steveloughran commented on code in PR #14297:
URL: https://github.com/apache/iceberg/pull/14297#discussion_r3167845288
##########
parquet/src/main/java/org/apache/iceberg/parquet/ParquetFormatModel.java:
##########
@@ -209,8 +254,45 @@ public FileAppender<D> build() throws IOException {
throw new IllegalArgumentException("Unknown file content: " +
content);
}
+ if (shredVariants) {
+ return buildShreddedAppender();
+ }
+
return internal.build();
}
+
+ /**
+ * Creates a {@link BufferedFileAppender} that buffers the first N rows,
runs variant shredding
+ * analysis on them, then creates the real Parquet appender with a
shredded schema.
+ *
+ * <p>Only top-level variant columns are shredded. Nested variants (inside
structs/lists/maps)
+ * fall through to unshredded 2-field layout because column index
resolution only applies to
+ * top-level fields.
+ */
+ private FileAppender<D> buildShreddedAppender() {
+ return new BufferedFileAppender<>(
+ bufferSize,
+ bufferedRows -> {
+ Map<Integer, Type> shreddedTypes =
+ variantAnalyzer.analyzeVariantColumns(bufferedRows, schema,
engineSchema);
+
+ if (!shreddedTypes.isEmpty()) {
Review Comment:
I think it would be prescient to add some logging at debug level here.
##########
parquet/src/main/java/org/apache/iceberg/parquet/VariantShreddingAnalyzer.java:
##########
@@ -0,0 +1,532 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you 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.
+ */
+package org.apache.iceberg.parquet;
+
+import java.math.BigDecimal;
+import java.util.List;
+import java.util.Map;
+import java.util.Set;
+import org.apache.iceberg.Schema;
+import org.apache.iceberg.relocated.com.google.common.collect.ImmutableMap;
+import org.apache.iceberg.relocated.com.google.common.collect.Lists;
+import org.apache.iceberg.relocated.com.google.common.collect.Maps;
+import org.apache.iceberg.relocated.com.google.common.collect.Sets;
+import org.apache.iceberg.types.Types.NestedField;
+import org.apache.iceberg.variants.PhysicalType;
+import org.apache.iceberg.variants.VariantArray;
+import org.apache.iceberg.variants.VariantObject;
+import org.apache.iceberg.variants.VariantPrimitive;
+import org.apache.iceberg.variants.VariantValue;
+import org.apache.parquet.schema.GroupType;
+import org.apache.parquet.schema.LogicalTypeAnnotation;
+import org.apache.parquet.schema.PrimitiveType;
+import org.apache.parquet.schema.Type;
+import org.apache.parquet.schema.Types;
+
+/**
+ * Analyzes variant data across buffered rows to determine an optimal
shredding schema.
+ *
+ * <p>Determinism contract: for a given set of variant values (regardless of
row arrival order),
+ * this analyzer produces the same shredded schema. When the number of
distinct fields at any level
+ * exceeds {@code MAX_INTERMEDIATE_FIELDS}, field tracking becomes
insertion-order dependent and
+ * determinism is not guaranteed.
+ *
+ * <ul>
+ * <li>Object fields use a TreeMap, so field ordering is alphabetical and
deterministic.
+ * <li>Type selection picks the most common type with explicit tie-break
priority (see
+ * TIE_BREAK_PRIORITY), not enum ordinal.
+ * <li>Integer types (INT8/16/32/64) and decimal types (DECIMAL4/8/16) are
each promoted to the
+ * widest observed before competing with other types.
+ * <li>Fields below {@code MIN_FIELD_FREQUENCY} are pruned. Above {@code
MAX_SHREDDED_FIELDS}, the
+ * most frequent are kept with alphabetical tie-breaking.
+ * <li>Recursion into nested objects/arrays stops at {@code
MAX_SHREDDING_DEPTH} (default 50).
+ * <li>New struct fields are not tracked once a node reaches {@code
MAX_INTERMEDIATE_FIELDS}
+ * (default 1000) to bound memory during inference.
+ * </ul>
+ *
+ * <p>This contract holds within a single batch. Different batches with
different distributions may
+ * produce different layouts; cross-batch stability requires schema pinning
(not yet implemented).
+ *
+ * <p>Subclasses implement {@link #extractVariantValues} to convert
engine-specific row types into
+ * {@link VariantValue} instances.
+ *
+ * @param <T> the engine-specific row type (e.g., Spark InternalRow, Flink
RowData)
+ * @param <S> the engine-specific schema type (e.g., Spark StructType, Flink
RowType)
+ */
+public abstract class VariantShreddingAnalyzer<T, S> {
+ private static final String TYPED_VALUE = "typed_value";
+ private static final String VALUE = "value";
+ private static final String ELEMENT = "element";
+ private static final double MIN_FIELD_FREQUENCY = 0.10;
+ private static final int MAX_SHREDDED_FIELDS = 300;
+ private static final int MAX_SHREDDING_DEPTH = 50;
+ private static final int MAX_INTERMEDIATE_FIELDS = 1000;
+
+ protected VariantShreddingAnalyzer() {}
+
+ /**
+ * Analyzes buffered variant values to determine the optimal shredding
schema.
