Spark is executing every single action two times

Question

I have created a simple Java application that uses Apache Spark to retrieve data from Cassandra, do some transformation on it and save it in another Cassandra table.

I am using Apache Spark 1.4.1 configured in a standalone cluster mode with a single master and slave, located on my machine.

DataFrame customers = sqlContext.cassandraSql("SELECT email, first_name, last_name FROM customer " +
    "WHERE CAST(store_id as string) = '" + storeId + "'");

DataFrame customersWhoOrderedTheProduct = sqlContext.cassandraSql("SELECT email FROM customer_bought_product " +
    "WHERE CAST(store_id as string) = '" + storeId + "' AND product_id = " + productId + "");

// We need only the customers who did not order the product
// We cache the DataFrame because we use it twice.
DataFrame customersWhoHaventOrderedTheProduct = customers
    .join(customersWhoOrderedTheProduct
    .select(customersWhoOrderedTheProduct.col("email")), customers.col("email").equalTo(customersWhoOrderedTheProduct.col("email")), "leftouter")
    .where(customersWhoOrderedTheProduct.col("email").isNull())
    .drop(customersWhoOrderedTheProduct.col("email"))
    .cache();

int numberOfCustomers = (int) customersWhoHaventOrderedTheProduct.count();

Date reportTime = new Date();

// Prepare the Broadcast values. They are used in the map below.
Broadcast<String> bStoreId = sparkContext.broadcast(storeId, classTag(String.class));
Broadcast<String> bReportName = sparkContext.broadcast(MessageBrokerQueue.report_did_not_buy_product.toString(), classTag(String.class));
Broadcast<java.sql.Timestamp> bReportTime = sparkContext.broadcast(new java.sql.Timestamp(reportTime.getTime()), classTag(java.sql.Timestamp.class));
Broadcast<Integer> bNumberOfCustomers = sparkContext.broadcast(numberOfCustomers, classTag(Integer.class));

// Map the customers to a custom class, thus adding new properties.
DataFrame storeCustomerReport = sqlContext.createDataFrame(customersWhoHaventOrderedTheProduct.toJavaRDD()
    .map(row -> new StoreCustomerReport(bStoreId.value(), bReportName.getValue(), bReportTime.getValue(), bNumberOfCustomers.getValue(), row.getString(0), row.getString(1), row.getString(2))), StoreCustomerReport.class);


// Save the DataFrame to cassandra
storeCustomerReport.write().mode(SaveMode.Append)
    .option("keyspace", "my_keyspace")
    .option("table", "my_report")
    .format("org.apache.spark.sql.cassandra")
    .save();

As you can see I cache the customersWhoHaventOrderedTheProduct DataFrame, after that I execute a countand call toJavaRDD.

By my calculations these actions should be executed only once. But when I go in the Spark UI for the current job I see the following stages:

As you can see every action is executed twice.

Am I doing something wrong? Is there any setting that I've missed?

Any ideas are greatly appreciated.

---

EDIT:

After I called System.out.println(storeCustomerReport.toJavaRDD().toDebugString());

This is the debug string:

(200) MapPartitionsRDD[43] at toJavaRDD at DidNotBuyProductReport.java:93 []
  |   MapPartitionsRDD[42] at createDataFrame at DidNotBuyProductReport.java:89 []
  |   MapPartitionsRDD[41] at map at DidNotBuyProductReport.java:90 []
  |   MapPartitionsRDD[40] at toJavaRDD at DidNotBuyProductReport.java:89 []
  |   MapPartitionsRDD[39] at toJavaRDD at DidNotBuyProductReport.java:89 []
  |   MapPartitionsRDD[38] at toJavaRDD at DidNotBuyProductReport.java:89 []
  |   ZippedPartitionsRDD2[37] at toJavaRDD at DidNotBuyProductReport.java:89 []
  |   MapPartitionsRDD[31] at toJavaRDD at DidNotBuyProductReport.java:89 []
  |   ShuffledRDD[30] at toJavaRDD at DidNotBuyProductReport.java:89 []
  +-(2) MapPartitionsRDD[29] at toJavaRDD at DidNotBuyProductReport.java:89 []
     |  MapPartitionsRDD[28] at toJavaRDD at DidNotBuyProductReport.java:89 []
     |  MapPartitionsRDD[27] at toJavaRDD at DidNotBuyProductReport.java:89 []
     |  MapPartitionsRDD[3] at cache at DidNotBuyProductReport.java:76 []
     |  CassandraTableScanRDD[2] at RDD at CassandraRDD.scala:15 []
  |   MapPartitionsRDD[36] at toJavaRDD at DidNotBuyProductReport.java:89 []
  |   ShuffledRDD[35] at toJavaRDD at DidNotBuyProductReport.java:89 []
  +-(2) MapPartitionsRDD[34] at toJavaRDD at DidNotBuyProductReport.java:89 []
     |  MapPartitionsRDD[33] at toJavaRDD at DidNotBuyProductReport.java:89 []
     |  MapPartitionsRDD[32] at toJavaRDD at DidNotBuyProductReport.java:89 []
     |  MapPartitionsRDD[5] at cache at DidNotBuyProductReport.java:76 []
     |  CassandraTableScanRDD[4] at RDD at CassandraRDD.scala:15 []

---

EDIT 2:

So after some research combined with trials and errors, I managed to optimize the job.

I created an RDD from customersWhoHaventOrderedTheProduct and I cache it before I call the count() action. (I moved the cache from the DataFrame to the RDD).

After that I use this RDD to create the storeCustomerReport DataFrame.

JavaRDD<Row> customersWhoHaventOrderedTheProductRdd = customersWhoHaventOrderedTheProduct.javaRDD().cache();

Now the stages look like this:

As you can see the two count and cache are now gone, but there are still two 'javaRDD' actions. I have no idea where they are coming from, as I call toJavaRDD only once in my code.

Answer 1

It looks like you are applying two actions in below code segment

// Map the customers to a custom class, thus adding new properties.
DataFrame storeCustomerReport = sqlContext.createDataFrame(customersWhoHaventOrderedTheProduct.toJavaRDD()
    .map(row -> new StoreCustomerReport(bStoreId.value(), bReportName.getValue(), bReportTime.getValue(), bNumberOfCustomers.getValue(), row.getString(0), row.getString(1), row.getString(2))), StoreCustomerReport.class);


// Save the DataFrame to cassandra
storeCustomerReport.write().mode(SaveMode.Append)
    .option("keyspace", "my_keyspace")

One at sqlContext.createDataFrame() and the other at storeCustomerReport.write() and both of these require customersWhoHaventOrderedTheProduct.toJavaRDD().

Persisting the RDD produced by should solve this issue.

JavaRDD cachedRdd = customersWhoHaventOrderedTheProduct.toJavaRDD().persist(StorageLevel.DISK_AND_MEMORY) //Or any other storage level