github-actions[bot] commented on code in PR #65674:
URL: https://github.com/apache/doris/pull/65674#discussion_r3609729393
##########
be/src/format_v2/parquet/parquet_file_context.cpp:
##########
@@ -129,10 +199,83 @@ bool should_use_merge_range_reader(const
std::vector<ParquetPageCacheRange>& ran
avg_io_size < io::MergeRangeFileReader::SMALL_IO;
}
+bool should_stage_small_http_file(std::string_view path, size_t file_size,
+ size_t in_memory_file_size) {
+ return file_size <= in_memory_file_size &&
+ (path.starts_with("http://";) || path.starts_with("https://";));
+}
+
} // namespace detail
namespace {
+constexpr uint8_t V2_PARQUET_MAGIC[4] = {'P', 'A', 'R', '1'};
+constexpr size_t V2_PARQUET_FOOTER_SIZE = 8;
+constexpr size_t V2_INITIAL_FOOTER_READ_SIZE = 48 * 1024;
+
+Status parse_native_parquet_footer(io::FileReaderSPtr file,
+ std::unique_ptr<NativeParquetMetadata>*
metadata,
+ size_t* footer_size, io::IOContext* io_ctx,
+ bool enable_mapping_varbinary,
+ bool enable_mapping_timestamp_tz) {
+ DORIS_CHECK(file != nullptr);
+ DORIS_CHECK(metadata != nullptr);
+ DORIS_CHECK(footer_size != nullptr);
+ const size_t file_size = file->size();
+ if (file_size < V2_PARQUET_FOOTER_SIZE) {
+ return Status::Corruption("Parquet v2 file is too small for a footer:
{}", file_size);
+ }
+
+ const size_t tail_size = std::min(file_size, V2_INITIAL_FOOTER_READ_SIZE);
+ std::vector<uint8_t> tail(tail_size);
+ size_t bytes_read = 0;
+ RETURN_IF_ERROR(file->read_at(file_size - tail_size, Slice(tail.data(),
tail.size()),
+ &bytes_read, io_ctx));
+ if (bytes_read != tail.size()) {
+ return Status::Corruption("Short Parquet v2 footer read: expected {},
got {}", tail.size(),
+ bytes_read);
+ }
+ const auto* magic = tail.data() + tail.size() - sizeof(V2_PARQUET_MAGIC);
+ if (memcmp(magic, V2_PARQUET_MAGIC, sizeof(V2_PARQUET_MAGIC)) != 0) {
+ return Status::Corruption("Invalid Parquet v2 footer magic in {}",
file->path().native());
+ }
+
+ const uint32_t serialized_size =
+ decode_fixed32_le(tail.data() + tail.size() -
V2_PARQUET_FOOTER_SIZE);
+ if (serialized_size > file_size - V2_PARQUET_FOOTER_SIZE) {
+ // Footer lengths are untrusted. Validate before
subtraction/allocation so a malformed
+ // small file cannot redirect the v2 reader or request an oversized
metadata buffer.
+ return Status::Corruption("Parquet v2 footer size {} exceeds file size
{}", serialized_size,
+ file_size);
+ }
+ std::vector<uint8_t> serialized_metadata(serialized_size);
Review Comment:
[P1] Bound the native footer before allocating it
`serialized_size` comes from the file tail and is checked only against the
object's length, so a large corrupt object can advertise a multi-gigabyte
footer and make every cold V2 open allocate and read that entire buffer before
Thrift can reject it. There is no footer-byte budget here, and concurrent
external-file splits can multiply the allocation. Enforce a configured/sane
metadata limit (and tracked reservation as appropriate) before constructing
this vector, and add an over-limit test that fails before the advertised footer
is allocated or read.