+ *
+ * @param bufferedRows the buffered rows to analyze
+ * @param variantFieldIndex the index of the variant field in the rows
+ * @return the shredded schema type, or null if no shredding should be
performed
+ */
+ public Type analyzeAndCreateSchema(List<T> bufferedRows, int
variantFieldIndex) {
+ List<VariantValue> variantValues = extractVariantValues(bufferedRows,
variantFieldIndex);
+ if (variantValues.isEmpty()) {
+ return null;
+ }
+
+ PathNode root = buildPathTree(variantValues);
+ PhysicalType rootType = root.info.getMostCommonType();
+ if (rootType == null) {
+ return null;
+ }
+
+ pruneInfrequentFields(root, root.info.observationCount);
+
+ return buildTypedValue(root, rootType);
+ }
+
+ protected abstract List<VariantValue> extractVariantValues(
+ List<T> bufferedRows, int variantFieldIndex);
+
+ /**
+ * Resolves a column name to its index in the engine-specific schema.
Returns -1 if the column is
+ * not found.
+ */
+ protected abstract int resolveColumnIndex(S engineSchema, String columnName);
+
+ /**
+ * Analyzes all variant columns in the schema, resolving column indices via
the engine-specific
+ * {@link #resolveColumnIndex} method.
+ *
+ * @param bufferedRows the buffered rows to analyze
+ * @param icebergSchema the Iceberg table schema
+ * @param engineSchema the engine-specific schema used to resolve column
indices
+ * @return a map from Iceberg field ID to the shredded Parquet type for each
variant column
+ */
+ public Map<Integer, Type> analyzeVariantColumns(
+ List<T> bufferedRows, Schema icebergSchema, S engineSchema) {
+ Map<Integer, Type> shreddedTypes = Maps.newHashMap();
+ for (NestedField col : icebergSchema.columns()) {
+ if (col.type().isVariantType()) {
+ int rowIndex = resolveColumnIndex(engineSchema, col.name());
+ if (rowIndex >= 0) {
+ Type typed = analyzeAndCreateSchema(bufferedRows, rowIndex);
+ if (typed != null) {
+ shreddedTypes.put(col.fieldId(), typed);
+ }
+ }
+ }
+ }
+
+ return shreddedTypes;
+ }
+
+ private static PathNode buildPathTree(List<VariantValue> variantValues) {
+ PathNode root = new PathNode(null);
+ root.info = new FieldInfo();
+
+ for (VariantValue value : variantValues) {
+ traverse(root, value, 0);
+ }
+
+ return root;
+ }
+
+ private static void pruneInfrequentFields(PathNode node, int totalRows) {
+ if (node.objectChildren.isEmpty() && node.arrayElement == null) {
+ return;
+ }
+
+ // Remove fields below frequency threshold
+ node.objectChildren
+ .entrySet()
+ .removeIf(
+ entry -> {
+ FieldInfo info = entry.getValue().info;
+ return info != null
+ && ((double) info.observationCount / totalRows) <
MIN_FIELD_FREQUENCY;
+ });
+
+ // Cap at MAX_SHREDDED_FIELDS, keep the most frequently observed
+ if (node.objectChildren.size() > MAX_SHREDDED_FIELDS) {
+ List<Map.Entry<String, PathNode>> sorted =
Lists.newArrayList(node.objectChildren.entrySet());
+ sorted.sort(
+ (a, b) -> {
+ int cmp =
+ Integer.compare(
+ b.getValue().info.observationCount,
a.getValue().info.observationCount);
+ return cmp != 0 ? cmp : a.getKey().compareTo(b.getKey());
+ });
+ Set<String> keep = Sets.newHashSet();
+ for (int i = 0; i < MAX_SHREDDED_FIELDS; i++) {
+ keep.add(sorted.get(i).getKey());
+ }
+ node.objectChildren.entrySet().removeIf(entry ->
!keep.contains(entry.getKey()));
+ }
+
+ // Recurse into remaining object children
+ for (PathNode child : node.objectChildren.values()) {
+ pruneInfrequentFields(child, totalRows);
+ }
+
+ // Recurse into array elements (arrays of objects need pruning too)
+ if (node.arrayElement != null) {
+ pruneInfrequentFields(node.arrayElement, totalRows);
+ }
+ }
+
+ private static void traverse(PathNode node, VariantValue value, int depth) {
+ if (value == null || value.type() == PhysicalType.NULL) {
+ return;
+ }
+
+ node.info.observe(value);
+
+ if (value.type() == PhysicalType.OBJECT && depth < MAX_SHREDDING_DEPTH) {
+ traverseObject(node, value.asObject(), depth);
+ } else if (value.type() == PhysicalType.ARRAY && depth <
MAX_SHREDDING_DEPTH) {
+ traverseArray(node, value.asArray(), depth);
+ }
+ }
+
+ private static void traverseObject(PathNode node, VariantObject obj, int
depth) {
+ for (String fieldName : obj.fieldNames()) {
+ VariantValue fieldValue = obj.get(fieldName);
+ if (fieldValue != null) {
+ PathNode childNode = node.objectChildren.get(fieldName);
+ if (childNode == null) {
+ if (node.objectChildren.size() >= MAX_INTERMEDIATE_FIELDS) {
+ continue;
+ }
+ childNode = new PathNode(fieldName);
+ childNode.info = new FieldInfo();
+ node.objectChildren.put(fieldName, childNode);
+ }
+ traverse(childNode, fieldValue, depth + 1);
+ }
+ }
+ }
+
+ // observationCount inside arrays counts per-element, not per-row, so fields
in long arrays
+ // have inflated frequency and resist pruning.
