fix(reader): ensure batch readers return zero-offset Arrow arrays (alibaba#440)

Author: lxy-9602

include/paimon/reader/batch_reader.h

@@ -40,6 +40,8 @@ class PAIMON_EXPORT BatchReader {
     /// If EOF is reached, returns an OK status with a nullptr array. Returns an error status only
     /// for critical failures (e.g., IO errors). Once an error is returned, this method must not be
     /// retried, as it will repeatedly return the same error code.
+    /// \note IMPORTANT: A non-EOF ArrowArray and all its nested child arrays must have offset 0 to
+    /// avoid potential issues during conversion through the Arrow C Data Interface.
     ///
     /// @return A result containing a `::ReadBatch`, which consists of a unique pointer to
     /// `ArrowArray` and a unique pointer to `ArrowSchema`. Returned array contains a `_VALUE_KIND`
@@ -52,6 +54,8 @@ class PAIMON_EXPORT BatchReader {
     /// If EOF is reached, returns an OK status with a nullptr array. Returns an error status only
     /// for critical failures (e.g., IO errors). Once an error is returned, this method must not be
     /// retried, as it will repeatedly return the same error code.
+    /// \note IMPORTANT: A non-EOF ArrowArray and all its nested child arrays must have offset 0 to
+    /// avoid potential issues during conversion through the Arrow C Data Interface.
     ///
     /// @return A result containing a `::ReadBatch` and a valid bitmap. `::ReadBatch` consists of a
     /// unique pointer to `ArrowArray` and a unique pointer to `ArrowSchema`. Returned array

src/paimon/common/utils/arrow/arrow_utils.cpp

@@ -18,13 +18,30 @@
 
 #include "arrow/array/array_base.h"
 #include "arrow/array/array_nested.h"
+#include "arrow/array/concatenate.h"
 #include "arrow/util/compression.h"
 #include "fmt/format.h"
 #include "paimon/common/utils/arrow/status_utils.h"
 #include "paimon/common/utils/string_utils.h"
 
 namespace paimon {
 
+namespace {
+
+bool HasNonZeroOffset(const std::shared_ptr<arrow::ArrayData>& data) {
+    if (data->offset != 0) {
+        return true;
+    }
+    for (const auto& child : data->child_data) {
+        if (HasNonZeroOffset(child)) {
+            return true;
+        }
+    }
+    return false;
+}
+
+}  // namespace
+
 const char* ArrowUtils::kArrowSchemaMetadataKey = "ARROW:schema";
 
 Result<std::shared_ptr<arrow::Schema>> ArrowUtils::DataTypeToSchema(
@@ -167,6 +184,28 @@ Result<std::shared_ptr<arrow::StructArray>> ArrowUtils::RemoveFieldFromStructArr
     return array;
 }
 
+Result<std::shared_ptr<arrow::RecordBatch>> ArrowUtils::NormalizeRecordBatchOffsets(
+    const std::shared_ptr<arrow::RecordBatch>& record_batch, arrow::MemoryPool* pool) {
+    arrow::ArrayVector normalized_columns;
+    for (int32_t i = 0; i < record_batch->num_columns(); ++i) {
+        const std::shared_ptr<arrow::Array>& column = record_batch->column(i);
+        if (!HasNonZeroOffset(column->data())) {
+            continue;
+        }
+        if (normalized_columns.empty()) {
+            normalized_columns = record_batch->columns();
+        }
+        PAIMON_ASSIGN_OR_RAISE_FROM_ARROW(std::shared_ptr<arrow::Array> normalized_column,
+                                          arrow::Concatenate({column}, pool));
+        normalized_columns[i] = std::move(normalized_column);
+    }
+    if (normalized_columns.empty()) {
+        return record_batch;
+    }
+    return arrow::RecordBatch::Make(record_batch->schema(), record_batch->num_rows(),
+                                    std::move(normalized_columns));
+}
+
 Result<arrow::Compression::type> ArrowUtils::GetCompressionType(const std::string& compression) {
     std::string normalized = StringUtils::ToLowerCase(compression);
     if (normalized.empty() || normalized == "none") {

src/paimon/common/utils/arrow/arrow_utils.h

@@ -48,6 +48,9 @@ class PAIMON_EXPORT ArrowUtils {
     static Result<std::shared_ptr<arrow::StructArray>> RemoveFieldFromStructArray(
         const std::shared_ptr<arrow::StructArray>& struct_array, const std::string& field_name);
 
