[SPARK-4987] [SQL] parquet timestamp type support

Author: Daoyuan Wang <[email protected]>

Closes #3820 from adrian-wang/parquettimestamp and squashes the following commits:

b1e2a0d [Daoyuan Wang] fix for nanos
4dadef1 [Daoyuan Wang] fix wrong read
93f438d [Daoyuan Wang] parquet timestamp support

Author: adrian-wang

docs/sql-programming-guide.md

@@ -580,6 +580,15 @@ Configuration of Parquet can be done using the `setConf` method on SQLContext or
     flag tells Spark SQL to interpret binary data as a string to provide compatibility with these systems.
   </td>
 </tr>
+<tr>
+  <td><code>spark.sql.parquet.int96AsTimestamp</code></td>
+  <td>true</td>
+  <td>
+    Some Parquet-producing systems, in particular Impala, store Timestamp into INT96. Spark would also
+    store Timestamp as INT96 because we need to avoid precision lost of the nanoseconds field. This
+    flag tells Spark SQL to interpret INT96 data as a timestamp to provide compatibility with these systems.
+  </td>
+</tr>
 <tr>
   <td><code>spark.sql.parquet.cacheMetadata</code></td>
   <td>true</td>

pom.xml

@@ -149,6 +149,7 @@
     <scala.binary.version>2.10</scala.binary.version>
     <jline.version>${scala.version}</jline.version>
     <jline.groupid>org.scala-lang</jline.groupid>
+    <jodd.version>3.6.3</jodd.version>
     <codehaus.jackson.version>1.8.8</codehaus.jackson.version>
     <snappy.version>1.1.1.6</snappy.version>
 

sql/core/pom.xml

@@ -66,6 +66,11 @@
       <artifactId>jackson-databind</artifactId>
       <version>2.3.0</version>
     </dependency>
+    <dependency>
+      <groupId>org.jodd</groupId>
+      <artifactId>jodd-core</artifactId>
+      <version>${jodd.version}</version>
+    </dependency>
     <dependency>
       <groupId>junit</groupId>
       <artifactId>junit</artifactId>

sql/core/src/main/scala/org/apache/spark/sql/SQLConf.scala

@@ -33,6 +33,7 @@ private[spark] object SQLConf {
   val DIALECT = "spark.sql.dialect"
 
   val PARQUET_BINARY_AS_STRING = "spark.sql.parquet.binaryAsString"
+  val PARQUET_INT96_AS_TIMESTAMP = "spark.sql.parquet.int96AsTimestamp"
   val PARQUET_CACHE_METADATA = "spark.sql.parquet.cacheMetadata"
   val PARQUET_COMPRESSION = "spark.sql.parquet.compression.codec"
   val PARQUET_FILTER_PUSHDOWN_ENABLED = "spark.sql.parquet.filterPushdown"
@@ -143,6 +144,12 @@ private[sql] class SQLConf extends Serializable {
   private[spark] def isParquetBinaryAsString: Boolean =
     getConf(PARQUET_BINARY_AS_STRING, "false").toBoolean
 
+  /**
+   * When set to true, we always treat INT96Values in Parquet files as timestamp.
+   */
+  private[spark] def isParquetINT96AsTimestamp: Boolean =
+    getConf(PARQUET_INT96_AS_TIMESTAMP, "true").toBoolean
+
   /**
    * When set to true, partition pruning for in-memory columnar tables is enabled.
    */

sql/core/src/main/scala/org/apache/spark/sql/parquet/ParquetConverter.scala

@@ -17,15 +17,20 @@
 
 package org.apache.spark.sql.parquet
 
+import java.sql.Timestamp
+import java.util.{TimeZone, Calendar}
+
 import scala.collection.mutable.{Buffer, ArrayBuffer, HashMap}
 
+import jodd.datetime.JDateTime
 import parquet.column.Dictionary
 import parquet.io.api.{PrimitiveConverter, GroupConverter, Binary, Converter}
 import parquet.schema.MessageType
 
 import org.apache.spark.sql.catalyst.expressions._
 import org.apache.spark.sql.parquet.CatalystConverter.FieldType
 import org.apache.spark.sql.types._
+import org.apache.spark.sql.parquet.timestamp.NanoTime
 