+ private static void traverseArray(PathNode node, VariantArray array, int
depth) {
+ int numElements = array.numElements();
+ if (node.arrayElement == null) {
+ node.arrayElement = new PathNode(null);
+ node.arrayElement.info = new FieldInfo();
+ }
+ for (int i = 0; i < numElements; i++) {
+ VariantValue element = array.get(i);
+ if (element != null) {
+ traverse(node.arrayElement, element, depth + 1);
+ }
+ }
+ }
+
+ private static Type buildFieldGroup(PathNode node) {
+ PhysicalType commonType = node.info.getMostCommonType();
+ if (commonType == null) {
+ return null;
+ }
+
+ Type typedValue = buildTypedValue(node, commonType);
+ if (typedValue == null) {
+ return null;
+ }
+
+ return Types.buildGroup(Type.Repetition.REQUIRED)
+ .optional(PrimitiveType.PrimitiveTypeName.BINARY)
+ .named(VALUE)
+ .addField(typedValue)
+ .named(node.fieldName);
+ }
+
+ private static Type buildTypedValue(PathNode node, PhysicalType
physicalType) {
+ return switch (physicalType) {
+ case ARRAY -> createArrayTypedValue(node);
+ case OBJECT -> createObjectTypedValue(node);
+ default -> createPrimitiveTypedValue(node.info, physicalType);
+ };
+ }
+
+ private static Type createObjectTypedValue(PathNode node) {
+ if (node.objectChildren.isEmpty()) {
+ return null;
+ }
+
+ Types.GroupBuilder<GroupType> builder =
Types.buildGroup(Type.Repetition.OPTIONAL);
+ boolean hasFields = false;
+ for (PathNode child : node.objectChildren.values()) {
+ Type fieldType = buildFieldGroup(child);
+ if (fieldType != null) {
+ builder.addField(fieldType);
+ hasFields = true;
+ }
+ }
+
+ return hasFields ? builder.named(TYPED_VALUE) : null;
+ }
+
+ private static Type createArrayTypedValue(PathNode node) {
+ PathNode elementNode = node.arrayElement;
+ if (elementNode == null) {
+ return null;
+ }
+ PhysicalType elementType = elementNode.info.getMostCommonType();
+ if (elementType == null) {
+ return null;
+ }
+ Type elementTypedValue = buildTypedValue(elementNode, elementType);
+ if (elementTypedValue == null) {
+ return null;
+ }
+
+ GroupType elementGroup =
+ Types.buildGroup(Type.Repetition.REQUIRED)
+ .optional(PrimitiveType.PrimitiveTypeName.BINARY)
+ .named(VALUE)
+ .addField(elementTypedValue)
+ .named(ELEMENT);
+
+ return Types.optionalList().element(elementGroup).named(TYPED_VALUE);
+ }
+
+ private static class PathNode {
+ private final String fieldName;
+ private final Map<String, PathNode> objectChildren = Maps.newTreeMap();
+ private PathNode arrayElement = null;
+ private FieldInfo info = null;
+
+ private PathNode(String fieldName) {
+ this.fieldName = fieldName;
+ }
+ }
+
+ /** Use DECIMAL with maximum precision and scale as the shredding type */
+ private static Type createDecimalTypedValue(FieldInfo info) {
+ int maxPrecision = Math.min(info.maxDecimalIntegerDigits +
info.maxDecimalScale, 38);
+ int maxScale = Math.min(info.maxDecimalScale, Math.max(0, 38 -
info.maxDecimalIntegerDigits));
+
+ if (maxPrecision <= 9) {
+ return Types.optional(PrimitiveType.PrimitiveTypeName.INT32)
+ .as(LogicalTypeAnnotation.decimalType(maxScale, maxPrecision))
+ .named(TYPED_VALUE);
+ } else if (maxPrecision <= 18) {
+ return Types.optional(PrimitiveType.PrimitiveTypeName.INT64)
+ .as(LogicalTypeAnnotation.decimalType(maxScale, maxPrecision))
+ .named(TYPED_VALUE);
+ } else {
+ return
Types.optional(PrimitiveType.PrimitiveTypeName.FIXED_LEN_BYTE_ARRAY)
+ .length(16)
+ .as(LogicalTypeAnnotation.decimalType(maxScale, maxPrecision))
+ .named(TYPED_VALUE);
+ }
+ }
+
+ private static Type createPrimitiveTypedValue(FieldInfo info, PhysicalType
primitiveType) {
+ return switch (primitiveType) {
+ case BOOLEAN_TRUE, BOOLEAN_FALSE ->
+
Types.optional(PrimitiveType.PrimitiveTypeName.BOOLEAN).named(TYPED_VALUE);
+ case INT8 ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT32)
+ .as(LogicalTypeAnnotation.intType(8, true))
+ .named(TYPED_VALUE);
+ case INT16 ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT32)
+ .as(LogicalTypeAnnotation.intType(16, true))
+ .named(TYPED_VALUE);
+ case INT32 ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT32)
+ .as(LogicalTypeAnnotation.intType(32, true))
+ .named(TYPED_VALUE);
+ case INT64 ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT64)
+ .as(LogicalTypeAnnotation.intType(64, true))
