Add insert overwrite dynamic support for iceberg.

iceberg/src/main/scala/com/nvidia/spark/rapids/iceberg/IcebergProviderImpl.scala

@@ -19,7 +19,7 @@ package com.nvidia.spark.rapids.iceberg
 import scala.reflect.ClassTag
 import scala.util.{Failure, Success, Try}
 
-import com.nvidia.spark.rapids.{AppendDataExecMeta, AtomicCreateTableAsSelectExecMeta, FileFormatChecks, GpuExec, GpuExpression, GpuRowToColumnarExec, GpuScan, IcebergFormatType, RapidsConf, ReadFileOp, ScanMeta, ScanRule, ShimReflectionUtils, StaticInvokeMeta, TargetSize, WriteFileOp}
+import com.nvidia.spark.rapids.{AppendDataExecMeta, AtomicCreateTableAsSelectExecMeta, FileFormatChecks, GpuExec, GpuExpression, GpuRowToColumnarExec, GpuScan, IcebergFormatType, OverwritePartitionsDynamicExecMeta, RapidsConf, ReadFileOp, ScanMeta, ScanRule, ShimReflectionUtils, StaticInvokeMeta, TargetSize, WriteFileOp}
 import org.apache.iceberg.spark.functions.{BucketFunction, GpuBucketExpression}
 import org.apache.iceberg.spark.source.{GpuSparkBatchQueryScan, GpuSparkWrite}
 import org.apache.iceberg.spark.supportsCatalog
@@ -28,7 +28,7 @@ import org.apache.spark.sql.catalyst.expressions.objects.StaticInvoke
 import org.apache.spark.sql.connector.read.Scan
 import org.apache.spark.sql.connector.write.Write
 import org.apache.spark.sql.execution.SparkPlan
-import org.apache.spark.sql.execution.datasources.v2.{AppendDataExec, AtomicCreateTableAsSelectExec, GpuAppendDataExec}
+import org.apache.spark.sql.execution.datasources.v2.{AppendDataExec, AtomicCreateTableAsSelectExec, GpuAppendDataExec, GpuOverwritePartitionsDynamicExec, OverwritePartitionsDynamicExec}
 import org.apache.spark.sql.execution.datasources.v2.rapids.GpuAtomicCreateTableAsSelectExec
 
 class IcebergProviderImpl extends IcebergProvider {
@@ -161,4 +161,33 @@ class IcebergProviderImpl extends IcebergProvider {
       cpuExec.refreshCache,
       GpuSparkWrite.convert(cpuExec.write))
   }
+
+  override def tagForGpu(cpuExec: OverwritePartitionsDynamicExec,
+                         meta: OverwritePartitionsDynamicExecMeta): Unit = {
+    if (!meta.conf.isIcebergEnabled) {
+      meta.willNotWorkOnGpu("Iceberg input and output has been disabled. To enable set " +
+        s"${RapidsConf.ENABLE_ICEBERG} to true")
+    }
+
+    if (!meta.conf.isIcebergWriteEnabled) {
+      meta.willNotWorkOnGpu("Iceberg output has been disabled. To enable set " +
+        s"${RapidsConf.ENABLE_ICEBERG_WRITE} to true")
+    }
+
+    FileFormatChecks.tag(meta, cpuExec.query.schema, IcebergFormatType, WriteFileOp)
+
+    GpuSparkWrite.tagForGpu(cpuExec.write, meta)
+  }
+
+  override def convertToGpu(cpuExec: OverwritePartitionsDynamicExec,
+                            meta: OverwritePartitionsDynamicExecMeta): GpuExec = {
+    var child: SparkPlan = meta.childPlans.head.convertIfNeeded()
+    if (!child.supportsColumnar) {
+      child = GpuRowToColumnarExec(child, TargetSize(meta.conf.gpuTargetBatchSizeBytes))
+    }
+    GpuOverwritePartitionsDynamicExec(
+      child,
+      cpuExec.refreshCache,
+      GpuSparkWrite.convert(cpuExec.write))
+  }
 }

iceberg/src/main/scala/org/apache/iceberg/spark/source/GpuSparkWrite.scala

@@ -53,7 +53,24 @@ class GpuSparkWrite(cpu: SparkWrite) extends GpuWrite with RequiresDistributionA
   private[source] val format: FileFormat = FieldUtils.readField(cpu, "format", true)
     .asInstanceOf[FileFormat]
 