+    static Result<std::shared_ptr<arrow::RecordBatch>> NormalizeRecordBatchOffsets(
+        const std::shared_ptr<arrow::RecordBatch>& record_batch, arrow::MemoryPool* pool);
+
     static bool EqualsIgnoreNullable(const std::shared_ptr<arrow::DataType>& type,
                                      const std::shared_ptr<arrow::DataType>& other_type);
 

src/paimon/common/utils/arrow/arrow_utils_test.cpp

@@ -379,6 +379,54 @@ TEST(ArrowUtilsTest, TestRemoveFieldFromStructArraySuccess) {
     ASSERT_TRUE(result->Equals(expected_struct_array));
 }
 
+TEST(ArrowUtilsTest, TestNormalizeRecordBatchOffsets) {
+    auto value_field = arrow::field("value", arrow::int32());
+    auto nested_field = arrow::field("nested", arrow::struct_({value_field}));
+    auto text_field = arrow::field("text", arrow::utf8());
+    auto clean_field = arrow::field("clean", arrow::boolean());
+    auto schema = arrow::schema({nested_field, text_field, clean_field});
+
+    std::shared_ptr<arrow::Array> values =
+        arrow::ipc::internal::json::ArrayFromJSON(arrow::int32(), "[0, 1, 2, 3, 4]").ValueOrDie();
+    std::shared_ptr<arrow::Array> sliced_values = values->Slice(1, 3);
+    std::shared_ptr<arrow::StructArray> nested_column =
+        arrow::StructArray::Make({sliced_values}, {value_field->name()}).ValueOrDie();
+    std::shared_ptr<arrow::Array> text =
+        arrow::ipc::internal::json::ArrayFromJSON(arrow::utf8(), R"(["a", "b", "c", "d", "e"])")
+            .ValueOrDie();
+    std::shared_ptr<arrow::Array> sliced_text = text->Slice(1, 3);
+    std::shared_ptr<arrow::Array> clean_column =
+        arrow::ipc::internal::json::ArrayFromJSON(arrow::boolean(), "[true, false, true]")
+            .ValueOrDie();
+    std::shared_ptr<arrow::RecordBatch> record_batch = arrow::RecordBatch::Make(
+        schema, /*num_rows=*/3, {nested_column, sliced_text, clean_column});
+
+    ASSERT_EQ(nested_column->offset(), 0);
+    ASSERT_EQ(nested_column->field(0)->offset(), 1);
+    ASSERT_EQ(sliced_text->offset(), 1);
+    ASSERT_EQ(clean_column->offset(), 0);
+
+    ASSERT_OK_AND_ASSIGN(
+        std::shared_ptr<arrow::RecordBatch> normalized_batch,
+        ArrowUtils::NormalizeRecordBatchOffsets(record_batch, arrow::default_memory_pool()));
+    ASSERT_NE(normalized_batch.get(), record_batch.get());
+    ASSERT_TRUE(normalized_batch->Equals(*record_batch));
+    std::shared_ptr<arrow::StructArray> normalized_nested =
+        std::static_pointer_cast<arrow::StructArray>(normalized_batch->column(0));
+    ASSERT_EQ(normalized_nested->offset(), 0);
+    ASSERT_EQ(normalized_nested->field(0)->offset(), 0);
+    ASSERT_EQ(normalized_batch->column(1)->offset(), 0);
+    ASSERT_EQ(normalized_batch->column(2)->offset(), 0);
+    ASSERT_NE(normalized_batch->column_data(0).get(), record_batch->column_data(0).get());
+    ASSERT_NE(normalized_batch->column_data(1).get(), record_batch->column_data(1).get());
+    ASSERT_EQ(normalized_batch->column_data(2).get(), record_batch->column_data(2).get());
+
+    ASSERT_OK_AND_ASSIGN(
+        std::shared_ptr<arrow::RecordBatch> unchanged_batch,
+        ArrowUtils::NormalizeRecordBatchOffsets(normalized_batch, arrow::default_memory_pool()));
+    ASSERT_EQ(unchanged_batch.get(), normalized_batch.get());
+}
+
 TEST(ArrowUtilsTest, TestEqualsIgnoreNullable) {
     {
         // test simple
@@ -481,6 +529,7 @@ TEST(ArrowUtilsTest, TestGetCompressionType) {
         ASSERT_EQ(type, arrow::Compression::GZIP);
     }
     {
+        // test invalid codec
         ASSERT_NOK(ArrowUtils::GetCompressionType("invalid_codec"));
     }
 }