+ .named(TYPED_VALUE);
+ case FLOAT ->
Types.optional(PrimitiveType.PrimitiveTypeName.FLOAT).named(TYPED_VALUE);
+ case DOUBLE ->
Types.optional(PrimitiveType.PrimitiveTypeName.DOUBLE).named(TYPED_VALUE);
+ case STRING ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.BINARY)
+ .as(LogicalTypeAnnotation.stringType())
+ .named(TYPED_VALUE);
+ case BINARY ->
Types.optional(PrimitiveType.PrimitiveTypeName.BINARY).named(TYPED_VALUE);
+ case TIME ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT64)
+ .as(LogicalTypeAnnotation.timeType(false,
LogicalTypeAnnotation.TimeUnit.MICROS))
+ .named(TYPED_VALUE);
+ case DATE ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT32)
+ .as(LogicalTypeAnnotation.dateType())
+ .named(TYPED_VALUE);
+ case TIMESTAMPTZ ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT64)
+ .as(LogicalTypeAnnotation.timestampType(true,
LogicalTypeAnnotation.TimeUnit.MICROS))
+ .named(TYPED_VALUE);
+ case TIMESTAMPNTZ ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT64)
+ .as(LogicalTypeAnnotation.timestampType(false,
LogicalTypeAnnotation.TimeUnit.MICROS))
+ .named(TYPED_VALUE);
+ case TIMESTAMPTZ_NANOS ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT64)
+ .as(LogicalTypeAnnotation.timestampType(true,
LogicalTypeAnnotation.TimeUnit.NANOS))
+ .named(TYPED_VALUE);
+ case TIMESTAMPNTZ_NANOS ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT64)
+ .as(LogicalTypeAnnotation.timestampType(false,
LogicalTypeAnnotation.TimeUnit.NANOS))
+ .named(TYPED_VALUE);
+ case DECIMAL4, DECIMAL8, DECIMAL16 -> createDecimalTypedValue(info);
+ case UUID ->
Review Comment:
really curious about the benefits of shredding this. Any use case where the
UUID is unique on every row has no compression, and the min/max stats are
useless too.
##########
parquet/src/main/java/org/apache/iceberg/parquet/VariantShreddingAnalyzer.java:
##########
@@ -0,0 +1,532 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you 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.
+ */
+package org.apache.iceberg.parquet;
+
+import java.math.BigDecimal;
+import java.util.List;
+import java.util.Map;
+import java.util.Set;
+import org.apache.iceberg.Schema;
+import org.apache.iceberg.relocated.com.google.common.collect.ImmutableMap;
+import org.apache.iceberg.relocated.com.google.common.collect.Lists;
+import org.apache.iceberg.relocated.com.google.common.collect.Maps;
+import org.apache.iceberg.relocated.com.google.common.collect.Sets;
+import org.apache.iceberg.types.Types.NestedField;
+import org.apache.iceberg.variants.PhysicalType;
+import org.apache.iceberg.variants.VariantArray;
+import org.apache.iceberg.variants.VariantObject;
+import org.apache.iceberg.variants.VariantPrimitive;
+import org.apache.iceberg.variants.VariantValue;
+import org.apache.parquet.schema.GroupType;
+import org.apache.parquet.schema.LogicalTypeAnnotation;
+import org.apache.parquet.schema.PrimitiveType;
+import org.apache.parquet.schema.Type;
+import org.apache.parquet.schema.Types;
+
+/**
+ * Analyzes variant data across buffered rows to determine an optimal
shredding schema.
+ *
+ * <p>Determinism contract: for a given set of variant values (regardless of
row arrival order),
+ * this analyzer produces the same shredded schema. When the number of
distinct fields at any level
+ * exceeds {@code MAX_INTERMEDIATE_FIELDS}, field tracking becomes
insertion-order dependent and
+ * determinism is not guaranteed.
+ *
+ * <ul>
+ * <li>Object fields use a TreeMap, so field ordering is alphabetical and
deterministic.
+ * <li>Type selection picks the most common type with explicit tie-break
priority (see
+ * TIE_BREAK_PRIORITY), not enum ordinal.
+ * <li>Integer types (INT8/16/32/64) and decimal types (DECIMAL4/8/16) are
each promoted to the
+ * widest observed before competing with other types.
+ * <li>Fields below {@code MIN_FIELD_FREQUENCY} are pruned. Above {@code
MAX_SHREDDED_FIELDS}, the
+ * most frequent are kept with alphabetical tie-breaking.
+ * <li>Recursion into nested objects/arrays stops at {@code
MAX_SHREDDING_DEPTH} (default 50).
+ * <li>New struct fields are not tracked once a node reaches {@code
MAX_INTERMEDIATE_FIELDS}
+ * (default 1000) to bound memory during inference.