##########
be/src/format_v2/parquet/parquet_file_context.cpp:
##########
@@ -535,83 +678,343 @@ Status arrow_status_to_doris_status(const arrow::Status&
status) {
}
Status ParquetFileContext::open(io::FileReaderSPtr input_file_reader,
io::IOContext* io_ctx,
- bool enable_page_cache,
- const io::FileDescription& file_description) {
+ bool enable_page_cache, const
io::FileDescription& file_description,
+ bool enable_mapping_timestamp_tz, bool
enable_mapping_varbinary) {
DORIS_CHECK(input_file_reader != nullptr);
- auto page_cache_file_key = build_page_cache_file_key(*input_file_reader,
file_description);
- arrow_file =
std::make_shared<DorisRandomAccessFile>(std::move(input_file_reader), io_ctx,
- enable_page_cache,
-
std::move(page_cache_file_key));
+ if
(detail::should_stage_small_http_file(input_file_reader->path().native(),
+ input_file_reader->size(),
+ config::in_memory_file_size)) {
+ // A metadata-cache hit can make the first physical read start inside
a tiny HTTP file.
+ // Read it from byte zero once so EOF-range quirks cannot make warm
scans less reliable
+ // than cold scans, while keeping this compatibility policy entirely
inside v2.
+ native_file =
std::make_shared<io::InMemoryFileReader>(std::move(input_file_reader));
+ } else {
+ native_file = std::move(input_file_reader);
+ }
+ native_io_ctx = io_ctx;
+
+ // V2 owns its footer payload and cache identity. Mapping flags affect the
parsed schema, and a
+ // distinct suffix prevents a v1 FileMetaData value from being cast as the
v2-owned type.
+ auto* meta_cache = ExecEnv::GetInstance()->file_meta_cache();
+ auto meta_cache_key = FileMetaCache::get_key(native_file,
file_description);
Review Comment:
[P1] Give the V2 footer cache a stable file identity
The appended mapping flags separate schema modes, but
`FileMetaCache::get_key()` still uses only the reader's native path plus mtime
(or size when mtime is unknown). HDFS readers strip the nameservice, so equal
`/x.parquet` paths in two `fs_name` namespaces with the same mtime/size
collide; an unknown-mtime mutable file overwritten at the same path and size
also reuses stale metadata. That cached schema, row-group offsets, statistics,
and index locations then drive reads and pruning for the second object. Build
the base key with `fs_name` and the same version/immutability policy already
used by the V2 page cache, bypass cache use for mutable unknown-version files,
and add cross-filesystem and overwrite tests.
##########
be/src/format_v2/parquet/parquet_statistics.cpp:
##########
@@ -1035,6 +1251,321 @@ Status select_row_groups_by_metadata(
namespace {
+bool native_metadata_predicate_is_type_safe(const ParquetColumnSchema&
column_schema) {
+ DORIS_CHECK(column_schema.type != nullptr);
+ // Raw VARBINARY file slots may feed table-side STRING casts. Footer/page
metadata is still in
+ // the pre-cast domain, so using it for a rewritten table predicate can
cause false negatives.
+ return remove_nullable(column_schema.type)->get_primitive_type() !=
TYPE_VARBINARY;
+}
+
+bool check_native_statistics(const tparquet::RowGroup& row_group,
+ const
std::vector<std::unique_ptr<ParquetColumnSchema>>& file_schema,
+ const format::FileScanRequest& request,
+ ParquetPruningStats* pruning_stats, const
cctz::time_zone* timezone) {
+ const auto slot_indexes =
collect_expr_zonemap_slot_indexes(request.conjuncts);
+ if (slot_indexes.empty()) {
+ return false;
+ }
+ ZoneMapEvalContext ctx;
+ for (const int slot_index : slot_indexes) {
+ const auto file_column_id = file_column_id_by_block_position(request,
slot_index);
+ if (!file_column_id.has_value()) {
+ continue;
+ }
+ const auto* column_schema = resolve_local_leaf_schema(file_schema,
*file_column_id);
+ if (column_schema == nullptr || column_schema->type == nullptr ||
+ !native_metadata_predicate_is_type_safe(*column_schema) ||
+ column_schema->leaf_column_id >=
static_cast<int>(row_group.columns.size())) {
+ continue;
+ }
+ const auto& chunk = row_group.columns[column_schema->leaf_column_id];
+ std::shared_ptr<segment_v2::ZoneMap> zone_map;
+ if (chunk.__isset.meta_data) {
+ const auto& column_metadata = chunk.meta_data;
+ const auto* statistics =
+ column_metadata.__isset.statistics ?
