Writing data through HWC

A step-by-step procedure walks you through connecting to HiveServer (HS2) to perform batch writes from Spark, which is recommended for production. You configure HWC for the managed table write, launch the Spark session, and write ACID, managed tables to Apache Hive.

  • Accept the default spark.datasource.hive.warehouse.load.staging.dir for the temporary staging location required by HWC.
  • Check that spark.hadoop.hive.zookeeper.quorum is configured.
  • Set Kerberos configurations for HWC, or for an unsecured cluster, set spark.security.credentials.hiveserver2.enabled=false.
Limitation: Only the ORC format is supported for writes.

The way data is written from HWC is not impacted by the read modes configured for HWC. For write operations, HWC writes to an intermediate location (as defined by the value of config spark.datasource.hive.warehouse.load.staging.dir) from Spark, followed by executing a "LOAD DATA" query in hive via JDBC. Exception: writing to dynamic partitions creates and intermediate temporary external table.

Using HWC to write data is recommended for production in CDP.

  1. Open a terminal window, start the Apache Spark session, and include the URL for HiveServer.
    spark-shell --jars /opt/cloudera/parcels/CDH/jars/hive-warehouse-connector-assembly-<version>.jar \
    -- conf spark.sql.hive.hiveserver2.jdbc.url=<JDBC endpoint for HiveServer>
    ...                
  2. Include in the launch string a configuration of the intermediate location to use as a staging directory.
    Example syntax:
    ...
    --conf spark.datasource.hive.warehouse.load.staging.dir=<path to directory>
  3. Write a Hive managed table.
    For example, in Scala:
    import com.hortonworks.hwc.HiveWarehouseSession
    import com.hortonworks.hwc.HiveWarehouseSession._
    
    val hive = HiveWarehouseSession.session(spark).build();
    hive.setDatabase("tpcds_bin_partitioned_orc_1000");
    val df = hive.executeQuery("select * from web_sales");
    df.createOrReplaceTempView("web_sales");
    hive.setDatabase("testDatabase");
    
    hive.createTable("newTable").ifNotExists()
     .column("ws_sold_time_sk", "bigint")
     .column("ws_ship_date_sk", "bigint")
     .create();
    
    sql("SELECT ws_sold_time_sk, ws_ship_date_sk FROM web_sales WHERE ws_sold_time_sk > 80000)
     .write.format(HIVE_WAREHOUSE_CONNECTOR)
     .mode("append")
     .option("table", "newTable")
     .save();           

    HWC internally fires the following query to Hive through JDBC:

    LOAD DATA INPATH '<spark.datasource.hive.warehouse.load.staging.dir>' INTO TABLE tpcds_bin_partitioned_orc_1000.newTable 
  4. Write to a statically partitioned, Hive managed table named t1 having two partitioned columns c1 and c2.
    df.write.format(HIVE_WAREHOUSE_CONNECTOR).mode("append").option("partition", "c1='val1',c2='val2'").option("table", "t1").save();
    HWC internally fires the following query to Hive through JDBC after writing data to a temporary location.
    LOAD DATA INPATH '<spark.datasource.hive.warehouse.load.staging.dir>' [OVERWRITE] INTO TABLE db.t1 PARTITION (c1='val1',c2='val2'); 
  5. Write to a dynamically partitioned table named t1 having two partitioned cols c1 and c2.
    df.write.format(HIVE_WAREHOUSE_CONNECTOR).mode("append").option("partition", "c1='val1',c2").option("table", "t1").save(); 
    HWC internally fires the following query to Hive through JDBC after writing data to a temporary location.
    CREATE TEMPORARY EXTERNAL TABLE db.job_id_table(cols....) STORED AS ORC LOCATION '<spark.datasource.hive.warehouse.load.staging.dir>';
    
    INSERT INTO TABLE t1 PARTITION (c1='val1',c2)  SELECT <cols> FROM db.job_id_table;         

    where <cols> should have comma separated list of columns in the table with dynamic partition columns being the last in the list and in the same order as the partition definition.