HIVE-29354: Projection and Filter Pushdown for Shredded VARIANT Columns

Author: deniskuzZ

iceberg/iceberg-handler/src/main/java/org/apache/iceberg/mr/hive/HiveIcebergFilterFactory.java

@@ -42,8 +42,10 @@
 import org.apache.iceberg.expressions.Expressions;
 import org.apache.iceberg.expressions.Literal;
 import org.apache.iceberg.expressions.UnboundTerm;
+import org.apache.iceberg.mr.hive.variant.VariantPathUtil;
 import org.apache.iceberg.relocated.com.google.common.base.Preconditions;
 import org.apache.iceberg.transforms.Transforms;
+import org.apache.iceberg.types.Type;
 import org.apache.iceberg.types.Types;
 import org.apache.iceberg.util.DateTimeUtil;
 import org.apache.iceberg.util.NaNUtil;
@@ -112,8 +114,8 @@ private static Expression translate(ExpressionTree tree, List<PredicateLeaf> lea
   private static Expression translateLeaf(PredicateLeaf leaf) {
     TransformSpec transformSpec = TransformSpec.fromStringWithColumnName(leaf.getColumnName());
     String columnName = transformSpec.getColumnName();
-    UnboundTerm<Object> column =
-        ObjectUtils.defaultIfNull(toTerm(columnName, transformSpec), Expressions.ref(columnName));
+
+    UnboundTerm<Object> column = resolveTerm(columnName, leaf, transformSpec);
 
     switch (leaf.getOperator()) {
       case EQUALS:
@@ -144,6 +146,15 @@ private static Expression translateLeaf(PredicateLeaf leaf) {
     }
   }
 
+  private static UnboundTerm<Object> resolveTerm(String columnName, PredicateLeaf leaf, TransformSpec transformSpec) {
+    UnboundTerm<Object> column = tryVariantExtractTerm(columnName, leaf);
+    if (column != null) {
+      return column;
+    }
+    return ObjectUtils.defaultIfNull(
+        toTerm(columnName, transformSpec), Expressions.ref(columnName));
+  }
+
   public static UnboundTerm<Object> toTerm(String columnName, TransformSpec transformSpec) {
     if (transformSpec == null) {
       return null;
@@ -168,6 +179,66 @@ public static UnboundTerm<Object> toTerm(String columnName, TransformSpec transf
     }
   }
 
+  /**
+   * Converts a shredded variant pseudo-column (e.g. {@code data.typed_value.age}) into an Iceberg variant extract term
+   * (e.g. {@code extract(data, "$.age", "long")}).
+   *
+   * <p>This enables Iceberg to prune manifests/files using variant metrics produced when variant shredding is enabled.
+   */
+  private static UnboundTerm<Object> tryVariantExtractTerm(String columnName, PredicateLeaf leaf) {
+    int typedIdx = columnName.indexOf(VariantPathUtil.TYPED_VALUE_SEGMENT);
+    if (typedIdx < 0) {
+      return null;
+    }
+
+    String variantColumn = columnName.substring(0, typedIdx);
+    String extractedPath =
+        columnName.substring(typedIdx + VariantPathUtil.TYPED_VALUE_SEGMENT.length());
+    if (variantColumn.isEmpty() || extractedPath.isEmpty()) {
+      return null;
+    }
+
+    Type.PrimitiveType icebergType = extractPrimitiveType(leaf);
+    if (icebergType == null) {
+      return null;
+    }
+
+    // Build an RFC9535 shorthand JSONPath-like path: "$.field" or "$.a.b"
+    String jsonPath = "$." + extractedPath;
+    try {
+      return Expressions.extract(variantColumn, jsonPath, icebergType.toString());
+    } catch (RuntimeException e) {
+      // Invalid path/type; fall back to normal reference handling.
+      return null;
+    }
+  }
+
+  private static Type.PrimitiveType extractPrimitiveType(PredicateLeaf leaf) {
+    // Returned types must serialize (via toString) into Iceberg primitive type strings accepted by
+    // Types.fromPrimitiveString.
+    switch (leaf.getType()) {
+      case LONG:
+        return Types.LongType.get();
+      case FLOAT:
+        // Hive SARG uses FLOAT for both float/double; using double is the safest default.
+        return Types.DoubleType.get();
+      case STRING:
+        return Types.StringType.get();
+      case BOOLEAN:
+        return Types.BooleanType.get();
+      case DATE:
+        return Types.DateType.get();
+      case TIMESTAMP:
+        // Iceberg timestamps are represented as micros in a long, but the Iceberg type is timestamp.
+        return Types.TimestampType.withoutZone();
+      case DECIMAL:
+        // Precision/scale are not available in the SARG leaf type.
+        return null;
+      default:
+        return null;
+    }
+  }
+
   // PredicateLeafImpl has a work-around for Kryo serialization with java.util.Date objects where it converts values to
   // Timestamp using Date#getTime. This conversion discards microseconds, so this is a necessary to avoid it.
   private static final DynFields.UnboundField<?> LITERAL_FIELD = DynFields.builder()

iceberg/iceberg-handler/src/main/java/org/apache/iceberg/mr/hive/HiveIcebergInputFormat.java