-  override def toBatch: BatchWrite = new GpuBatchAppend(this)
+  override def toBatch: BatchWrite = {
+    // Call the CPU version's toBatch to get the appropriate BatchWrite implementation
+    // Iceberg's SparkWrite returns different implementations based on write mode:
+    // - BatchAppend for append operations
+    // - DynamicOverwrite for dynamic partition overwrite
+    // Since these are private classes, we check the class name to determine which GPU version
+    // to use
+    val cpuBatch = cpu.toBatch
+    val cpuBatchClassName = cpuBatch.getClass.getSimpleName
+
+    cpuBatchClassName match {
+      case "BatchAppend" => new GpuBatchAppend(this)
+      case "DynamicOverwrite" => new GpuDynamicOverwrite(this, cpuBatch)
+      case _ => 
+        throw new UnsupportedOperationException(
+          s"Unsupported Iceberg batch write type: $cpuBatchClassName")
+    }
+  }
 
   override def toStreaming: StreamingWrite = throw new UnsupportedOperationException(
     "GpuSparkWrite does not support streaming write")

iceberg/src/main/scala/org/apache/iceberg/spark/source/write.scala

@@ -39,4 +39,15 @@ class GpuBatchAppend(write: GpuSparkWrite) extends GpuBaseBatchWrite(write) {
 
     write.commitOperation(append, s"append with $numFiles new data files")
   }
+}
+
+class GpuDynamicOverwrite(
+    write: GpuSparkWrite,
+    cpuBatchWrite: BatchWrite) extends GpuBaseBatchWrite(write) {
+
+  override def commit(messages: Array[WriterCommitMessage]): Unit = {
+    // Delegate to the CPU version's commit method
+    // The CPU version handles all the complex logic for dynamic partition overwrite
+    cpuBatchWrite.commit(messages)
+  }
 }

integration_tests/src/main/python/iceberg/iceberg_insert_overwrite_dynamic_test.py

