Slow Performance of Elastic Search with Spark

Elastic Stack Elasticsearch
1

I am trying to read the data from Elastic Search into a dataframe using Java ES-Spark-connector.. Now when I try to execute a query for example count() on the dataframe, the performance is dismal. For a 3Gb data it takes around 6 mins. On the other hand if I save the data in hadoop/hdfs it and then read it from there it takes around 3s. Can some one tell me a work around for this. The code I am using is as below.

SparkConf conf = new SparkConf().setAppName("Simple App").setMaster("local[*]");
        conf.set("es.index.auto.create", "true");
        JavaSparkContext javaSparkContext = new JavaSparkContext(conf);
       
        SparkSession spark = new SparkSession(javaSparkContext.sc());
        SQLContext sqlContext = new SQLContext(spark);
        Dataset<Row> ds = spark.read().format("org.elasticsearch.spark.sql").load("index_name");
       count = ds.count();   // this takes around 6 mins for 3GB data
2

I am also having performance problems, the same Spark query I do from the raw JSON files is 12-15 times faster than the same query via ES.
I have about this is on the github issues but they told me to ask in here, anyway we need to find a way to inspect the push down query to understand what's going on.

4

And even if you do this:

ds.count().explain(extended=True)

does not show you the ES queries.

I have another discussion open here for monitoring queries in the backend which seems the only solution for now.

5

Please read this too because what is essentially happening is that the ES driver is not that smart and does just a full scroll of the data instead of leveraging the ES capabilities.