Writing Data to Hudi Tables In Batches

Scenario

Hudi provides multiple write modes. For details, see the configuration item hoodie.datasource.write.operation. This section describes upsert, insert, and bulk_insert.

• insert: The operation process is similar to upsert. The query on updated file partitions is not based on indexes. Therefore, insert is faster than upsert. This operation is recommended for data sources that do not contain updated data. If the data source contains updated data, the data lake will have duplicate data.
• bulk_insert (insert in batches): It is used for initial dataset loading. This operation sorts primary keys and then inserts data into a Hudi table by writing data to a common Parquet table. It has the best performance but cannot control small files. The upsert and insert operations can control small files by using heuristics.
• upsert (insert and update): It is the default operation type. Hudi determines whether historical data exists based on the primary key. Historical data is updated, and other data is inserted. This operation is recommended for data sources, such as change data capture (CDC), that include updated data.

• Primary keys are not sorted during insert. Therefore, you are not advised to use insert during dataset initialization.
• You are advised to use insert if data is new, use upsert if data needs to be updated, and use bulk_insert if datasets need to be initialized.

Writing Data to Hudi Tables In Batches

1. Import the Hudi package to generate test data. For details, see 2 to 4 in Creating a Hudi Table Using Spark Shell.
2. Write data to the Hudi table. Add the option("hoodie.datasource.write.operation", "bulk_insert") parameter to the write command and set the write mode to bulk_insert. For details about how to specify other write modes, see Table 1.

   df.write.format("org.apache.hudi").
   options(getQuickstartWriteConfigs).
   option("hoodie.datasource.write.precombine.field", "ts").
   option("hoodie.datasource.write.recordkey.field", "uuid").
   option("hoodie.datasource.write.partitionpath.field", "").
   option("hoodie.datasource.write.operation", "bulk_insert").
   option("hoodie.table.name", tableName).
   option("hoodie.datasource.write.keygenerator.class", "org.apache.hudi.keygen.NonpartitionedKeyGenerator").
   option("hoodie.datasource.hive_sync.enable", "true").
   option("hoodie.datasource.hive_sync.partition_fields", "").
   option("hoodie.datasource.hive_sync.partition_extractor_class", "org.apache.hudi.hive.NonPartitionedExtractor").
   option("hoodie.datasource.hive_sync.table", tableName).
   option("hoodie.datasource.hive_sync.use_jdbc", "false").
   option("hoodie.bulkinsert.shuffle.parallelism", 4).
   mode(Overwrite).
   save(basePath)

   

   • For details about the parameters in the example, see Table 1.
   • If the Spark DataSource API is used to update the MOR table, small files of the updated data may be merged when a small volume of data is inserted. As a result, some updated data can be found in the read-optimized view of the MOR table.
   • If the base file of the data to be updated is a small file, the data to be inserted and new data for update are merged with the base file to generate a new base file instead of being written to logs.

Configuring Partitions

Hudi supports multiple partitioning modes, such as multi-level partitioning, non-partitioning, single-level partitioning, and partitioning by date. You can select a proper partitioning mode as required. The following describes how to configure different partitioning modes for Hudi.