Support unix_timestamp on GPU for subset of formats (NVIDIA#1113)

* Support unix_timestamp on GPU for subset of formats Signed-off-by: Andy Grove <[email protected]> * close scalar value Signed-off-by: Andy Grove <[email protected]> * compatible formats will now run on GPU without requiring incompatibleOps to be set Signed-off-by: Andy Grove <[email protected]> * code cleanup and address more review comments Signed-off-by: Andy Grove <[email protected]> * add specific config option for enabling incompatible date formats on GPU * update documentation Signed-off-by: Andy Grove <[email protected]> * improve docs Signed-off-by: Andy Grove <[email protected]> * use constants for special dates Signed-off-by: Andy Grove <[email protected]> * Add support for more date formats and remove incompat from to_unix_timestamp Signed-off-by: Andy Grove <[email protected]> * remove debug print Signed-off-by: Andy Grove <[email protected]> * Revert unnecessary change Signed-off-by: Andy Grove <[email protected]> * Make ToUnixTimestamp consistent with UnixTimestamp Signed-off-by: Andy Grove <[email protected]> * refactor to remove duplicate code Signed-off-by: Andy Grove <[email protected]> * fix resource leaks and fix regressions in python tests Signed-off-by: Andy Grove <[email protected]> * scalstyle Signed-off-by: Andy Grove <[email protected]> * update docs Signed-off-by: Andy Grove <[email protected]> * fix error in handling of legacyTimeParserPolicy=EXCEPTION Signed-off-by: Andy Grove <[email protected]> * fix test failures against Spark 3.1.0 Signed-off-by: Andy Grove <[email protected]>
sperlingxx · Nov 20, 2020 · b58dc2a · b58dc2a
1 parent e6a7cff
commit b58dc2a
Show file tree

Hide file tree

Showing 10 changed files with 433 additions and 60 deletions.
diff --git a/docs/compatibility.md b/docs/compatibility.md
@@ -212,6 +212,29 @@ window functions like `row_number`, `lead`, and `lag` can produce different resu
 includes both `-0.0` and `0.0`, or if the ordering is ambiguous. Spark can produce
 different results from one run to another if the ordering is ambiguous on a window function too.
 
+## Parsing strings as dates or timestamps
+
+When converting strings to dates or timestamps using functions like `to_date` and `unix_timestamp`,
+only a subset of possible formats are supported on GPU with full compatibility with Spark. The
+supported formats are:
+
+- `dd/MM/yyyy`
+- `yyyy/MM`
+- `yyyy/MM/dd`
+- `yyyy-MM`
+- `yyyy-MM-dd`
+- `yyyy-MM-dd HH:mm:ss`
+
+Other formats may result in incorrect results and will not run on the GPU by default. Some
+specific issues with other formats are:
+
+- Spark supports partial microseconds but the plugin does not
+- The plugin will produce incorrect results for input data that is not in the correct format in
+some cases
+
+To enable all formats on GPU, set
+[`spark.rapids.sql.incompatibleDateFormats.enabled`](configs.md#sql.incompatibleDateFormats.enabled) to `true`.
+
 ## Casting between types
 
 In general, performing `cast` and `ansi_cast` operations on the GPU is compatible with the same operations on the CPU. However, there are some exceptions. For this reason, certain casts are disabled on the GPU by default and require configuration options to be specified to enable them. 