&column_metadata.statistics : nullptr;
+ if (statistics != nullptr &&
!detail::can_use_native_footer_min_max(
+
column_schema->type_descriptor, *statistics)) {
+ statistics = nullptr;
+ }
+ zone_map = ParquetStatisticsUtils::MakeZoneMap(
+ ParquetStatisticsUtils::TransformColumnStatistics(
+ *column_schema, statistics,
column_metadata.num_values, timezone));
+ }
+ add_slot_zonemap(&ctx, slot_index, column_schema->type,
std::move(zone_map));
+ }
+ const auto result =
VExprContext::evaluate_zonemap_filter(request.conjuncts, ctx);
+ accumulate_zonemap_stats(ctx, pruning_stats);
+ return result == ZoneMapFilterResult::kNoMatch;
+}
+
+bool is_native_dictionary_data_encoding(tparquet::Encoding::type encoding) {
+ return encoding == tparquet::Encoding::PLAIN_DICTIONARY ||
+ encoding == tparquet::Encoding::RLE_DICTIONARY;
+}
+
+bool is_native_level_encoding(tparquet::Encoding::type encoding) {
+ return encoding == tparquet::Encoding::RLE || encoding ==
tparquet::Encoding::BIT_PACKED;
+}
+
+bool is_native_dictionary_encoded_chunk(const tparquet::ColumnMetaData&
metadata) {
+ if (!metadata.__isset.dictionary_page_offset ||
metadata.dictionary_page_offset < 0) {
+ return false;
+ }
+ if (metadata.__isset.encoding_stats && !metadata.encoding_stats.empty()) {
+ bool has_dictionary_data_page = false;
+ for (const auto& encoding_stat : metadata.encoding_stats) {
+ if ((encoding_stat.page_type != tparquet::PageType::DATA_PAGE &&
+ encoding_stat.page_type != tparquet::PageType::DATA_PAGE_V2)
||
+ encoding_stat.count <= 0) {
+ continue;
+ }
+ if (!is_native_dictionary_data_encoding(encoding_stat.encoding)) {
+ return false;
+ }
+ has_dictionary_data_page = true;
+ }
+ return has_dictionary_data_page;
+ }
+ bool has_dictionary_encoding = false;
+ for (const auto encoding : metadata.encodings) {
+ if (is_native_dictionary_data_encoding(encoding)) {
+ has_dictionary_encoding = true;
+ } else if (!is_native_level_encoding(encoding)) {
+ return false;
+ }
+ }
+ return has_dictionary_encoding;
+}
+
+const format::LocalColumnIndex* find_request_projection(const
format::FileScanRequest& request,
+ format::LocalColumnId
file_column_id) {
+ for (const auto& projection : request.predicate_columns) {
+ if (projection.local_id() == file_column_id.value()) {
+ return &projection;
+ }
+ }
+ for (const auto& projection : request.non_predicate_columns) {
+ if (projection.local_id() == file_column_id.value()) {
+ return &projection;
+ }
+ }
+ return nullptr;
+}
+
+ParquetRowGroupPruneReason native_dictionary_prune_reason(
+ const tparquet::RowGroup& row_group, int row_group_idx,
+ const std::vector<std::unique_ptr<ParquetColumnSchema>>& file_schema,
+ const format::FileScanRequest& request, const cctz::time_zone*
timezone,
+ ParquetFileContext* file_context) {
+ if (file_context == nullptr || file_context->native_metadata == nullptr) {
+ return ParquetRowGroupPruneReason::NONE;
+ }
+ const auto conjuncts_by_slot = collect_conjuncts_by_single_slot(
+ request.conjuncts, expr_zonemap::single_slot_dictionary_index);
+ for (const auto& [slot_index, conjuncts] : conjuncts_by_slot) {
+ const auto file_column_id = file_column_id_by_block_position(request,
slot_index);
+ if (!file_column_id.has_value()) {
+ continue;
+ }
+ const auto* column_schema = resolve_local_leaf_schema(file_schema,
*file_column_id);
+ const auto* projection = find_request_projection(request,
*file_column_id);
+ if (column_schema == nullptr || projection == nullptr ||
column_schema->type == nullptr ||
+ !column_schema->type_descriptor.is_string_like ||
+ column_schema->leaf_column_id >=
static_cast<int>(row_group.columns.size())) {
+ continue;
+ }
+ if (!native_metadata_predicate_is_type_safe(*column_schema)) {
+ // The file-local VARBINARY may feed a table-side STRING cast.