src/paimon/format/parquet/file_reader_wrapper.cpp

@@ -24,6 +24,7 @@
 #include "arrow/record_batch.h"
 #include "arrow/util/range.h"
 #include "fmt/format.h"
+#include "paimon/common/utils/arrow/arrow_utils.h"
 #include "paimon/format/parquet/column_index_filter.h"
 #include "paimon/format/parquet/page_filtered_row_group_reader.h"
 #include "paimon/format/parquet/parquet_format_defs.h"
@@ -273,6 +274,10 @@ Result<std::shared_ptr<arrow::RecordBatch>> FileReaderWrapper::NextFullyMatched(
     if (!record_batch) {
         return std::shared_ptr<arrow::RecordBatch>();
     }
+    // Large binary columns (exceed 2GB) may split at different row boundaries. TableBatchReader
+    // aligns their chunks by slicing columns, which may leave non-zero child offsets.
+    PAIMON_ASSIGN_OR_RAISE(record_batch,
+                           ArrowUtils::NormalizeRecordBatchOffsets(record_batch, pool_.get()));
 
     int32_t rg_id = target_row_groups_[current_row_group_idx_].GetRowGroupIndex();
     uint64_t rg_end = all_row_group_ranges_[rg_id].second;

src/paimon/format/parquet/page_filtered_row_group_reader.cpp

@@ -26,6 +26,7 @@
 #include "arrow/table.h"
 #include "arrow/util/future.h"
 #include "fmt/format.h"
+#include "paimon/common/utils/arrow/arrow_utils.h"
 #include "paimon/common/utils/arrow/status_utils.h"
 #include "parquet/arrow/reader_internal.h"
 #include "parquet/metadata.h"
@@ -36,12 +37,13 @@ namespace paimon::parquet {
 namespace {
 
 /// Wraps an arrow::Table + TableBatchReader as a RecordBatchReader so the caller can
-/// stream zero-copy-sliced batches without deep-copying multi-chunk columns. The Table
-/// is held to keep its ChunkedArrays alive for the inner TableBatchReader.
+/// stream batches while ensuring every returned array offset is zero. The Table is held
+/// to keep its ChunkedArrays alive for the inner TableBatchReader.
 class TableRecordBatchReader : public arrow::RecordBatchReader {
  public:
-    TableRecordBatchReader(std::shared_ptr<arrow::Table> table, int64_t chunksize)
-        : table_(std::move(table)), inner_(*table_) {
+    TableRecordBatchReader(std::shared_ptr<arrow::Table> table, int64_t chunksize,
+                           std::shared_ptr<arrow::MemoryPool> pool)
+        : table_(std::move(table)), inner_(*table_), pool_(std::move(pool)) {
         inner_.set_chunksize(chunksize);
     }
 
@@ -50,12 +52,26 @@ class TableRecordBatchReader : public arrow::RecordBatchReader {
     }
 
     arrow::Status ReadNext(std::shared_ptr<arrow::RecordBatch>* out) override {
-        return inner_.ReadNext(out);
+        ARROW_RETURN_NOT_OK(inner_.ReadNext(out));
+        if (!*out) {
+            return arrow::Status::OK();
+        }
+
+        // Page filtering may produce columns with different chunk boundaries. TableBatchReader
+        // aligns them by slicing columns, which may leave non-zero child offsets.
+        Result<std::shared_ptr<arrow::RecordBatch>> normalized_result =
+            ArrowUtils::NormalizeRecordBatchOffsets(*out, pool_.get());
+        if (!normalized_result.ok()) {
+            return ToArrowStatus(normalized_result.status());
+        }
+        *out = std::move(normalized_result).value();
+        return arrow::Status::OK();
     }
 
  private:
     std::shared_ptr<arrow::Table> table_;
     arrow::TableBatchReader inner_;
+    std::shared_ptr<arrow::MemoryPool> pool_;
 };
 
 }  // namespace
@@ -240,7 +256,8 @@ Result<std::unique_ptr<arrow::RecordBatchReader>> PageFilteredRowGroupReader::Re
     if (row_ranges.IsEmpty()) {
         PAIMON_ASSIGN_OR_RAISE_FROM_ARROW(std::shared_ptr<arrow::Table> empty_table,
                                           arrow::Table::MakeEmpty(arrow_schema, pool.get()));
-        return std::make_unique<TableRecordBatchReader>(std::move(empty_table), max_chunksize);
+        return std::make_unique<TableRecordBatchReader>(std::move(empty_table), max_chunksize,
+                                                        pool);
     }
 
     int64_t expected_rows = row_ranges.RowCount();
@@ -283,7 +300,8 @@ Result<std::unique_ptr<arrow::RecordBatchReader>> PageFilteredRowGroupReader::Re
     }
 
     auto table = arrow::Table::Make(arrow_schema, std::move(columns), expected_rows);
-    return std::make_unique<TableRecordBatchReader>(std::move(table), max_chunksize);
+    return std::make_unique<TableRecordBatchReader>(std::move(table), max_chunksize,
+                                                    std::move(pool));
 }
 
 std::vector<::arrow::io::ReadRange> PageFilteredRowGroupReader::ComputePageRanges(

src/paimon/format/parquet/page_filtered_row_group_reader_test.cpp

@@ -661,14 +661,13 @@ TEST_F(PageFilteredRowGroupReaderTest, ComputePageRangesMultiplePages) {
     ASSERT_LT(ranges[0].offset, ranges[1].offset);
 }
 