+ * </ul>
+ *
+ * <p>This contract holds within a single batch. Different batches with
different distributions may
+ * produce different layouts; cross-batch stability requires schema pinning
(not yet implemented).
+ *
+ * <p>Subclasses implement {@link #extractVariantValues} to convert
engine-specific row types into
+ * {@link VariantValue} instances.
+ *
+ * @param <T> the engine-specific row type (e.g., Spark InternalRow, Flink
RowData)
+ * @param <S> the engine-specific schema type (e.g., Spark StructType, Flink
RowType)
+ */
+public abstract class VariantShreddingAnalyzer<T, S> {
+ private static final String TYPED_VALUE = "typed_value";
+ private static final String VALUE = "value";
+ private static final String ELEMENT = "element";
+ private static final double MIN_FIELD_FREQUENCY = 0.10;
+ private static final int MAX_SHREDDED_FIELDS = 300;
Review Comment:
This is going to explode schema complexity; I can see on my query work that
this has an impact on Rowgroup skipping as there's more overhead on parsing
this ... I'm trying hard to optimise it w/ lazy compute, but it's still
expensive.
any record with, 2+ variant objects could now have 600+ child elements.
##########
parquet/src/main/java/org/apache/iceberg/parquet/ParquetFormatModel.java:
##########
@@ -119,13 +154,21 @@ public ModelWriteBuilder<D, S> engineSchema(S newSchema) {
@Override
public ModelWriteBuilder<D, S> set(String property, String value) {
+ if (SHRED_VARIANTS_KEY.equals(property)) {
Review Comment:
what about a switch statement here?
##########
docs/docs/configuration.md:
##########
@@ -50,6 +50,8 @@ Iceberg tables support table properties to configure table
behavior, like the de
| write.parquet.dict-size-bytes | 2097152 (2 MB)
| Parquet dictionary page size
|
| write.parquet.compression-codec | zstd
| Parquet compression codec: zstd, brotli, lz4, gzip, snappy, uncompressed
|
| write.parquet.compression-level | null
| Parquet compression level
|
+| write.parquet.variant.shred | false
| When true, variant columns are written with shredded Parquet encoding
for improved query performance
|
Review Comment:
"for better compression and possibly also improved query performance".
Compression holds, but query perf needs
* ongoing work through the stack
* queries actually applied to the shredded values
##########
parquet/src/main/java/org/apache/iceberg/parquet/VariantShreddingAnalyzer.java:
##########
@@ -0,0 +1,532 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you 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.
+ */
+package org.apache.iceberg.parquet;
+
+import java.math.BigDecimal;
+import java.util.List;
+import java.util.Map;
+import java.util.Set;
+import org.apache.iceberg.Schema;
+import org.apache.iceberg.relocated.com.google.common.collect.ImmutableMap;
+import org.apache.iceberg.relocated.com.google.common.collect.Lists;
+import org.apache.iceberg.relocated.com.google.common.collect.Maps;
+import org.apache.iceberg.relocated.com.google.common.collect.Sets;
+import org.apache.iceberg.types.Types.NestedField;
+import org.apache.iceberg.variants.PhysicalType;
+import org.apache.iceberg.variants.VariantArray;
+import org.apache.iceberg.variants.VariantObject;
+import org.apache.iceberg.variants.VariantPrimitive;
+import org.apache.iceberg.variants.VariantValue;
+import org.apache.parquet.schema.GroupType;
+import org.apache.parquet.schema.LogicalTypeAnnotation;
+import org.apache.parquet.schema.PrimitiveType;
+import org.apache.parquet.schema.Type;
+import org.apache.parquet.schema.Types;
+
+/**
+ * Analyzes variant data across buffered rows to determine an optimal
shredding schema.
+ *
+ * <p>Determinism contract: for a given set of variant values (regardless of
row arrival order),
+ * this analyzer produces the same shredded schema. When the number of
distinct fields at any level
+ * exceeds {@code MAX_INTERMEDIATE_FIELDS}, field tracking becomes
insertion-order dependent and
+ * determinism is not guaranteed.
+ *
+ * <ul>
+ * <li>Object fields use a TreeMap, so field ordering is alphabetical and
deterministic.
+ * <li>Type selection picks the most common type with explicit tie-break
priority (see
+ * TIE_BREAK_PRIORITY), not enum ordinal.
+ * <li>Integer types (INT8/16/32/64) and decimal types (DECIMAL4/8/16) are
each promoted to the
+ * widest observed before competing with other types.
+ * <li>Fields below {@code MIN_FIELD_FREQUENCY} are pruned. Above {@code
MAX_SHREDDED_FIELDS}, the
+ * most frequent are kept with alphabetical tie-breaking.
+ * <li>Recursion into nested objects/arrays stops at {@code
MAX_SHREDDING_DEPTH} (default 50).
+ * <li>New struct fields are not tracked once a node reaches {@code
MAX_INTERMEDIATE_FIELDS}
+ * (default 1000) to bound memory during inference.
+ * </ul>
+ *
+ * <p>This contract holds within a single batch. Different batches with
different distributions may
+ * produce different layouts; cross-batch stability requires schema pinning
(not yet implemented).