Pruning before that cast
+ // can compare different Field kinds and incorrectly discard a
matching row group.
+ continue;
+ }
+ const auto& chunk = row_group.columns[column_schema->leaf_column_id];
+ if (!chunk.__isset.meta_data ||
+ (chunk.meta_data.type != tparquet::Type::BYTE_ARRAY &&
+ chunk.meta_data.type != tparquet::Type::FIXED_LEN_BYTE_ARRAY) ||
+ !is_native_dictionary_encoded_chunk(chunk.meta_data)) {
+ continue;
+ }
+ std::unique_ptr<ParquetColumnReader> reader;
+ const std::vector<RowRange> ranges {{0, row_group.num_rows}};
+ const std::unordered_map<int, tparquet::OffsetIndex> offset_indexes;
+ const auto status = NativeColumnReader::create(
+ *column_schema, projection, file_context->native_file,
+ file_context->native_metadata, row_group_idx, ranges,
offset_indexes, timezone,
+ file_context->native_io_ctx, nullptr,
file_context->native_page_cache_enabled,
+ file_context->native_page_cache_file_key, true, {}, &reader);
+ if (!status.ok() || reader == nullptr) {
+ continue;
+ }
+ auto dictionary_result = reader->dictionary_values();
+ if (!dictionary_result.has_value()) {
+ continue;
+ }
+ auto dictionary = std::move(dictionary_result).value();
+ std::vector<Field> values(dictionary->size());
+ for (size_t value_idx = 0; value_idx < dictionary->size();
++value_idx) {
+ dictionary->get(value_idx, values[value_idx]);
+ }
+ DictionaryEvalContext ctx;
+ ctx.slots.emplace(slot_index, DictionaryEvalContext::SlotDictionary {
+ .data_type = column_schema->type,
+ .values = std::move(values),
+ });
+ if (VExprContext::evaluate_dictionary_filter(conjuncts, ctx) ==
+ ZoneMapFilterResult::kNoMatch) {
+ return ParquetRowGroupPruneReason::DICTIONARY;
+ }
+ }
+ return ParquetRowGroupPruneReason::NONE;
+}
+
+ParquetRowGroupPruneReason native_bloom_filter_prune_reason(
+ const tparquet::RowGroup& row_group,
+ const std::vector<std::unique_ptr<ParquetColumnSchema>>& file_schema,
+ const format::FileScanRequest& request, ParquetFileContext*
file_context,
+ ParquetPruningStats* pruning_stats) {
+ if (file_context == nullptr || file_context->native_file == nullptr) {
+ return ParquetRowGroupPruneReason::NONE;
+ }
+ const auto conjuncts_by_slot = collect_conjuncts_by_single_slot(
+ request.conjuncts, expr_zonemap::single_slot_bloom_filter_index);
+ for (const auto& [slot_index, conjuncts] : conjuncts_by_slot) {
+ const auto file_column_id = file_column_id_by_block_position(request,
slot_index);
+ if (!file_column_id.has_value()) {
+ continue;
+ }
+ const auto* column_schema = resolve_local_leaf_schema(file_schema,
*file_column_id);
+ if (column_schema == nullptr || column_schema->type == nullptr ||
+ !native_metadata_predicate_is_type_safe(*column_schema) ||
+ !bloom_filter_supported(*column_schema) ||
+ column_schema->leaf_column_id >=
static_cast<int>(row_group.columns.size())) {
+ continue;
+ }
+ const auto& chunk = row_group.columns[column_schema->leaf_column_id];
+ if (!chunk.__isset.meta_data) {
+ continue;
+ }
+ std::unique_ptr<native::BlockSplitBloomFilter> bloom_filter;
+ Status status;
+ {
+ int64_t timer_sink = 0;
+ SCOPED_RAW_TIMER(pruning_stats == nullptr ? &timer_sink
+ :
&pruning_stats->bloom_filter_read_time);
+ status = read_native_bloom_filter(chunk.meta_data,
file_context->native_file,
+ file_context->native_io_ctx,
&bloom_filter);
+ }
+ if (!status.ok() || bloom_filter == nullptr) {
+ continue;
+ }
+ BloomFilterEvalContext ctx;
+ ctx.slots.emplace(slot_index, BloomFilterEvalContext::SlotBloomFilter {
Review Comment:
[P1] Normalize Bloom probes to the Parquet physical type
This installs the raw split-block filter in `BloomFilterEvalContext`, so
equality/IN literals are hashed in their Doris logical representation. A valid
Parquet `UINT32` leaf is exposed as Doris `BIGINT`: `4,000,000,000` arrives as
eight bytes, while the file Bloom contains the four-byte INT32 carrier, so the
hashes differ and `kNoMatch` can discard the row group that contains the value.