@@ -54,6 +54,7 @@
 import org.apache.iceberg.expressions.ResidualEvaluator;
 import org.apache.iceberg.hive.HiveVersion;
 import org.apache.iceberg.mr.InputFormatConfig;
+import org.apache.iceberg.mr.hive.variant.VariantFilterRewriter;
 import org.apache.iceberg.mr.mapred.AbstractMapredIcebergRecordReader;
 import org.apache.iceberg.mr.mapred.Container;
 import org.apache.iceberg.mr.mapred.MapredIcebergInputFormat;
@@ -91,65 +92,129 @@ public class HiveIcebergInputFormat extends MapredIcebergInputFormat<Record>
   }
 
   /**
-   * Converts the Hive filter found in the job conf to an Iceberg filter expression.
-   * @param conf - job conf
-   * @return - Iceberg data filter expression
+   * Encapsulates planning-time and reader-time Iceberg filter expressions derived from Hive predicates.
    */
-  static Expression icebergDataFilterFromHiveConf(Configuration conf) {
-    Expression icebergFilter = SerializationUtil.deserializeFromBase64(conf.get(InputFormatConfig.FILTER_EXPRESSION));
-    if (icebergFilter != null) {
-      // in case we already have it prepared..
-      return icebergFilter;
+  private static final class FilterExpressions {
+
+    private static Expression planningFilter(Configuration conf) {
+      // Planning-safe filter (extract removed) may already be serialized for reuse.
+      Expression planningFilter = SerializationUtil
+          .deserializeFromBase64(conf.get(InputFormatConfig.FILTER_EXPRESSION));
+      if (planningFilter != null) {
+        // in case we already have it prepared..
+        return planningFilter;
+      }
+      // Reader filter should retain extract(...) for row-group pruning. Rebuild from Hive predicate to avoid losing
+      // variant rewrites when planningFilter was stripped.
+      Expression readerFilter = icebergDataFilterFromHiveConf(conf);
+      if (readerFilter != null) {
+        return VariantFilterRewriter.stripVariantExtractPredicates(readerFilter);
+      }
+      return null;
     }
-    String hiveFilter = conf.get(TableScanDesc.FILTER_EXPR_CONF_STR);
-    if (hiveFilter != null) {
-      ExprNodeGenericFuncDesc exprNodeDesc = SerializationUtilities
-          .deserializeObject(hiveFilter, ExprNodeGenericFuncDesc.class);
-      return getFilterExpr(conf, exprNodeDesc);
+
+    private static Expression icebergDataFilterFromHiveConf(Configuration conf) {
+      // Build an Iceberg filter from Hive's serialized predicate so we can preserve extract(...) terms for
+      // reader-level pruning (e.g. Parquet shredded VARIANT row-group pruning).
+      //
+      // This intentionally does NOT consult FILTER_EXPRESSION, because FILTER_EXPRESSION must remain safe for
+      // Iceberg planning-time utilities (some of which cannot stringify extract(...) terms).
+      String hiveFilter = conf.get(TableScanDesc.FILTER_EXPR_CONF_STR);
+      if (hiveFilter != null) {
+        ExprNodeGenericFuncDesc exprNodeDesc =
+            SerializationUtilities.deserializeObject(hiveFilter, ExprNodeGenericFuncDesc.class);
+        return getFilterExpr(conf, exprNodeDesc);
+      }
+      return null;
+    }
+
+    private static Expression planningResidual(FileScanTask task, Configuration conf) {
+      return residual(task, conf, planningFilter(conf));
+    }
+
+    private static Expression readerResidual(FileScanTask task, Configuration conf) {
+      return residual(task, conf, icebergDataFilterFromHiveConf(conf));
+    }
+
+    private static Expression residual(FileScanTask task, Configuration conf, Expression filter) {
+      if (filter == null) {
+        return Expressions.alwaysTrue();
+      }
+      boolean caseSensitive = conf.getBoolean(
+          InputFormatConfig.CASE_SENSITIVE, InputFormatConfig.CASE_SENSITIVE_DEFAULT);
+
+      return ResidualEvaluator.of(task.spec(), filter, caseSensitive)
+          .residualFor(task.file().partition());
     }
-    return null;
   }
 