+ *
+ * <p>Subclasses implement {@link #extractVariantValues} to convert
engine-specific row types into
+ * {@link VariantValue} instances.
+ *
+ * @param <T> the engine-specific row type (e.g., Spark InternalRow, Flink
RowData)
+ * @param <S> the engine-specific schema type (e.g., Spark StructType, Flink
RowType)
+ */
+public abstract class VariantShreddingAnalyzer<T, S> {
+ private static final String TYPED_VALUE = "typed_value";
+ private static final String VALUE = "value";
+ private static final String ELEMENT = "element";
+ private static final double MIN_FIELD_FREQUENCY = 0.10;
+ private static final int MAX_SHREDDED_FIELDS = 300;
+ private static final int MAX_SHREDDING_DEPTH = 50;
+ private static final int MAX_INTERMEDIATE_FIELDS = 1000;
+
+ protected VariantShreddingAnalyzer() {}
+
+ /**
+ * Analyzes buffered variant values to determine the optimal shredding
schema.
+ *
+ * @param bufferedRows the buffered rows to analyze
+ * @param variantFieldIndex the index of the variant field in the rows
+ * @return the shredded schema type, or null if no shredding should be
performed
+ */
+ public Type analyzeAndCreateSchema(List<T> bufferedRows, int
variantFieldIndex) {
+ List<VariantValue> variantValues = extractVariantValues(bufferedRows,
variantFieldIndex);
+ if (variantValues.isEmpty()) {
+ return null;
+ }
+
+ PathNode root = buildPathTree(variantValues);
+ PhysicalType rootType = root.info.getMostCommonType();
+ if (rootType == null) {
+ return null;
+ }
+
+ pruneInfrequentFields(root, root.info.observationCount);
+
+ return buildTypedValue(root, rootType);
+ }
+
+ protected abstract List<VariantValue> extractVariantValues(
+ List<T> bufferedRows, int variantFieldIndex);
+
+ /**
+ * Resolves a column name to its index in the engine-specific schema.
Returns -1 if the column is
+ * not found.
+ */
+ protected abstract int resolveColumnIndex(S engineSchema, String columnName);
+
+ /**
+ * Analyzes all variant columns in the schema, resolving column indices via
the engine-specific
+ * {@link #resolveColumnIndex} method.
+ *
+ * @param bufferedRows the buffered rows to analyze
+ * @param icebergSchema the Iceberg table schema
+ * @param engineSchema the engine-specific schema used to resolve column
indices
+ * @return a map from Iceberg field ID to the shredded Parquet type for each
variant column
+ */
+ public Map<Integer, Type> analyzeVariantColumns(
+ List<T> bufferedRows, Schema icebergSchema, S engineSchema) {
+ Map<Integer, Type> shreddedTypes = Maps.newHashMap();
+ for (NestedField col : icebergSchema.columns()) {
+ if (col.type().isVariantType()) {
+ int rowIndex = resolveColumnIndex(engineSchema, col.name());
+ if (rowIndex >= 0) {
+ Type typed = analyzeAndCreateSchema(bufferedRows, rowIndex);
+ if (typed != null) {
+ shreddedTypes.put(col.fieldId(), typed);
+ }
+ }
+ }
+ }
+
+ return shreddedTypes;
+ }
+
+ private static PathNode buildPathTree(List<VariantValue> variantValues) {
+ PathNode root = new PathNode(null);
+ root.info = new FieldInfo();
+
+ for (VariantValue value : variantValues) {
+ traverse(root, value, 0);
+ }
+
+ return root;
+ }
+
+ private static void pruneInfrequentFields(PathNode node, int totalRows) {
+ if (node.objectChildren.isEmpty() && node.arrayElement == null) {
+ return;
+ }
+
+ // Remove fields below frequency threshold
+ node.objectChildren
+ .entrySet()
+ .removeIf(
+ entry -> {
+ FieldInfo info = entry.getValue().info;
+ return info != null
+ && ((double) info.observationCount / totalRows) <
MIN_FIELD_FREQUENCY;
+ });
+
+ // Cap at MAX_SHREDDED_FIELDS, keep the most frequently observed
+ if (node.objectChildren.size() > MAX_SHREDDED_FIELDS) {
+ List<Map.Entry<String, PathNode>> sorted =
Lists.newArrayList(node.objectChildren.entrySet());
+ sorted.sort(
+ (a, b) -> {
+ int cmp =
+ Integer.compare(
+ b.getValue().info.observationCount,
a.getValue().info.observationCount);
+ return cmp != 0 ? cmp : a.getKey().compareTo(b.getKey());
+ });
+ Set<String> keep = Sets.newHashSet();
+ for (int i = 0; i < MAX_SHREDDED_FIELDS; i++) {
+ keep.add(sorted.get(i).getKey());
+ }
+ node.objectChildren.entrySet().removeIf(entry ->
!keep.contains(entry.getKey()));
+ }
+
+ // Recurse into remaining object children
+ for (PathNode child : node.objectChildren.values()) {
+ pruneInfrequentFields(child, totalRows);
+ }
+
+ // Recurse into array elements (arrays of objects need pruning too)
+ if (node.arrayElement != null) {
+ pruneInfrequentFields(node.arrayElement, totalRows);
+ }
+ }
+
+ private static void traverse(PathNode node, VariantValue value, int depth) {
+ if (value == null || value.type() == PhysicalType.NULL) {
+ return;
+ }
+
+ node.info.observe(value);
+
+ if (value.type() == PhysicalType.OBJECT && depth < MAX_SHREDDING_DEPTH) {
+ traverseObject(node, value.asObject(), depth);
+ } else if (value.type() == PhysicalType.ARRAY && depth <
MAX_SHREDDING_DEPTH) {
+ traverseArray(node, value.asArray(), depth);
+ }
+ }
+
+ private static void traverseObject(PathNode node, VariantObject obj, int
depth) {
+ for (String fieldName : obj.fieldNames()) {
+ VariantValue fieldValue = obj.get(fieldName);
+ if (fieldValue != null) {
+ PathNode childNode = node.objectChildren.get(fieldName);
+ if (childNode == null) {
+ if (node.objectChildren.size() >= MAX_INTERMEDIATE_FIELDS) {
+ continue;
+ }
+ childNode = new PathNode(fieldName);
+ childNode.info = new FieldInfo();
+ node.objectChildren.put(fieldName, childNode);
+ }
+ traverse(childNode, fieldValue, depth + 1);
+ }
+ }
+ }
+
+ // observationCount inside arrays counts per-element, not per-row, so fields
in long arrays
+ // have inflated frequency and resist pruning.