@@ -0,0 +1,301 @@
+# Copyright (c) 2025, NVIDIA CORPORATION.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Callable, Any
+
+import pytest
+
+from asserts import assert_equal_with_local_sort, assert_gpu_fallback_collect
+from data_gen import gen_df, copy_and_update
+from iceberg import create_iceberg_table, iceberg_base_table_cols, iceberg_gens_list, \
+    get_full_table_name, iceberg_full_gens_list, iceberg_write_enabled_conf
+from marks import iceberg, ignore_order, allow_non_gpu
+from spark_session import with_gpu_session, with_cpu_session, is_spark_35x
+
+pytestmark = pytest.mark.skipif(not is_spark_35x(),
+                       reason="Current spark-rapids only support spark 3.5.x")
+
+
+# Seed for initial data to ensure different data from overwrite operations
+INITIAL_DATA_SEED = 42
+
+# Configuration for dynamic partition overwrite mode
+dynamic_overwrite_conf = copy_and_update(iceberg_write_enabled_conf, {
+    "spark.sql.sources.partitionOverwriteMode": "dynamic"
+})
+
+
+def do_test_insert_overwrite_dynamic(spark_tmp_table_factory,
+                                      create_table_func: Callable[[str], Any]):
+    """Test INSERT OVERWRITE with dynamic partition overwrite mode."""
+    base_table_name = get_full_table_name(spark_tmp_table_factory)
+    cpu_table_name = f"{base_table_name}_cpu"
+    gpu_table_name = f"{base_table_name}_gpu"
+
+    create_table_func(cpu_table_name)
+    create_table_func(gpu_table_name)
+
+    def initial_insert(spark, table_name: str):
+        """Insert initial data to have something to overwrite."""
+        # Use a specific seed for initial data to ensure different data from overwrite
+        df = gen_df(spark, list(zip(iceberg_base_table_cols, iceberg_gens_list)), seed=INITIAL_DATA_SEED)
+        view_name = spark_tmp_table_factory.get()
+        df.createOrReplaceTempView(view_name)
+        spark.sql(f"INSERT INTO {table_name} SELECT * FROM {view_name}")
+
+    # Insert initial data on both CPU and GPU
+    with_gpu_session(lambda spark: initial_insert(spark, gpu_table_name),
+                     conf=iceberg_write_enabled_conf)
+    with_cpu_session(lambda spark: initial_insert(spark, cpu_table_name),
+                     conf=iceberg_write_enabled_conf)
+
+    def overwrite_data(spark, table_name: str):
+        """Perform INSERT OVERWRITE with dynamic mode - only overwrites touched partitions."""
+        # Use default seed for overwrite data - different from initial
+        df = gen_df(spark, list(zip(iceberg_base_table_cols, iceberg_gens_list)))
+        view_name = spark_tmp_table_factory.get()
+        df.createOrReplaceTempView(view_name)
+        spark.sql(f"INSERT OVERWRITE TABLE {table_name} SELECT * FROM {view_name}")
+
+    # Perform overwrite on both tables with dynamic mode
+    with_gpu_session(lambda spark: overwrite_data(spark, gpu_table_name),
+                     conf=dynamic_overwrite_conf)
+    with_cpu_session(lambda spark: overwrite_data(spark, cpu_table_name),
+                     conf=dynamic_overwrite_conf)
+
+    cpu_data = with_cpu_session(lambda spark: spark.table(cpu_table_name).collect())
+    gpu_data = with_cpu_session(lambda spark: spark.table(gpu_table_name).collect())
+    assert_equal_with_local_sort(cpu_data, gpu_data)
+
+
+@iceberg
+@ignore_order(local=True)
+@pytest.mark.parametrize("format_version", ["1", "2"], ids=lambda x: f"format_version={x}")
+@pytest.mark.parametrize("write_distribution_mode", ["none", "hash", "range"],
+                         ids=lambda x: f"write_distribution_mode={x}")
+def test_insert_overwrite_dynamic_unpartitioned_table(spark_tmp_table_factory, format_version, write_distribution_mode):
+    """Test dynamic overwrite on unpartitioned tables - should run on GPU."""
+    table_prop = {"format-version": format_version,
+                  "write.distribution-mode": write_distribution_mode}
+
+    do_test_insert_overwrite_dynamic(
+        spark_tmp_table_factory,
+        lambda table_name: create_iceberg_table(table_name, table_prop=table_prop))
+
+
+@iceberg
+@ignore_order(local=True)
+@pytest.mark.parametrize("format_version", ["1", "2"], ids=lambda x: f"format_version={x}")
+@pytest.mark.parametrize("write_distribution_mode", ["none", "hash", "range"],
+                         ids=lambda x: f"write_distribution_mode={x}")
+@pytest.mark.parametrize("fanout", [True, False], ids=lambda x: f"fanout={x}")
+def test_insert_overwrite_dynamic_bucket_partitioned(spark_tmp_table_factory, format_version, write_distribution_mode, fanout):
+    """Test dynamic overwrite with bucket partitioning - should run on GPU."""
+    table_prop = {"format-version": format_version,
+                  "write.distribution-mode": write_distribution_mode,
+                  "write.spark.fanout.enabled": str(fanout).lower()}
+
+    def create_table_with_partition(table_name: str):
+        create_iceberg_table(
+            table_name,
+            partition_col_sql="bucket(16, _c2), bucket(16, _c3)",