The existing Arrow adapter converts logical integers back to the physical
representation. Reuse that descriptor-aware normalization for native pruning
and add a native row-group test with a present UINT32 above `INT32_MAX`.
##########
be/src/format_v2/parquet/reader/native/column_chunk_reader.cpp:
##########
@@ -0,0 +1,1296 @@
+// 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.
+
+#include "format_v2/parquet/reader/native/column_chunk_reader.h"
+
+#include <cctz/time_zone.h>
+#include <gen_cpp/parquet_types.h>
+#include <glog/logging.h>
+#include <parquet/metadata.h>
+#include <string.h>
+
+#include <algorithm>
+#include <cstdint>
+#include <limits>
+#include <memory>
+#include <utility>
+
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "core/column/column.h"
+#include "core/column/column_vector.h"
+#include "core/custom_allocator.h"
+#include "core/data_type_serde/data_type_serde.h"
+#include "core/data_type_serde/parquet_timestamp.h"
+#include "format_v2/parquet/native_schema_desc.h"
+#include "format_v2/parquet/reader/native/decoder.h"
+#include "format_v2/parquet/reader/native/level_decoder.h"
+#include "format_v2/parquet/reader/native/page_reader.h"
+#include "io/fs/buffered_reader.h"
+#include "runtime/runtime_profile.h"
+#include "storage/cache/page_cache.h"
+#include "util/bit_util.h"
+#include "util/block_compression.h"
+#include "util/unaligned.h"
+
+namespace cctz {
+class time_zone;
+} // namespace cctz
+namespace doris {
+namespace io {
+class BufferedStreamReader;
+struct IOContext;
+} // namespace io
+} // namespace doris
+
+namespace doris::format::parquet::native {
+
+bool can_prepare_page_cache_payload(bool session_cache_enabled, bool
storage_cache_disabled,
+ bool cache_available, bool
header_available) {
+ return session_cache_enabled && !storage_cache_disabled && cache_available
&& header_available;
+}
+
+ParquetReaderCompat parquet_reader_compat(const std::string& created_by) {
+ if (created_by.empty()) {
+ return {};
+ }
+ const ::parquet::ApplicationVersion version(created_by);
+ return {.parquet_816_padding =
+
version.VersionLt(::parquet::ApplicationVersion::PARQUET_816_FIXED_VERSION()),
+ .data_page_v2_always_compressed = version.VersionLt(
+
::parquet::ApplicationVersion::PARQUET_CPP_10353_FIXED_VERSION())};
+}
+
+Status compute_column_chunk_range(const tparquet::ColumnMetaData& metadata,
size_t file_size,
+ bool parquet_816_padding, ColumnChunkRange*
range) {
+ DORIS_CHECK(range != nullptr);
+ int64_t start = metadata.data_page_offset;
+ if (metadata.__isset.dictionary_page_offset &&
metadata.dictionary_page_offset > 0 &&
+ metadata.dictionary_page_offset < start) {
+ // Some writers use dictionary_page_offset=0 as an absence sentinel.
Range validation must
+ // follow has_dict_page() or the native reader starts at the Parquet
magic bytes.
+ start = metadata.dictionary_page_offset;
+ }
+ const int64_t length = metadata.total_compressed_size;
+ if (UNLIKELY(start < 0 || length < 0)) {
+ return Status::Corruption("Parquet column chunk has a negative offset
or length");
+ }
+ const uint64_t unsigned_start = static_cast<uint64_t>(start);
+ const uint64_t unsigned_length = static_cast<uint64_t>(length);
+ if (UNLIKELY(unsigned_start > file_size || unsigned_length > file_size -
unsigned_start)) {
+ // Thrift range fields are signed and untrusted; validate before
converting them to the
+ // unsigned stream-reader coordinates so overflow cannot wrap back
into the file.