+ private static void traverseArray(PathNode node, VariantArray array, int
depth) {
+ int numElements = array.numElements();
+ if (node.arrayElement == null) {
+ node.arrayElement = new PathNode(null);
+ node.arrayElement.info = new FieldInfo();
+ }
+ for (int i = 0; i < numElements; i++) {
+ VariantValue element = array.get(i);
+ if (element != null) {
+ traverse(node.arrayElement, element, depth + 1);
+ }
+ }
+ }
+
+ private static Type buildFieldGroup(PathNode node) {
+ PhysicalType commonType = node.info.getMostCommonType();
+ if (commonType == null) {
+ return null;
+ }
+
+ Type typedValue = buildTypedValue(node, commonType);
+ if (typedValue == null) {
+ return null;
+ }
+
+ return Types.buildGroup(Type.Repetition.REQUIRED)
+ .optional(PrimitiveType.PrimitiveTypeName.BINARY)
+ .named(VALUE)
+ .addField(typedValue)
+ .named(node.fieldName);
+ }
+
+ private static Type buildTypedValue(PathNode node, PhysicalType
physicalType) {
+ return switch (physicalType) {
+ case ARRAY -> createArrayTypedValue(node);
+ case OBJECT -> createObjectTypedValue(node);
+ default -> createPrimitiveTypedValue(node.info, physicalType);
+ };
+ }
+
+ private static Type createObjectTypedValue(PathNode node) {
+ if (node.objectChildren.isEmpty()) {
+ return null;
+ }
+
+ Types.GroupBuilder<GroupType> builder =
Types.buildGroup(Type.Repetition.OPTIONAL);
+ boolean hasFields = false;
+ for (PathNode child : node.objectChildren.values()) {
+ Type fieldType = buildFieldGroup(child);
+ if (fieldType != null) {
+ builder.addField(fieldType);
+ hasFields = true;
+ }
+ }
+
+ return hasFields ? builder.named(TYPED_VALUE) : null;
+ }
+
+ private static Type createArrayTypedValue(PathNode node) {
+ PathNode elementNode = node.arrayElement;
+ if (elementNode == null) {
+ return null;
+ }
+ PhysicalType elementType = elementNode.info.getMostCommonType();
+ if (elementType == null) {
+ return null;
+ }
+ Type elementTypedValue = buildTypedValue(elementNode, elementType);
+ if (elementTypedValue == null) {
+ return null;
+ }
+
+ GroupType elementGroup =
+ Types.buildGroup(Type.Repetition.REQUIRED)
+ .optional(PrimitiveType.PrimitiveTypeName.BINARY)
+ .named(VALUE)
+ .addField(elementTypedValue)
+ .named(ELEMENT);
+
+ return Types.optionalList().element(elementGroup).named(TYPED_VALUE);
+ }
+
+ private static class PathNode {
+ private final String fieldName;
+ private final Map<String, PathNode> objectChildren = Maps.newTreeMap();
+ private PathNode arrayElement = null;
+ private FieldInfo info = null;
+
+ private PathNode(String fieldName) {
+ this.fieldName = fieldName;
+ }
+ }
+
+ /** Use DECIMAL with maximum precision and scale as the shredding type */
+ private static Type createDecimalTypedValue(FieldInfo info) {
+ int maxPrecision = Math.min(info.maxDecimalIntegerDigits +
info.maxDecimalScale, 38);
+ int maxScale = Math.min(info.maxDecimalScale, Math.max(0, 38 -
info.maxDecimalIntegerDigits));
+
+ if (maxPrecision <= 9) {
+ return Types.optional(PrimitiveType.PrimitiveTypeName.INT32)
+ .as(LogicalTypeAnnotation.decimalType(maxScale, maxPrecision))
+ .named(TYPED_VALUE);
+ } else if (maxPrecision <= 18) {
+ return Types.optional(PrimitiveType.PrimitiveTypeName.INT64)
+ .as(LogicalTypeAnnotation.decimalType(maxScale, maxPrecision))
+ .named(TYPED_VALUE);
+ } else {
+ return
Types.optional(PrimitiveType.PrimitiveTypeName.FIXED_LEN_BYTE_ARRAY)
+ .length(16)
+ .as(LogicalTypeAnnotation.decimalType(maxScale, maxPrecision))
+ .named(TYPED_VALUE);
+ }
+ }
+
+ private static Type createPrimitiveTypedValue(FieldInfo info, PhysicalType
primitiveType) {
+ return switch (primitiveType) {
+ case BOOLEAN_TRUE, BOOLEAN_FALSE ->