 /**
  * Collection of converters of Parquet types (group and primitive types) that
@@ -123,6 +128,12 @@ private[sql] object CatalystConverter {
             parent.updateDecimal(fieldIndex, value, d)
         }
       }
+      case TimestampType => {
+        new CatalystPrimitiveConverter(parent, fieldIndex) {
+          override def addBinary(value: Binary): Unit =
+            parent.updateTimestamp(fieldIndex, value)
+        }
+      }
       // All other primitive types use the default converter
       case ctype: PrimitiveType => { // note: need the type tag here!
         new CatalystPrimitiveConverter(parent, fieldIndex)
@@ -197,9 +208,11 @@ private[parquet] abstract class CatalystConverter extends GroupConverter {
   protected[parquet] def updateString(fieldIndex: Int, value: String): Unit =
     updateField(fieldIndex, value)
 
-  protected[parquet] def updateDecimal(fieldIndex: Int, value: Binary, ctype: DecimalType): Unit = {
+  protected[parquet] def updateTimestamp(fieldIndex: Int, value: Binary): Unit =
+    updateField(fieldIndex, readTimestamp(value))
+
+  protected[parquet] def updateDecimal(fieldIndex: Int, value: Binary, ctype: DecimalType): Unit =
     updateField(fieldIndex, readDecimal(new Decimal(), value, ctype))
-  }
 
   protected[parquet] def isRootConverter: Boolean = parent == null
 
@@ -232,6 +245,13 @@ private[parquet] abstract class CatalystConverter extends GroupConverter {
     unscaled = (unscaled << (64 - numBits)) >> (64 - numBits)
     dest.set(unscaled, precision, scale)
   }
+
+  /**
+   * Read a Timestamp value from a Parquet Int96Value
+   */
+  protected[parquet] def readTimestamp(value: Binary): Timestamp = {
+    CatalystTimestampConverter.convertToTimestamp(value)
+  }
 }
 
 /**
@@ -384,6 +404,9 @@ private[parquet] class CatalystPrimitiveRowConverter(
   override protected[parquet] def updateString(fieldIndex: Int, value: String): Unit =
     current.setString(fieldIndex, value)
 
+  override protected[parquet] def updateTimestamp(fieldIndex: Int, value: Binary): Unit =
+    current.update(fieldIndex, readTimestamp(value))
+
   override protected[parquet] def updateDecimal(
       fieldIndex: Int, value: Binary, ctype: DecimalType): Unit = {
     var decimal = current(fieldIndex).asInstanceOf[Decimal]
@@ -454,6 +477,73 @@ private[parquet] object CatalystArrayConverter {
   val INITIAL_ARRAY_SIZE = 20
 }
 
