[HUDI-4430] Incorrect type casting while reading HUDI table created with CustomKeyGenerator and unixtimestamp paritioning field

• Type: Bug
• Status: Open
• Priority: Critical
• Resolution: Unresolved
• Affects Version/s: 0.12.0
• Fix Version/s: 1.0.0
• Component/s: writer-core
• Story Points: 6
• Assignee: Lokesh Jain
• Reporter: Volodymyr Burenin
• Created: 20/Jul/22 20:56
• Updated: 12/Nov/24 21:35

Description

Hi,

I have discovered an issue that doesn't play nicely with the custom key generatosr, basically anything that is not TimestampBasedKeyGenerator or TimestampBasedAvroKeyGenerator.

While trying to read a table that was created with these parameters(the rest don't matter):

hoodie.datasource.write.recordkey.field=query_id,event_type
hoodie.datasource.write.keygenerator.class=org.apache.hudi.keygen.CustomKeyGenerator
hoodie.datasource.write.partitionpath.field=create_time_epoch_seconds:timestamp
hoodie.deltastreamer.keygen.timebased.timestamp.type=UNIX_TIMESTAMP
hoodie.deltastreamer.keygen.timebased.output.dateformat=yyyy/MM/dd

I get and error that looks like:

22/07/20 20:32:48 DEBUG Spark32HoodieParquetFileFormat: Appending StructType(StructField(create_time_epoch_seconds,LongType,true)) [2022/07/13]
22/07/20 20:32:48 ERROR Executor: Exception in task 0.0 in stage 5.0 (TID 5)
java.lang.ClassCastException: org.apache.spark.unsafe.types.UTF8String cannot be cast to java.lang.Long
    at scala.runtime.BoxesRunTime.unboxToLong(BoxesRunTime.java:107)
    at org.apache.spark.sql.catalyst.expressions.BaseGenericInternalRow.getLong(rows.scala:42)
    at org.apache.spark.sql.catalyst.expressions.BaseGenericInternalRow.getLong$(rows.scala:42)
    at org.apache.spark.sql.catalyst.expressions.GenericInternalRow.getLong(rows.scala:195)
    at org.apache.spark.sql.execution.vectorized.ColumnVectorUtils.populate(ColumnVectorUtils.java:66)
    at org.apache.spark.sql.execution.datasources.parquet.VectorizedParquetRecordReader.initBatch(VectorizedParquetRecordReader.java:245) 

Apparently the issue is in _partitionSchemaFromProperties function in file: hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/hudi/SparkHoodieTableFileIndex.scala

that checks for a class type it uses StructType of String for.
Once it is any non Timestamp based known class it basically uses whatever type it is and then fails to retrieve the value for.

I have a proposal here which we probably need: Give a user a way to force a string type if needed and add ability to add a prefixed column that contains a processed partition value. It could be done as two separate features.

This problem is critical for me, so I have to change Hoodie source code on my end temporary to make it work.

Here is how I roughly changed the referenced function:

 

/**
 * Get the partition schema from the hoodie.properties.
 */
private lazy val _partitionSchemaFromProperties: StructType = {
  val tableConfig = metaClient.getTableConfig
  val partitionColumns = tableConfig.getPartitionFields

  if (partitionColumns.isPresent) {
    val partitionFields = partitionColumns.get().map(column => StructField("_hoodie_"+column, StringType))
    StructType(partitionFields)
  } else {
    // If the partition columns have not stored in hoodie.properties(the table that was
    // created earlier), we trait it as a non-partitioned table.
    logWarning("No partition columns available from hoodie.properties." +
      " Partition pruning will not work")
    new StructType()
  }
} 

Issue Links

• is fixed by: HUDI-7996 Store partition type with partition fields in table configs
• is fixed by: HUDI-8036 Handle partition schema for custom key gen in SparkHoodieTableFileIndex
• relates to: HUDI-3204 Allow original partition column value to be retrieved when using TimestampBasedKeyGen