-/// Test: variable-length columns are streamed across multiple zero-copy-sliced
+/// Test: variable-length columns are streamed across multiple offset-normalized
 /// RecordBatches when batch_size is smaller than the matched row count, instead of
 /// being concatenated into a single RecordBatch via CombineChunks.
 ///
 /// This verifies the alignment with Arrow's standard TableBatchReader path:
-/// multi-chunk binary/string columns split along chunk + batch_size boundaries,
-/// with no deep copy. Asserts both correctness (total rows + full content order) and
-/// the multi-batch shape (more than one chunk in the collected ChunkedArray).
+/// multi-chunk binary/string columns split along chunk + batch_size boundaries. It
+/// asserts correctness and the multi-batch shape.
 TEST_F(PageFilteredRowGroupReaderTest, StringColumnMultiBatchStreaming) {
     std::string file_name = dir_->Str() + "/string_multi_batch.parquet";
 
@@ -721,6 +720,23 @@ TEST_F(PageFilteredRowGroupReaderTest, StringColumnMultiBatchStreaming) {
     ASSERT_EQ(40, seen);
 }
 
+TEST_F(PageFilteredRowGroupReaderTest, NormalizesSlicedBatchOffsets) {
+    std::string file_name = dir_->Str() + "/normalized_sliced_offsets.parquet";
+    std::shared_ptr<arrow::StructArray> data = MakeSequentialIntData(60);
+    WriteTestFile(file_name, data, /*write_batch_size=*/10, /*max_row_group_length=*/60);
+
+    std::shared_ptr<arrow::Schema> read_schema =
+        arrow::schema({arrow::field("val", arrow::int32())});
+    std::shared_ptr<Predicate> predicate = PredicateBuilder::GreaterOrEqual(
+        /*field_index=*/0, /*field_name=*/"val", FieldType::INT, Literal(20));
+
+    std::shared_ptr<arrow::ChunkedArray> result;
+    ReadWithPredicateImpl(file_name, read_schema, predicate, &result, /*batch_size=*/7);
+    ASSERT_TRUE(result);
+    ASSERT_EQ(40, result->length());
+    ASSERT_GT(result->num_chunks(), 1);
+}
+
 /// Test: end-to-end page-filtered read produces correct results when using page-level PreBuffer.
 ///
 /// This exercises the full path: ComputePageRanges → PreBufferRanges → CachedInputStream →

src/paimon/testing/utils/read_result_collector.h

@@ -180,6 +180,7 @@ class ReadResultCollector {
                 if (BatchReader::IsEofBatch(batch_with_bitmap)) {
                     return std::shared_ptr<arrow::Array>();
                 }
+                PAIMON_RETURN_NOT_OK(CheckBatchOffset(batch_with_bitmap.first));
                 assert(!batch_with_bitmap.second.IsEmpty());
                 PAIMON_ASSIGN_OR_RAISE(
                     batch, ReaderUtils::ApplyBitmapToReadBatch(std::move(batch_with_bitmap),
@@ -192,12 +193,30 @@ class ReadResultCollector {
             if (BatchReader::IsEofBatch(batch)) {
                 return std::shared_ptr<arrow::Array>();
             }
+            PAIMON_RETURN_NOT_OK(CheckBatchOffset(batch));
         }
         auto& [c_array, c_schema] = batch;
         assert(c_array->length > 0);
         PAIMON_ASSIGN_OR_RAISE_FROM_ARROW(auto result_array,
                                           arrow::ImportArray(c_array.get(), c_schema.get()));
         return result_array;
     }
+
+    static Status CheckBatchOffset(const BatchReader::ReadBatch& batch) {
+        assert(!BatchReader::IsEofBatch(batch));
+        return CheckArrayOffset(batch.first.get());
+    }
+
+    static Status CheckArrayOffset(const ArrowArray* array) {
+        assert(array);
+        if (array->offset != 0) {
+            return Status::Invalid("BatchReader returned an array with non-zero offset " +
+                                   std::to_string(array->offset));
+        }
+        for (int64_t i = 0; i < array->n_children; i++) {
+            PAIMON_RETURN_NOT_OK(CheckArrayOffset(array->children[i]));
+        }
+        return Status::OK();
+    }
 };
 }  // namespace paimon::test