+            table_prop=table_prop)
+
+    do_test_insert_overwrite_dynamic(
+        spark_tmp_table_factory,
+        create_table_with_partition)
+
+
+@iceberg
+@ignore_order(local=True)
+@allow_non_gpu('OverwritePartitionsDynamicExec', 'ShuffleExchangeExec', 'SortExec', 'ProjectExec')
+@pytest.mark.parametrize("format_version", ["1", "2"], ids=lambda x: f"format_version={x}")
+@pytest.mark.parametrize("write_distribution_mode", ["none", "hash", "range"],
+                         ids=lambda x: f"write_distribution_mode={x}")
+def test_insert_overwrite_dynamic_unsupported_data_types_fallback(spark_tmp_table_factory, format_version, write_distribution_mode):
+    """Test that INSERT OVERWRITE falls back to CPU with unsupported data types."""
+    table_prop = {"format-version": format_version,
+                  "write.distribution-mode": write_distribution_mode}
+
+    def this_gen_df(spark):
+        cols = [f"_c{idx}" for idx, _ in enumerate(iceberg_full_gens_list)]
+        return gen_df(spark, list(zip(cols, iceberg_full_gens_list)))
+
+    table_name = get_full_table_name(spark_tmp_table_factory)
+
+    # Create table
+    create_iceberg_table(table_name, table_prop=table_prop, df_gen=this_gen_df)
+
+    # Initial insert (CPU only to set up the table)
+    def initial_insert(spark, table_name: str):
+        # Use a specific seed for initial data to ensure different data from overwrite
+        cols = [f"_c{idx}" for idx, _ in enumerate(iceberg_full_gens_list)]
+        df = gen_df(spark, list(zip(cols, iceberg_full_gens_list)), seed=INITIAL_DATA_SEED)
+        view_name = spark_tmp_table_factory.get()
+        df.createOrReplaceTempView(view_name)
+        spark.sql(f"INSERT INTO {table_name} SELECT * FROM {view_name}")
+
+    with_cpu_session(lambda spark: initial_insert(spark, table_name),
+                     conf=iceberg_write_enabled_conf)
+
+    # Overwrite with dynamic mode (assert GPU falls back to CPU)
+    def overwrite_data(spark, table_name: str):
+        # Use default seed for overwrite data - different from initial
+        df = this_gen_df(spark)
+        view_name = spark_tmp_table_factory.get()
+        df.createOrReplaceTempView(view_name)
+        return spark.sql(f"INSERT OVERWRITE TABLE {table_name} SELECT * FROM {view_name}")
+
+    assert_gpu_fallback_collect(lambda spark: overwrite_data(spark, table_name),
+                                "OverwritePartitionsDynamicExec",
+                                conf=dynamic_overwrite_conf)
+
+
+@iceberg
+@ignore_order(local=True)
+@allow_non_gpu('OverwritePartitionsDynamicExec', 'ShuffleExchangeExec', 'SortExec', 'ProjectExec')
+@pytest.mark.parametrize("format_version", ["1", "2"], ids=lambda x: f"format_version={x}")
+@pytest.mark.parametrize("write_distribution_mode", ["none", "hash", "range"],
+                         ids=lambda x: f"write_distribution_mode={x}")
+@pytest.mark.parametrize("partition_col_sql", [
+    pytest.param("_c2", id="identity"),
+    pytest.param("truncate(5, _c6)", id="truncate"),
+    pytest.param("year(_c9)", id="year"),
+    pytest.param("month(_c9)", id="month"),
+    pytest.param("day(_c9)", id="day"),
+    pytest.param("hour(_c9)", id="hour"),
+    pytest.param("bucket(8, _c6)", id="bucket_unsupported_type"),
+])
+def test_insert_overwrite_dynamic_unsupported_partition_fallback(
+        spark_tmp_table_factory, format_version, write_distribution_mode, partition_col_sql):
+    """Test that INSERT OVERWRITE falls back to CPU with unsupported partition functions."""
+    table_prop = {"format-version": format_version,
+                  "write.distribution-mode": write_distribution_mode}
+
+    table_name = get_full_table_name(spark_tmp_table_factory)
+
+    # Create table
+    create_iceberg_table(table_name,
+                         partition_col_sql=partition_col_sql,
+                         table_prop=table_prop)
+
+    # Initial insert (CPU only to set up the table)
+    def initial_insert(spark, table_name: str):
+        # Use a specific seed for initial data to ensure different data from overwrite
+        df = gen_df(spark, list(zip(iceberg_base_table_cols, iceberg_gens_list)), seed=INITIAL_DATA_SEED)
+        view_name = spark_tmp_table_factory.get()
+        df.createOrReplaceTempView(view_name)
+        spark.sql(f"INSERT INTO {table_name} SELECT * FROM {view_name}")
+
+    with_cpu_session(lambda spark: initial_insert(spark, table_name),
+                     conf=iceberg_write_enabled_conf)
+
+    # Overwrite with dynamic mode (assert GPU falls back to CPU)
+    def overwrite_data(spark, table_name: str):
+        # Use default seed for overwrite data - different from initial
+        df = gen_df(spark, list(zip(iceberg_base_table_cols, iceberg_gens_list)))
+        view_name = spark_tmp_table_factory.get()
+        df.createOrReplaceTempView(view_name)
+        return spark.sql(f"INSERT OVERWRITE TABLE {table_name} SELECT * FROM {view_name}")
+