+ return Status::Corruption("Parquet column chunk [{}, {}) exceeds file
size {}", start,
+ unsigned_start + unsigned_length, file_size);
+ }
+ size_t bounded_length = static_cast<size_t>(unsigned_length);
+ if (parquet_816_padding) {
+ // parquet-mr before PARQUET-816 under-reported the chunk by up to 100
bytes. Padding stays
+ // file-bounded and is only enabled for the affected writer versions.
+ bounded_length += std::min<size_t>(100, file_size - unsigned_start -
unsigned_length);
+ }
+ range->offset = static_cast<size_t>(unsigned_start);
+ range->length = bounded_length;
+ return Status::OK();
+}
+
+bool validate_offset_index(const tparquet::OffsetIndex& index, const
ColumnChunkRange& chunk_range,
+ int64_t data_page_offset, int64_t row_count) {
+ if (index.page_locations.empty() || data_page_offset < 0 || row_count < 0
||
+ index.page_locations.front().first_row_index != 0 ||
+ index.page_locations.front().offset != data_page_offset ||
+ chunk_range.length > std::numeric_limits<size_t>::max() -
chunk_range.offset) {
+ return false;
+ }
+ // Row indexes alone cannot detect a uniformly shifted OffsetIndex. Anchor
its first location
+ // to the owning metadata so page-to-row mapping cannot silently move by
one physical page.
+ const uint64_t chunk_begin = chunk_range.offset;
+ const uint64_t chunk_end = chunk_begin + chunk_range.length;
+ uint64_t previous_end = chunk_begin;
+ int64_t previous_row = -1;
+ for (const auto& location : index.page_locations) {
+ if (location.first_row_index <= previous_row ||
location.first_row_index >= row_count ||
+ location.offset < 0 || location.compressed_page_size <= 0) {
+ return false;
+ }
+ const uint64_t begin = static_cast<uint64_t>(location.offset);
+ const uint64_t size =
static_cast<uint64_t>(location.compressed_page_size);
+ if (begin < chunk_begin || begin < previous_end || begin > chunk_end ||
+ size > chunk_end - begin) {
+ return false;
+ }
+ previous_row = location.first_row_index;
+ previous_end = begin + size;
+ }
+ return true;
+}
+
+namespace {
+
+Status append_v2_int96_datetime(ColumnDateTimeV2::Container& data,
+ const ParquetInt96Timestamp& value,
+ const cctz::time_zone& timezone) {
+ static constexpr int64_t MICROS_PER_SECOND = 1000000LL;
+
+ int64_t micros = 0;
+ // Keep the fast ColumnDateTimeV2 path on the shared INT96 contract so
plain, dictionary, and
+ // sparse selections cannot materialize an invalid nanos-of-day that SerDe
would reject.