diff --git a/docs/configs.md b/docs/configs.md
@@ -71,6 +71,7 @@ Name | Description | Default Value
 <a name="sql.hashOptimizeSort.enabled"></a>spark.rapids.sql.hashOptimizeSort.enabled|Whether sorts should be inserted after some hashed operations to improve output ordering. This can improve output file sizes when saving to columnar formats.|false
 <a name="sql.improvedFloatOps.enabled"></a>spark.rapids.sql.improvedFloatOps.enabled|For some floating point operations spark uses one way to compute the value and the underlying cudf implementation can use an improved algorithm. In some cases this can result in cudf producing an answer when spark overflows. Because this is not as compatible with spark, we have it disabled by default.|false
 <a name="sql.improvedTimeOps.enabled"></a>spark.rapids.sql.improvedTimeOps.enabled|When set to true, some operators will avoid overflowing by converting epoch days  directly to seconds without first converting to microseconds|false
+<a name="sql.incompatibleDateFormats.enabled"></a>spark.rapids.sql.incompatibleDateFormats.enabled|When parsing strings as dates and timestamps in functions like unix_timestamp, setting this to true will force all parsing onto GPU even for formats that can result in incorrect results when parsing invalid inputs.|false
 <a name="sql.incompatibleOps.enabled"></a>spark.rapids.sql.incompatibleOps.enabled|For operations that work, but are not 100% compatible with the Spark equivalent set if they should be enabled by default or disabled by default.|false
 <a name="sql.python.gpu.enabled"></a>spark.rapids.sql.python.gpu.enabled|This is an experimental feature and is likely to change in the future. Enable (true) or disable (false) support for scheduling Python Pandas UDFs with GPU resources. When enabled, pandas UDFs are assumed to share the same GPU that the RAPIDs accelerator uses and will honor the python GPU configs|false
 <a name="sql.reader.batchSizeBytes"></a>spark.rapids.sql.reader.batchSizeBytes|Soft limit on the maximum number of bytes the reader reads per batch. The readers will read chunks of data until this limit is met or exceeded. Note that the reader may estimate the number of bytes that will be used on the GPU in some cases based on the schema and number of rows in each batch.|2147483647
@@ -219,12 +220,12 @@ Name | SQL Function(s) | Description | Default Value | Notes
 <a name="sql.expression.TimeSub"></a>spark.rapids.sql.expression.TimeSub| |Subtracts interval from timestamp|true|None|
 <a name="sql.expression.ToDegrees"></a>spark.rapids.sql.expression.ToDegrees|`degrees`|Converts radians to degrees|true|None|
 <a name="sql.expression.ToRadians"></a>spark.rapids.sql.expression.ToRadians|`radians`|Converts degrees to radians|true|None|
-<a name="sql.expression.ToUnixTimestamp"></a>spark.rapids.sql.expression.ToUnixTimestamp|`to_unix_timestamp`|Returns the UNIX timestamp of the given time|false|This is not 100% compatible with the Spark version because Incorrectly formatted strings and bogus dates produce garbage data instead of null|
+<a name="sql.expression.ToUnixTimestamp"></a>spark.rapids.sql.expression.ToUnixTimestamp|`to_unix_timestamp`|Returns the UNIX timestamp of the given time|true|None|
 <a name="sql.expression.UnaryMinus"></a>spark.rapids.sql.expression.UnaryMinus|`negative`|Negate a numeric value|true|None|
 <a name="sql.expression.UnaryPositive"></a>spark.rapids.sql.expression.UnaryPositive|`positive`|A numeric value with a + in front of it|true|None|
 <a name="sql.expression.UnboundedFollowing$"></a>spark.rapids.sql.expression.UnboundedFollowing$| |Special boundary for a window frame, indicating all rows preceding the current row|true|None|
 <a name="sql.expression.UnboundedPreceding$"></a>spark.rapids.sql.expression.UnboundedPreceding$| |Special boundary for a window frame, indicating all rows preceding the current row|true|None|
-<a name="sql.expression.UnixTimestamp"></a>spark.rapids.sql.expression.UnixTimestamp|`unix_timestamp`|Returns the UNIX timestamp of current or specified time|false|This is not 100% compatible with the Spark version because Incorrectly formatted strings and bogus dates produce garbage data instead of null|
+<a name="sql.expression.UnixTimestamp"></a>spark.rapids.sql.expression.UnixTimestamp|`unix_timestamp`|Returns the UNIX timestamp of current or specified time|true|None|
 <a name="sql.expression.Upper"></a>spark.rapids.sql.expression.Upper|`upper`, `ucase`|String uppercase operator|false|This is not 100% compatible with the Spark version because in some cases unicode characters change byte width when changing the case. The GPU string conversion does not support these characters. For a full list of unsupported characters see https://github.com/rapidsai/cudf/issues/3132|
 <a name="sql.expression.WeekDay"></a>spark.rapids.sql.expression.WeekDay|`weekday`|Returns the day of the week (0 = Monday...6=Sunday)|true|None|
 <a name="sql.expression.WindowExpression"></a>spark.rapids.sql.expression.WindowExpression| |Calculates a return value for every input row of a table based on a group (or "window") of rows|true|None|

diff --git a/integration_tests/src/main/python/date_time_test.py b/integration_tests/src/main/python/date_time_test.py
@@ -160,26 +160,23 @@ def test_dayofyear(data_gen):
     assert_gpu_and_cpu_are_equal_collect(
             lambda spark : unary_op_df(spark, data_gen).select(f.dayofyear(f.col('a'))))
 