+
Types.optional(PrimitiveType.PrimitiveTypeName.BOOLEAN).named(TYPED_VALUE);
+ case INT8 ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT32)
+ .as(LogicalTypeAnnotation.intType(8, true))
+ .named(TYPED_VALUE);
+ case INT16 ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT32)
+ .as(LogicalTypeAnnotation.intType(16, true))
+ .named(TYPED_VALUE);
+ case INT32 ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT32)
+ .as(LogicalTypeAnnotation.intType(32, true))
+ .named(TYPED_VALUE);
+ case INT64 ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT64)
+ .as(LogicalTypeAnnotation.intType(64, true))
+ .named(TYPED_VALUE);
+ case FLOAT ->
Types.optional(PrimitiveType.PrimitiveTypeName.FLOAT).named(TYPED_VALUE);
+ case DOUBLE ->
Types.optional(PrimitiveType.PrimitiveTypeName.DOUBLE).named(TYPED_VALUE);
+ case STRING ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.BINARY)
+ .as(LogicalTypeAnnotation.stringType())
+ .named(TYPED_VALUE);
+ case BINARY ->
Types.optional(PrimitiveType.PrimitiveTypeName.BINARY).named(TYPED_VALUE);
+ case TIME ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT64)
+ .as(LogicalTypeAnnotation.timeType(false,
LogicalTypeAnnotation.TimeUnit.MICROS))
+ .named(TYPED_VALUE);
+ case DATE ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT32)
+ .as(LogicalTypeAnnotation.dateType())
+ .named(TYPED_VALUE);
+ case TIMESTAMPTZ ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT64)
+ .as(LogicalTypeAnnotation.timestampType(true,
LogicalTypeAnnotation.TimeUnit.MICROS))
+ .named(TYPED_VALUE);
+ case TIMESTAMPNTZ ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT64)
+ .as(LogicalTypeAnnotation.timestampType(false,
LogicalTypeAnnotation.TimeUnit.MICROS))
+ .named(TYPED_VALUE);
+ case TIMESTAMPTZ_NANOS ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT64)
+ .as(LogicalTypeAnnotation.timestampType(true,
LogicalTypeAnnotation.TimeUnit.NANOS))
+ .named(TYPED_VALUE);
+ case TIMESTAMPNTZ_NANOS ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.INT64)
+ .as(LogicalTypeAnnotation.timestampType(false,
LogicalTypeAnnotation.TimeUnit.NANOS))
+ .named(TYPED_VALUE);
+ case DECIMAL4, DECIMAL8, DECIMAL16 -> createDecimalTypedValue(info);
+ case UUID ->
+ Types.optional(PrimitiveType.PrimitiveTypeName.FIXED_LEN_BYTE_ARRAY)
+ .length(16)
+ .as(LogicalTypeAnnotation.uuidType())
+ .named(TYPED_VALUE);
+ default ->
+ throw new UnsupportedOperationException(
+ "Unknown primitive physical type: " + primitiveType);
+ };
+ }
+
+ /** Tracks occurrence count and types for a single field. */
+ private static class FieldInfo {
+ private final Map<PhysicalType, Integer> typeCounts = Maps.newHashMap();
+ private int maxDecimalScale = 0;
+ private int maxDecimalIntegerDigits = 0;
+ private int observationCount = 0;
+
+ private static final Map<PhysicalType, Integer> INTEGER_PRIORITY =
+ ImmutableMap.of(
+ PhysicalType.INT8, 0,
+ PhysicalType.INT16, 1,
+ PhysicalType.INT32, 2,
+ PhysicalType.INT64, 3);
+
+ private static final Map<PhysicalType, Integer> DECIMAL_PRIORITY =
+ ImmutableMap.of(
+ PhysicalType.DECIMAL4, 0,
+ PhysicalType.DECIMAL8, 1,
+ PhysicalType.DECIMAL16, 2);
+
+ private static final Map<PhysicalType, Integer> TIE_BREAK_PRIORITY =
Review Comment:
explain role in javadocs. And if it is "what to prefer to shred", I'd put
STRING above BINARY as string compares char sequences;
binary just looks at bytes and then length.
IF I was being ruthless I'd put binary below nano timestamps.
It's just that I'd expect a lot more strings to go in than binaries, and in
future those nano timestamps.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