   /**
-   * getFilterExpr extracts search argument from ExprNodeGenericFuncDesc and returns Iceberg Filter Expression
+   * Builds an Iceberg filter expression from a Hive predicate expression node.
    * @param conf - job conf
    * @param exprNodeDesc - Describes a GenericFunc node
    * @return Iceberg Filter Expression
    */
   static Expression getFilterExpr(Configuration conf, ExprNodeGenericFuncDesc exprNodeDesc) {
-    if (exprNodeDesc != null) {
-      SearchArgument sarg = ConvertAstToSearchArg.create(conf, exprNodeDesc);
-      try {
-        return HiveIcebergFilterFactory.generateFilterExpression(sarg);
-      } catch (UnsupportedOperationException e) {
-        LOG.warn("Unable to create Iceberg filter, proceeding without it (will be applied by Hive later): ", e);
+    if (exprNodeDesc == null) {
+      return null;
+    }
+
+    ExprNodeGenericFuncDesc exprForSarg = exprNodeDesc;
+    if (Boolean.parseBoolean(conf.get(InputFormatConfig.VARIANT_SHREDDING_ENABLED))) {
+      ExprNodeGenericFuncDesc rewritten = VariantFilterRewriter.rewriteForShredding(exprNodeDesc);
+      if (rewritten != null) {
+        exprForSarg = rewritten;
       }
     }
-    return null;
+
+    SearchArgument sarg = ConvertAstToSearchArg.create(conf, exprForSarg);
+    if (sarg == null) {
+      return null;
+    }
+
+    try {
+      return HiveIcebergFilterFactory.generateFilterExpression(sarg);
+    } catch (UnsupportedOperationException e) {
+      LOG.warn(
+          "Unable to create Iceberg filter, proceeding without it (will be applied by Hive later): ",
+          e);
+      return null;
+    }
   }
 
   /**
-   * Converts Hive filter found in the passed job conf to an Iceberg filter expression. Then evaluates this
-   * against the task's partition value producing a residual filter expression.
+   * Returns a residual expression that is safe to apply as a record-level filter.
+   *
+   * <p>This residual is derived from the task-level Iceberg planning filter (already extract-free) after
+   * evaluating it against the task's partition value.
    * @param task - file scan task to evaluate the expression against
    * @param conf - job conf
    * @return - Iceberg residual filter expression
    */
   public static Expression residualForTask(FileScanTask task, Configuration conf) {
-    Expression dataFilter = icebergDataFilterFromHiveConf(conf);
-    if (dataFilter == null) {
-      return Expressions.alwaysTrue();
-    }
-    return ResidualEvaluator.of(
-        task.spec(), dataFilter,
-        conf.getBoolean(InputFormatConfig.CASE_SENSITIVE, InputFormatConfig.CASE_SENSITIVE_DEFAULT)
-    ).residualFor(task.file().partition());
+    return FilterExpressions.planningResidual(task, conf);
+  }
+
+  /**
+   * Returns a residual expression intended only for reader-level pruning (best-effort).
+   *
+   * <p>This residual is derived from the task-level Iceberg filter after evaluating it against the task's
+   * partition value. It may include {@code extract(...)} predicates and is suitable for formats/readers that
+   * can leverage such terms for pruning (e.g. Parquet row-group pruning using shredded VARIANT columns).
+   *
+   * <p><strong>Do not</strong> use this for record-level residual filtering, as {@code extract} cannot be
+   * evaluated at record level in Iceberg readers.
+   */
+  public static Expression residualForReaderPruning(FileScanTask task, Configuration conf) {
+    return FilterExpressions.readerResidual(task, conf);
   }
 
   @Override
   public InputSplit[] getSplits(JobConf job, int numSplits) throws IOException {
-    Expression filter = icebergDataFilterFromHiveConf(job);
+    Expression filter = FilterExpressions.planningFilter(job);
     if (filter != null) {
+      // Iceberg planning-time utilities may attempt to stringify the filter. Ensure the planning filter never
+      // contains extract(...) or shredded typed_value references.
       job.set(InputFormatConfig.FILTER_EXPRESSION, SerializationUtil.serializeToBase64(filter));
     }
 

iceberg/iceberg-handler/src/main/java/org/apache/iceberg/mr/hive/HiveIcebergStorageHandler.java

@@ -187,7 +187,6 @@
 import org.apache.iceberg.mr.hive.actions.HiveIcebergDeleteOrphanFiles;
 import org.apache.iceberg.mr.hive.plan.IcebergBucketFunction;
 import org.apache.iceberg.mr.hive.udf.GenericUDFIcebergZorder;
-import org.apache.iceberg.parquet.VariantUtil;
 import org.apache.iceberg.puffin.Blob;
 import org.apache.iceberg.puffin.BlobMetadata;
 import org.apache.iceberg.puffin.Puffin;
@@ -1752,8 +1751,7 @@ private void fallbackToNonVectorizedModeBasedOnProperties(Properties tableProps)
     if (FileFormat.AVRO == IcebergTableUtil.defaultFileFormat(tableProps::getProperty) ||
         isValidMetadataTable(tableProps.getProperty(IcebergAcidUtil.META_TABLE_PROPERTY)) ||
         hasOrcTimeInSchema(tableProps, tableSchema) ||
-        !hasParquetNestedTypeWithinListOrMap(tableProps, tableSchema) ||
-        VariantUtil.shouldUseVariantShredding(tableProps::getProperty, tableSchema)) {
+        !hasParquetNestedTypeWithinListOrMap(tableProps, tableSchema)) {
       // disable vectorization
       SessionStateUtil.getQueryState(conf).ifPresent(queryState ->
           queryState.getConf().setBoolVar(ConfVars.HIVE_VECTORIZATION_ENABLED, false));