-@incompat #Really only the string is
 @pytest.mark.parametrize('data_gen', date_n_time_gens, ids=idfn)
 def test_unix_timestamp(data_gen):
     assert_gpu_and_cpu_are_equal_collect(
             lambda spark : unary_op_df(spark, data_gen).select(f.unix_timestamp(f.col('a'))))
 
-@incompat #Really only the string is
 @pytest.mark.parametrize('data_gen', date_n_time_gens, ids=idfn)
 def test_to_unix_timestamp(data_gen):
     assert_gpu_and_cpu_are_equal_collect(
             lambda spark : unary_op_df(spark, data_gen).selectExpr("to_unix_timestamp(a)"))
 
-@incompat #Really only the string is
 @pytest.mark.parametrize('data_gen', date_n_time_gens, ids=idfn)
 def test_unix_timestamp_improved(data_gen):
-    conf = {"spark.rapids.sql.improvedTimeOps.enabled": "true"}
+    conf = {"spark.rapids.sql.improvedTimeOps.enabled": "true",
+            "spark.sql.legacy.timeParserPolicy": "CORRECTED"}
     assert_gpu_and_cpu_are_equal_collect(
             lambda spark : unary_op_df(spark, data_gen).select(f.unix_timestamp(f.col('a'))), conf)
 
-@incompat #Really only the string is
 @pytest.mark.parametrize('data_gen', date_n_time_gens, ids=idfn)
 def test_to_unix_timestamp_improved(data_gen):
     conf = {"spark.rapids.sql.improvedTimeOps.enabled": "true"}
@@ -190,14 +187,11 @@ def test_to_unix_timestamp_improved(data_gen):
         (StringGen('[0-9]{4}/[01][12]/[0-2][1-8]'),'yyyy/MM/dd'),
         (ConvertGen(DateGen(nullable=False), lambda d: d.strftime('%Y/%m').zfill(7), data_type=StringType()), 'yyyy/MM')]
 
-@incompat
 @pytest.mark.parametrize('data_gen,date_form', str_date_and_format_gen, ids=idfn)
 def test_string_to_unix_timestamp(data_gen, date_form):
-    print("date: " + date_form)
     assert_gpu_and_cpu_are_equal_collect(
         lambda spark : unary_op_df(spark, data_gen, seed=1).selectExpr("to_unix_timestamp(a, '{}')".format(date_form)))
 
-@incompat
 @pytest.mark.parametrize('data_gen,date_form', str_date_and_format_gen, ids=idfn)
 def test_string_unix_timestamp(data_gen, date_form):
     assert_gpu_and_cpu_are_equal_collect(

diff --git a/integration_tests/src/test/scala/com/nvidia/spark/rapids/TimeOperatorsSuite.scala b/integration_tests/src/test/scala/com/nvidia/spark/rapids/TimeOperatorsSuite.scala
@@ -16,6 +16,7 @@
 
 package com.nvidia.spark.rapids
 
+import org.apache.spark.SparkConf
 import org.apache.spark.sql.functions._
 
 class TimeOperatorsSuite extends SparkQueryCompareTestSuite {
@@ -28,7 +29,8 @@ class TimeOperatorsSuite extends SparkQueryCompareTestSuite {
   }
 
   testSparkResultsAreEqual(
-      "Test from_unixtime with alternative month and two digit year", datesPostEpochDf) {
+      "Test from_unixtime with alternative month and two digit year", datesPostEpochDf,
+      conf = new SparkConf().set(RapidsConf.INCOMPATIBLE_DATE_FORMATS.key, "true")) {
     frame => frame.select(from_unixtime(col("dates"),"dd/LL/yy HH:mm:ss.SSSSSS"))
   }
 

diff --git a/sql-plugin/src/main/scala/com/nvidia/spark/rapids/DateUtils.scala b/sql-plugin/src/main/scala/com/nvidia/spark/rapids/DateUtils.scala
@@ -36,6 +36,12 @@ object DateUtils {
     "MM" -> "%m", "LL" -> "%m", "dd" -> "%d", "mm" -> "%M", "ss" -> "%S", "HH" -> "%H",
     "yy" -> "%y", "yyyy" -> "%Y", "SSSSSS" -> "%f")
 