+    assert_gpu_fallback_collect(lambda spark: overwrite_data(spark, table_name),
+                                "OverwritePartitionsDynamicExec",
+                                conf=dynamic_overwrite_conf)
+
+
+@iceberg
+@ignore_order(local=True)
+@allow_non_gpu('OverwritePartitionsDynamicExec', 'ShuffleExchangeExec', 'SortExec', 'ProjectExec')
+@pytest.mark.parametrize("format_version", ["1", "2"], ids=lambda x: f"format_version={x}")
+@pytest.mark.parametrize("file_format", ["orc", "avro"], ids=lambda x: f"file_format={x}")
+@pytest.mark.parametrize("write_distribution_mode", ["none", "hash", "range"],
+                         ids=lambda x: f"write_distribution_mode={x}")
+def test_insert_overwrite_dynamic_unsupported_file_format_fallback(
+        spark_tmp_table_factory, format_version, file_format, write_distribution_mode):
+    """Test that INSERT OVERWRITE falls back to CPU with unsupported file formats."""
+    table_prop = {"format-version": format_version,
+                  "write.distribution-mode": write_distribution_mode,
+                  "write.format.default": file_format}
+
+    table_name = get_full_table_name(spark_tmp_table_factory)
+
+    # Create table
+    create_iceberg_table(table_name, table_prop=table_prop)
+
+    # Initial insert (CPU only to set up the table)
+    def initial_insert(spark, table_name: str):
+        # Use a specific seed for initial data to ensure different data from overwrite
+        df = gen_df(spark, list(zip(iceberg_base_table_cols, iceberg_gens_list)), seed=INITIAL_DATA_SEED)
+        view_name = spark_tmp_table_factory.get()
+        df.createOrReplaceTempView(view_name)
+        spark.sql(f"INSERT INTO {table_name} SELECT * FROM {view_name}")
+
+    with_cpu_session(lambda spark: initial_insert(spark, table_name),
+                     conf=iceberg_write_enabled_conf)
+
+    # Overwrite with dynamic mode (assert GPU falls back to CPU)
+    def overwrite_data(spark, table_name: str):
+        # Use default seed for overwrite data - different from initial
+        df = gen_df(spark, list(zip(iceberg_base_table_cols, iceberg_gens_list)))
+        view_name = spark_tmp_table_factory.get()
+        df.createOrReplaceTempView(view_name)
+        return spark.sql(f"INSERT OVERWRITE TABLE {table_name} SELECT * FROM {view_name}")
+
+    assert_gpu_fallback_collect(lambda spark: overwrite_data(spark, table_name),
+                                "OverwritePartitionsDynamicExec",
+                                conf=dynamic_overwrite_conf)
+
+
+@iceberg
+@ignore_order(local=True)
+@allow_non_gpu('OverwritePartitionsDynamicExec', 'ShuffleExchangeExec', 'SortExec', 'ProjectExec')
+@pytest.mark.parametrize("format_version", ["1", "2"], ids=lambda x: f"format_version={x}")
+@pytest.mark.parametrize("write_distribution_mode", ["none", "hash", "range"],
+                         ids=lambda x: f"write_distribution_mode={x}")
+@pytest.mark.parametrize("conf_key", ["spark.rapids.sql.format.iceberg.enabled",
+                                      "spark.rapids.sql.format.iceberg.write.enabled"],
+                         ids=lambda x: f"{x}=False")
+def test_insert_overwrite_dynamic_fallback_when_conf_disabled(
+        spark_tmp_table_factory, format_version, write_distribution_mode, conf_key):
+    """Test that INSERT OVERWRITE falls back to CPU when Iceberg write is disabled."""
+    table_prop = {"format-version": format_version,
+                  "write.distribution-mode": write_distribution_mode}
+
+    table_name = get_full_table_name(spark_tmp_table_factory)
+
+    # Create table
+    create_iceberg_table(table_name, table_prop=table_prop)
+
+    # Initial insert (CPU only to set up the table)
+    def initial_insert(spark, table_name: str):
+        # Use a specific seed for initial data to ensure different data from overwrite
+        df = gen_df(spark, list(zip(iceberg_base_table_cols, iceberg_gens_list)), seed=INITIAL_DATA_SEED)
+        view_name = spark_tmp_table_factory.get()
+        df.createOrReplaceTempView(view_name)
+        spark.sql(f"INSERT INTO {table_name} SELECT * FROM {view_name}")
+
+    with_cpu_session(lambda spark: initial_insert(spark, table_name),
+                     conf=iceberg_write_enabled_conf)
+
+    # Overwrite with dynamic mode (assert GPU falls back to CPU)
+    def overwrite_data(spark, table_name: str):
+        # Use default seed for overwrite data - different from initial
+        df = gen_df(spark, list(zip(iceberg_base_table_cols, iceberg_gens_list)))
+        view_name = spark_tmp_table_factory.get()
+        df.createOrReplaceTempView(view_name)
+        return spark.sql(f"INSERT OVERWRITE TABLE {table_name} SELECT * FROM {view_name}")
+
+    updated_conf = copy_and_update(dynamic_overwrite_conf, {conf_key: "false"})
+    assert_gpu_fallback_collect(lambda spark: overwrite_data(spark, table_name),
+                                "OverwritePartitionsDynamicExec",
+                                conf=updated_conf)