HTTP method override mechanism.
Sample cURL invocation that employs the X-HTTP-Method-Override request header:
curl -X POST -H "X-HTTP-Method-Override: DELETE" http://localhost:8080/openscoring/model/DecisionTreeIris
Sample cURL invocation that employs the _method query parameter:
curl -X POST http://localhost:8080/openscoring/model/DecisionTreeIris?_method=DELETE
Gets the snapshot of the metric set of a model.
The response body is a JSON serialized form of an org.openscoring.common.MetricSetResponse (source) object.
Response status codes:
- 200 OK. The evaluation was successful.
- 403 Forbidden. The acting user does not have an "admin" role.
- 404 Not Found. The requested model was not found.
Sample cURL invocation:
curl -X GET http://localhost:8080/openscoring/metric/model/DecisionTreeIris
Sample response:
{
"version" : "3.0.0",
"counters" : {
"records" : {
"count" : 1
}
},
"gauges" : { },
"histograms" : { },
"meters" : { },
"timers" : {
"evaluate" : {
"count" : 1,
"max" : 0.008521913,
"mean" : 0.008521913,
"min" : 0.008521913,
"p50" : 0.008521913,
"p75" : 0.008521913,
"p95" : 0.008521913,
"p98" : 0.008521913,
"p99" : 0.008521913,
"p999" : 0.008521913,
"stddev" : 0.0,
"m15_rate" : 0.19237151525464488,
"m1_rate" : 0.11160702915400945,
"m5_rate" : 0.17797635419760474,
"mean_rate" : 0.023793073545863026,
"duration_units" : "seconds",
"rate_units" : "calls/second"
}
}
}Openscoring is dual-licensed under the GNU Affero General Public License (AGPL) version 3.0, and a commercial license.
Openscoring is developed and maintained by Openscoring Ltd, Estonia.
Interested in using Openscoring software in your application? Please contact info@openscoring.io