+  val ONE_SECOND_MICROSECONDS = 1000000
+
+  val ONE_DAY_SECONDS = 86400L
+
+  val ONE_DAY_MICROSECONDS = 86400000000L
+
   case class FormatKeywordToReplace(word: String, startIndex: Int, endIndex: Int)
 
   /**

diff --git a/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuCast.scala b/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuCast.scala
@@ -96,8 +96,6 @@ object GpuCast {
     "\\A\\d{4}\\-\\d{2}\\-\\d{2}[ T]\\d{2}:\\d{2}:\\d{2}\\.\\d{6}Z\\Z"
   private val TIMESTAMP_REGEX_NO_DATE = "\\A[T]?(\\d{2}:\\d{2}:\\d{2}\\.\\d{6}Z)\\Z"
 
-  private val ONE_DAY_MICROSECONDS = 86400000000L
-
   /**
    * Regex for identifying strings that contain numeric values that can be casted to integral
    * types. This includes floating point numbers but not numbers containing exponents.
@@ -122,6 +120,12 @@ object GpuCast {
 
   val INVALID_FLOAT_CAST_MSG = "At least one value is either null or is an invalid number"
 
+  val EPOCH = "epoch"
+  val NOW = "now"
+  val TODAY = "today"
+  val YESTERDAY = "yesterday"
+  val TOMORROW = "tomorrow"
+
   /**
    * Returns true iff we can cast `from` to `to` using the GPU.
    */
@@ -182,6 +186,17 @@ object GpuCast {
       case _ => false
     }
   }
+
+  def calculateSpecialDates: Map[String, Int] = {
+    val now = DateTimeUtils.currentDate(ZoneId.of("UTC"))
+    Map(
+      EPOCH -> 0,
+      NOW -> now,
+      TODAY -> now,
+      YESTERDAY -> (now - 1),
+      TOMORROW -> (now + 1)
+    )
+  }
 }
 
 /**
@@ -655,16 +670,6 @@ case class GpuCast(
       }
     }
 
-    // special dates
-    val now = DateTimeUtils.currentDate(ZoneId.of("UTC"))
-    val specialDates: Map[String, Int] = Map(
-      "epoch" -> 0,
-      "now" -> now,
-      "today" -> now,
-      "yesterday" -> (now - 1),
-      "tomorrow" -> (now + 1)
-    )
-
     var sanitizedInput = input.incRefCount()
 
     // replace partial months
@@ -677,6 +682,8 @@ case class GpuCast(
       cv.stringReplaceWithBackrefs("-([0-9])([ T](:?[\\r\\n]|.)*)?\\Z", "-0\\1")
     }
 
+    val specialDates = calculateSpecialDates
+
     withResource(sanitizedInput) { sanitizedInput =>
 
       // convert dates that are in valid formats yyyy, yyyy-mm, yyyy-mm-dd
@@ -756,11 +763,11 @@ case class GpuCast(
     val today: Long = cal.getTimeInMillis * 1000
     val todayStr = new SimpleDateFormat("yyyy-MM-dd").format(cal.getTime)
     val specialDates: Map[String, Long] = Map(
-      "epoch" -> 0,
-      "now" -> today,
-      "today" -> today,
-      "yesterday" -> (today - ONE_DAY_MICROSECONDS),
-      "tomorrow" -> (today + ONE_DAY_MICROSECONDS)
+      GpuCast.EPOCH -> 0,
+      GpuCast.NOW -> today,
+      GpuCast.TODAY -> today,
+      GpuCast.YESTERDAY -> (today - DateUtils.ONE_DAY_MICROSECONDS),
+      GpuCast.TOMORROW -> (today + DateUtils.ONE_DAY_MICROSECONDS)
     )
 
     var sanitizedInput = input.incRefCount()

