SparkML-spring-scoring-poc

POC of socring rest service od Spark ML Pipelines. The

Motivation

Serve Apache spark pipelines from Kubernetes

Architecure

REST JSON Request -> K8s service -> Apache Spark

Getting started

in the code data/pipline-archive it pyspark pipeline

1. Create serving pipeline

Create simple pipeline and Save pipeline

from pyspark.ml import Pipeline
from pyspark.ml.classification import LogisticRegression
from pyspark.ml.feature import HashingTF, Tokenizer
from pyspark import SparkContext
from pyspark.streaming import StreamingContext
from pyspark.sql.session import SparkSession

sc = SparkContext("local[2]", "SparkML")
spark = SparkSession(sc)

# Prepare training documents from a list of (id, text, label) tuples.
training = spark.createDataFrame([
    (0, "a b c d e spark", 1.0),
    (1, "b d", 0.0),
    (2, "spark f g h", 1.0),
    (3, "hadoop mapreduce", 0.0)
], ["id", "text", "label"])

# Configure an ML pipeline, which consists of three stages: tokenizer, hashingTF, and lr.
tokenizer = Tokenizer(inputCol="text", outputCol="words")
hashingTF = HashingTF(inputCol=tokenizer.getOutputCol(), outputCol="features")
lr = LogisticRegression(maxIter=10, regParam=0.001)
pipeline = Pipeline(stages=[tokenizer, hashingTF, lr])

# Fit the pipeline to training documents.
model = pipeline.fit(training)

# This is the save pipeline method
model.save("/some-where-ml/project1/pipeline")

2. Adding /some-where-ml/project1/mlserving.json

{
    "name":"Test pipeline",
    "pipeline":"simple_pipline1",
    "schema":[
        {"name":"name","type":"STRING", "isNullable":false}
    ],
    "sample":{
        "name":"Bob"
    },
    "output": ["name", "features"]
}

3. Zip

No zip the folder /some-where-ml/project1 and put it in your pipelines.folder

zip -r spark-sample-pipeline.zip /some-where-ml/project1 

4. Runing & Executing

in application.properties change the pipelines.folder to pipeline store folder -OR- run as paramete (like below)

# Build mvn
mvn install package

# Run
java -Dpipelines.folder=/some-where-ml -jar target/mlserver.jar

5. Call REST

curl -X POST \
	 -H "Content-Type: application/json" \
	 -d '[{"text":"Yehuda"}]' \
	 http://localhost:9900/predict/spark-sample-pipeline

6. (OPTIONAL) What next?

• add Warm-up for modules
• dockerize and serve it as service

Dockerization

To do some docker build

mvn clean install package && docker build  -t alefbt/spring-mlspark-serving .

Then run

docker run  --rm -it -p 9900:8080  alefbt/spring-mlspark-serving

Notes

• This is code is not optimized to sub-second serving, it's possibol <,i did it on other project ;-) in order to do it, you need do some cacheing>
• Code contribution is welcome !
• Remember: this porject is POC.

Known issues

• FIXED. Snappy (xerial.snappy package) doing some problems when running on vanilla Docker openjdk:8-jdk-alpine - see the Dockerfile explaines the walkaround

Licence

MIT - Free.

Use and Contribute