+private[parquet] object CatalystTimestampConverter {
+  // TODO most part of this comes from Hive-0.14
+  // Hive code might have some issues, so we need to keep an eye on it.
+  // Also we use NanoTime and Int96Values from parquet-examples.
+  // We utilize jodd to convert between NanoTime and Timestamp
+  val parquetTsCalendar = new ThreadLocal[Calendar]
+  def getCalendar = {
+    // this is a cache for the calendar instance.
+    if (parquetTsCalendar.get == null) {
+      parquetTsCalendar.set(Calendar.getInstance(TimeZone.getTimeZone("GMT")))
+    }
+    parquetTsCalendar.get
+  }
+  val NANOS_PER_SECOND: Long = 1000000000
+  val SECONDS_PER_MINUTE: Long = 60
+  val MINUTES_PER_HOUR: Long = 60
+  val NANOS_PER_MILLI: Long = 1000000
+
+  def convertToTimestamp(value: Binary): Timestamp = {
+    val nt = NanoTime.fromBinary(value)
+    val timeOfDayNanos = nt.getTimeOfDayNanos
+    val julianDay = nt.getJulianDay
+    val jDateTime = new JDateTime(julianDay.toDouble)
+    val calendar = getCalendar
+    calendar.set(Calendar.YEAR, jDateTime.getYear)
+    calendar.set(Calendar.MONTH, jDateTime.getMonth - 1)
+    calendar.set(Calendar.DAY_OF_MONTH, jDateTime.getDay)
+
+    // written in command style
+    var remainder = timeOfDayNanos
+    calendar.set(
+      Calendar.HOUR_OF_DAY,
+      (remainder / (NANOS_PER_SECOND * SECONDS_PER_MINUTE * MINUTES_PER_HOUR)).toInt)
+    remainder = remainder % (NANOS_PER_SECOND * SECONDS_PER_MINUTE * MINUTES_PER_HOUR)
+    calendar.set(
+      Calendar.MINUTE, (remainder / (NANOS_PER_SECOND * SECONDS_PER_MINUTE)).toInt)
+    remainder = remainder % (NANOS_PER_SECOND * SECONDS_PER_MINUTE)
+    calendar.set(Calendar.SECOND, (remainder / NANOS_PER_SECOND).toInt)
+    val nanos = remainder % NANOS_PER_SECOND
+    val ts = new Timestamp(calendar.getTimeInMillis)
+    ts.setNanos(nanos.toInt)
+    ts
+  }
+
+  def convertFromTimestamp(ts: Timestamp): Binary = {
+    val calendar = getCalendar
+    calendar.setTime(ts)
+    val jDateTime = new JDateTime(calendar.get(Calendar.YEAR),
+      calendar.get(Calendar.MONTH) + 1, calendar.get(Calendar.DAY_OF_MONTH))
+    // Hive-0.14 didn't set hour before get day number, while the day number should
+    // has something to do with hour, since julian day number grows at 12h GMT
+    // here we just follow what hive does.
+    val julianDay = jDateTime.getJulianDayNumber
+
+    val hour = calendar.get(Calendar.HOUR_OF_DAY)
+    val minute = calendar.get(Calendar.MINUTE)
+    val second = calendar.get(Calendar.SECOND)
+    val nanos = ts.getNanos
+    // Hive-0.14 would use hours directly, that might be wrong, since the day starts
+    // from 12h in Julian. here we just follow what hive does.
+    val nanosOfDay = nanos + second * NANOS_PER_SECOND +
+      minute * NANOS_PER_SECOND * SECONDS_PER_MINUTE +
+      hour * NANOS_PER_SECOND * SECONDS_PER_MINUTE * MINUTES_PER_HOUR
+    NanoTime(julianDay, nanosOfDay).toBinary
+  }
+}
+
 /**
  * A `parquet.io.api.GroupConverter` that converts a single-element groups that
  * match the characteristics of an array (see

sql/core/src/main/scala/org/apache/spark/sql/parquet/ParquetRelation.scala

@@ -65,8 +65,8 @@ private[sql] case class ParquetRelation(
     ParquetTypesConverter.readSchemaFromFile(
       new Path(path.split(",").head),
       conf,
-      sqlContext.conf.isParquetBinaryAsString)
-
+      sqlContext.conf.isParquetBinaryAsString,
+      sqlContext.conf.isParquetINT96AsTimestamp)
   lazy val attributeMap = AttributeMap(output.map(o => o -> o))
 
   override def newInstance() = ParquetRelation(path, conf, sqlContext).asInstanceOf[this.type]

sql/core/src/main/scala/org/apache/spark/sql/parquet/ParquetTableSupport.scala

@@ -83,7 +83,8 @@ private[parquet] class RowReadSupport extends ReadSupport[Row] with Logging {
     // TODO: Why it can be null?
     if (schema == null)  {
       log.debug("falling back to Parquet read schema")
-      schema = ParquetTypesConverter.convertToAttributes(parquetSchema, false)
+      schema = ParquetTypesConverter.convertToAttributes(
+        parquetSchema, false, true)
     }
     log.debug(s"list of attributes that will be read: $schema")
     new RowRecordMaterializer(parquetSchema, schema)
@@ -184,12 +185,12 @@ private[parquet] class RowWriteSupport extends WriteSupport[Row] with Logging {
         case t @ StructType(_) => writeStruct(
           t,
           value.asInstanceOf[CatalystConverter.StructScalaType[_]])
-        case _ => writePrimitive(schema.asInstanceOf[PrimitiveType], value)
+        case _ => writePrimitive(schema.asInstanceOf[NativeType], value)
       }
     }
   }
 