diff --git a/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuOverrides.scala b/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuOverrides.scala
@@ -46,6 +46,7 @@ import org.apache.spark.sql.execution.exchange.{BroadcastExchangeExec, ShuffleEx
 import org.apache.spark.sql.execution.joins._
 import org.apache.spark.sql.execution.python._
 import org.apache.spark.sql.execution.window.WindowExec
+import org.apache.spark.sql.internal.SQLConf
 import org.apache.spark.sql.rapids._
 import org.apache.spark.sql.rapids.catalyst.expressions.GpuRand
 import org.apache.spark.sql.rapids.execution.{GpuBroadcastMeta, GpuBroadcastNestedLoopJoinMeta, GpuCustomShuffleReaderExec, GpuShuffleMeta}
@@ -1142,28 +1143,24 @@ object GpuOverrides {
         override def convertToGpu(lhs: Expression, rhs: Expression): GpuExpression = {
           if (conf.isImprovedTimestampOpsEnabled) {
             // passing the already converted strf string for a little optimization
-            GpuToUnixTimestampImproved(lhs, rhs, strfFormat)
+            GpuToUnixTimestampImproved(lhs, rhs, sparkFormat, strfFormat)
           } else {
-            GpuToUnixTimestamp(lhs, rhs, strfFormat)
+            GpuToUnixTimestamp(lhs, rhs, sparkFormat, strfFormat)
           }
         }
-      })
-      .incompat("Incorrectly formatted strings and bogus dates produce garbage data" +
-        " instead of null"),
+      }),
     expr[UnixTimestamp](
       "Returns the UNIX timestamp of current or specified time",
       (a, conf, p, r) => new UnixTimeExprMeta[UnixTimestamp](a, conf, p, r){
         override def convertToGpu(lhs: Expression, rhs: Expression): GpuExpression = {
           if (conf.isImprovedTimestampOpsEnabled) {
             // passing the already converted strf string for a little optimization
-            GpuUnixTimestampImproved(lhs, rhs, strfFormat)
+            GpuUnixTimestampImproved(lhs, rhs, sparkFormat, strfFormat)
           } else {
-            GpuUnixTimestamp(lhs, rhs, strfFormat)
+            GpuUnixTimestamp(lhs, rhs, sparkFormat, strfFormat)
           }
         }
-      })
-      .incompat("Incorrectly formatted strings and bogus dates produce garbage data" +
-        " instead of null"),
+      }),
     expr[Hour](
       "Returns the hour component of the string/timestamp",
       (a, conf, p, r) => new UnaryExprMeta[Hour](a, conf, p, r) {
@@ -2094,6 +2091,16 @@ object GpuOverrides {
   ).map(r => (r.getClassFor.asSubclass(classOf[SparkPlan]), r)).toMap
   val execs: Map[Class[_ <: SparkPlan], ExecRule[_ <: SparkPlan]] =
     commonExecs ++ ShimLoader.getSparkShims.getExecs
+
+  def getTimeParserPolicy: TimeParserPolicy = {
+    val policy = SQLConf.get.getConfString(SQLConf.LEGACY_TIME_PARSER_POLICY.key, "EXCEPTION")
+    policy match {
+      case "LEGACY" => LegacyTimeParserPolicy
+      case "EXCEPTION" => ExceptionTimeParserPolicy
+      case "CORRECTED" => CorrectedTimeParserPolicy
+    }
+  }
+
 }
 /** Tag the initial plan when AQE is enabled */
 case class GpuQueryStagePrepOverrides() extends Rule[SparkPlan] with Logging {

diff --git a/sql-plugin/src/main/scala/com/nvidia/spark/rapids/RapidsConf.scala b/sql-plugin/src/main/scala/com/nvidia/spark/rapids/RapidsConf.scala
@@ -427,6 +427,13 @@ object RapidsConf {
     .booleanConf
     .createWithDefault(false)
 
+  val INCOMPATIBLE_DATE_FORMATS = conf("spark.rapids.sql.incompatibleDateFormats.enabled")
+      .doc("When parsing strings as dates and timestamps in functions like unix_timestamp, " +
+          "setting this to true will force all parsing onto GPU even for formats that can " +
+          "result in incorrect results when parsing invalid inputs.")
+      .booleanConf
+      .createWithDefault(false)
+
   val IMPROVED_FLOAT_OPS = conf("spark.rapids.sql.improvedFloatOps.enabled")
     .doc("For some floating point operations spark uses one way to compute the value " +
       "and the underlying cudf implementation can use an improved algorithm. " +
@@ -922,6 +929,8 @@ class RapidsConf(conf: Map[String, String]) extends Logging {
 
   lazy val isIncompatEnabled: Boolean = get(INCOMPATIBLE_OPS)
 
+  lazy val incompatDateFormats: Boolean = get(INCOMPATIBLE_DATE_FORMATS)
+
   lazy val includeImprovedFloat: Boolean = get(IMPROVED_FLOAT_OPS)
 
   lazy val pinnedPoolSize: Long = get(PINNED_POOL_SIZE)