ClassCastException in Spark when retrieving arrays from ClickHouse via clickhouse-java driver

[maxim-lixakov]

trafficstars

Describe the bug

In the current implementation of Spark when working with arrays of data returned from ClickHouse through the clickhouse-java driver, there is a type conversion error. When the driver returns an array of primitive types (e.g., bytes or integers), Spark attempts to cast them to an array of objects, which is not possible in Java and Scala, as arrays of primitives are not subtypes of arrays of objects. This leads to a ClassCastException.

Steps to reproduce

CREATE TABLE statements for tables involved:

CREATE TABLE example_table (
  id UInt32,
  data Array(Int32)
) ENGINE = MergeTree()
ORDER BY id;

INSERT INTO example_table VALUES (1, [1, 2, 3]), (2, [4, 5, 6]);

At first we implement custom Clickhouse dialect that handles Array(T) because in native Spark, Array(T) is unsupported type

import scala.util.matching.Regex
import org.apache.spark.sql.jdbc.{JdbcDialect, JdbcType}
import org.apache.spark.sql.execution.datasources.jdbc.{JdbcUtils}
import org.apache.spark.sql.types._
import org.slf4j.LoggerFactory
import java.sql.Types

private object ClickhouseDialectExtension extends JdbcDialect {

  private val logger = LoggerFactory.getLogger(getClass)

  private val arrayTypePattern: Regex = "^Array\\((.*)\\)$".r

  override def canHandle(url: String): Boolean = {
    url.startsWith("jdbc:clickhouse")
  }

  override def getCatalystType(
      sqlType: Int,
      typeName: String,
      size: Int,
      md: MetadataBuilder): Option[DataType] = {
    sqlType match {
      case Types.ARRAY =>
        arrayTypePattern.findFirstMatchIn(typeName) match {
          case Some(m) =>
            val elementType = m.group(1)
            JdbcUtils.getCommonJDBCType(elementType).map(dt => ArrayType(dt))
          case None => None
        }
      case _ => None
    }
  }

  override def getJDBCType(dt: DataType): Option[JdbcType] = dt match {
    case ArrayType(et, _) =>
      logger.debug("Custom mapping applied: Array[T_1] for ArrayType(T_0)")
      getJDBCType(et)
        .orElse(JdbcUtils.getCommonJDBCType(et))
        .map(jdbcType => JdbcType(s"Array(${jdbcType.databaseTypeDefinition})", Types.ARRAY))
    case _ => None
  }
}

val spark = SparkSession
  .builder()
  .master("local[*]")
  .appName("Spark Clickhouse Dialect Test Session")
  .config("spark.jars", jarPaths)  // include the JAR file containing the compiled custom dialect
  .getOrCreate()

// Register custom Clickhouse dialect
JdbcDialects.registerDialect(ClickhouseDialectExtension)

val df = spark.read
    .format("jdbc")
    .option("url", s"jdbc:clickhouse://$jdbcHostname:${jdbcPort}/$database")
    .option("dbtable", "example_table")
    .load()

df.collect() // raises error:  java.lang.ClassCastException: [B cannot be cast to [Ljava.lang.Object;

Expected behaviour

The driver should return an array of objects instead of an array of primitives to avoid ClassCastException in Spark.

Code example

Example code snippet in Spark that demonstrates the issue:

case ArrayType(et, _) =>
  val elementConversion = et match {
    case TimestampType =>
      (array: Object) =>
        array.asInstanceOf[Array[java.sql.Timestamp]].map { timestamp =>
          nullSafeConvert(timestamp, DateTimeUtils.fromJavaTimestamp)
        }

    case StringType =>
      (array: Object) =>
        array.asInstanceOf[Array[java.lang.Object]]
          .map(obj => if (obj == null) null else UTF8String.fromString(obj.toString))

    case DateType =>
      (array: Object) =>
        array.asInstanceOf[Array[java.sql.Date]].map { date =>
          nullSafeConvert(date, DateTimeUtils.fromJavaDate)
        }

    case dt: DecimalType =>
      (array: Object) =>
        array.asInstanceOf[Array[java.math.BigDecimal]].map { decimal =>
          nullSafeConvert[java.math.BigDecimal](
            decimal, d => Decimal(d, dt.precision, dt.scale))
        }

    case LongType if metadata.contains("binarylong") =>
      throw QueryExecutionErrors.unsupportedArrayElementTypeBasedOnBinaryError(dt)

    case ArrayType(_, _) =>
      throw QueryExecutionErrors.nestedArraysUnsupportedError()

    case _ => (array: Object) => array.asInstanceOf[Array[Any]]
  }

Error log

java.lang.ClassCastException: [B cannot be cast to [Ljava.lang.Object;

Configuration

Environment

• Client version: com.clickhouse.clickhouse-jdbc.0.6.0-patch5
• Language version: Java 1.8.0, Scala 2.12
• OS: macOS 13.4.1 (22F82)

ClickHouse server

• ClickHouse Server version: docker image clickhouse-server:latest-alpine
• ClickHouse Server non-default settings, if any: None

[chernser]

@maxim-lixakov, thank you for reporting the issue. It will look into it. Btw, ClickHouse has own Spark connector. Please see https://clickhouse.com/docs/en/integrations/apache-spark

Thanks!

[dolfinus]

This issue prevents our custom Clickhouse dialect for Spark from reading Arrays from Clickhouse: https://github.com/MobileTeleSystems/spark-dialect-extension/blob/b4ba3724bf036394457985588ea891c05c90e5b1/src/test/scala/io/github/mtsongithub/doetl/sparkdialectextensions/clickhouse/ClickhouseDialectTest.scala#L478-L500

[dolfinus]

Doesn' reproduce on 0.9.2

[chernser]

Good day, @dolfinus ! Thank you for the update!

I will verify it anyway then on the HEAD.

[chernser]

@dolfinus @maxim-lixakov

The driver should return an array of objects instead of an array of primitives to avoid ClassCastException in Spark. What method do you use to get array value?

java.sql.ResultSet#getObject(int) - returns java.sql.Array java.sql.ResultSet#getObject(int, java.lang.Class<T>) - tried to return Class<T>

I think HEAD fixed some behavior at this may be working not as expected in 0.9.3

Thanks!