-  private[parquet] def writePrimitive(schema: PrimitiveType, value: Any): Unit = {
+  private[parquet] def writePrimitive(schema: DataType, value: Any): Unit = {
     if (value != null) {
       schema match {
         case StringType => writer.addBinary(
@@ -202,6 +203,7 @@ private[parquet] class RowWriteSupport extends WriteSupport[Row] with Logging {
         case IntegerType => writer.addInteger(value.asInstanceOf[Int])
         case ShortType => writer.addInteger(value.asInstanceOf[Short])
         case LongType => writer.addLong(value.asInstanceOf[Long])
+        case TimestampType => writeTimestamp(value.asInstanceOf[java.sql.Timestamp])
         case ByteType => writer.addInteger(value.asInstanceOf[Byte])
         case DoubleType => writer.addDouble(value.asInstanceOf[Double])
         case FloatType => writer.addFloat(value.asInstanceOf[Float])
@@ -307,6 +309,10 @@ private[parquet] class RowWriteSupport extends WriteSupport[Row] with Logging {
     writer.addBinary(Binary.fromByteArray(scratchBytes, 0, numBytes))
   }
 
+  private[parquet] def writeTimestamp(ts: java.sql.Timestamp): Unit = {
+    val binaryNanoTime = CatalystTimestampConverter.convertFromTimestamp(ts)
+    writer.addBinary(binaryNanoTime)
+  }
 }
 
 // Optimized for non-nested rows
@@ -351,6 +357,7 @@ private[parquet] class MutableRowWriteSupport extends RowWriteSupport {
       case DoubleType => writer.addDouble(record.getDouble(index))
       case FloatType => writer.addFloat(record.getFloat(index))
       case BooleanType => writer.addBoolean(record.getBoolean(index))
+      case TimestampType => writeTimestamp(record(index).asInstanceOf[java.sql.Timestamp])
       case d: DecimalType =>
         if (d.precisionInfo == None || d.precisionInfo.get.precision > 18) {
           sys.error(s"Unsupported datatype $d, cannot write to consumer")

sql/core/src/main/scala/org/apache/spark/sql/parquet/ParquetTestData.scala

@@ -25,7 +25,7 @@ import org.apache.hadoop.mapreduce.Job
 import org.apache.spark.sql.test.TestSQLContext
 
 import parquet.example.data.{GroupWriter, Group}
-import parquet.example.data.simple.SimpleGroup
+import parquet.example.data.simple.{NanoTime, SimpleGroup}
 import parquet.hadoop.{ParquetReader, ParquetFileReader, ParquetWriter}
 import parquet.hadoop.api.WriteSupport
 import parquet.hadoop.api.WriteSupport.WriteContext
@@ -63,6 +63,7 @@ private[sql] object ParquetTestData {
       optional int64 mylong;
       optional float myfloat;
       optional double mydouble;
+      optional int96 mytimestamp;
       }"""
 
   // field names for test assertion error messages
@@ -72,7 +73,8 @@ private[sql] object ParquetTestData {
     "mystring:String",
     "mylong:Long",
     "myfloat:Float",
-    "mydouble:Double"
+    "mydouble:Double",
+    "mytimestamp:Timestamp"
   )
 
   val subTestSchema =
@@ -98,6 +100,7 @@ private[sql] object ParquetTestData {
       optional int64 myoptlong;
       optional float myoptfloat;
       optional double myoptdouble;
+      optional int96 mytimestamp;
       }
     """
 
@@ -236,6 +239,7 @@ private[sql] object ParquetTestData {
       record.add(3, i.toLong << 33)
       record.add(4, 2.5F)
       record.add(5, 4.5D)
+      record.add(6, new NanoTime(1,2))
       writer.write(record)
     }
     writer.close()