+ RETURN_IF_ERROR(parquet_int96_timestamp_micros(value, µs));
+ int64_t epoch_seconds = micros / MICROS_PER_SECOND;
+ int64_t micros_of_second = micros % MICROS_PER_SECOND;
+ if (micros_of_second < 0) {
+ micros_of_second += MICROS_PER_SECOND;
+ --epoch_seconds;
+ }
+ DateV2Value<DateTimeV2ValueType> datetime;
+ datetime.from_unixtime(epoch_seconds, timezone);
+ datetime.set_microsecond(static_cast<uint32_t>(micros_of_second));
+ if (!datetime.is_valid_date()) {
+ return Status::DataQualityError("Parquet INT96 timestamp is outside
the Doris range");
+ }
+ data.push_back(datetime);
+ return Status::OK();
+}
+
+class V2Int96DateTimeConsumer final : public ParquetFixedValueConsumer {
+public:
+ V2Int96DateTimeConsumer(IColumn& column, const ParquetDecodeContext&
context,
+ ParquetMaterializationState* state)
+ : _data(assert_cast<ColumnDateTimeV2&>(column).get_data()),
_state(state) {
+ static const auto utc = cctz::utc_time_zone();
+ _timezone = context.timezone == nullptr ? &utc : context.timezone;
+ }
+
+ Status consume(const uint8_t* values, size_t num_values, size_t
value_width) override {
+ DORIS_CHECK_EQ(value_width, sizeof(ParquetInt96Timestamp));
+ const size_t old_size = _data.size();
+ for (size_t row = 0; row < num_values; ++row) {
+ const auto value = unaligned_load<ParquetInt96Timestamp>(
+ values + row * sizeof(ParquetInt96Timestamp));
+ const auto status = append_v2_int96_datetime(_data, value,
*_timezone);
+ if (!status.ok()) {
+ if (_state != nullptr &&
_state->mark_conversion_failure(_data.size())) {
+ _data.emplace_back();
+ continue;
+ }
+ _data.resize(old_size);
+ return status;
+ }
+ }
+ return Status::OK();
+ }
+
+private:
+ ColumnDateTimeV2::Container& _data;
+ ParquetMaterializationState* _state;
+ const cctz::time_zone* _timezone = nullptr;
+};
+
+class RejectV2Int96BinaryConsumer final : public ParquetBinaryValueConsumer {
+public:
+ Status consume(const StringRef*, size_t) override {
+ return Status::NotSupported("INT96 cannot be decoded from binary
Parquet values");
+ }
+};
+
+Status read_v2_int96_datetime(IColumn& column, ParquetDecodeSource& source,
+ const ParquetDecodeContext& context, size_t
num_values,
+ ParquetMaterializationState& state) {
+ V2Int96DateTimeConsumer consumer(column, context, &state);
+ if (context.encoding != ParquetValueEncoding::DICTIONARY) {
+ return source.decode_fixed_values(num_values, consumer);
+ }
+ if (state.dictionary_generation != source.dictionary_generation()) {
+ state.typed_dictionary = column.clone_empty();
+ auto* output_null_map =
state.begin_dictionary_conversion(source.dictionary_size());
+ V2Int96DateTimeConsumer dictionary_consumer(*state.typed_dictionary,
context, &state);
+ RejectV2Int96BinaryConsumer binary_consumer;
+ const auto dictionary_status =
+ source.decode_dictionary(dictionary_consumer, binary_consumer);
+ state.end_dictionary_conversion(output_null_map);
+ RETURN_IF_ERROR(dictionary_status);
+ DORIS_CHECK_EQ(state.typed_dictionary->size(),
source.dictionary_size());
+ state.dictionary_generation = source.dictionary_generation();
+ }
+ RETURN_IF_ERROR(source.decode_dictionary_indices(num_values,
&state.dictionary_indices));
+ DORIS_CHECK_EQ(state.dictionary_indices.size(), num_values);
+ const size_t old_size = column.size();
+ column.insert_indices_from(*state.typed_dictionary,
state.dictionary_indices.data(),
+ state.dictionary_indices.data() + num_values);
+ if (state.can_insert_null_on_conversion_failure()) {
+ for (size_t row = 0; row < num_values; ++row) {
+ if (!state.dictionary_conversion_failures.empty() &&
+
state.dictionary_conversion_failures[state.dictionary_indices[row]] != 0) {
+ state.mark_conversion_failure(old_size + row);
+ }
+ }
+ }
+ return Status::OK();
+}
+
+Status read_native_or_serde(IColumn& column, const DataTypeSerDe& serde,
+ ParquetDecodeSource& source, const
ParquetDecodeContext& context,
+ size_t num_values, ParquetMaterializationState&
state) {
+ if (context.physical_type == ParquetPhysicalType::INT96 &&
+ check_and_get_column<ColumnDateTimeV2>(&column) != nullptr) {
+ return read_v2_int96_datetime(column, source, context, num_values,
state);
+ }
+ return serde.read_column_from_parquet(column, source, context, num_values,
state);
+}
+
+Status translate_value_encoding(tparquet::Encoding::type encoding,
+ ParquetValueEncoding* translated) {
+ DORIS_CHECK(translated != nullptr);
+ switch (encoding) {
+ case tparquet::Encoding::PLAIN:
+ *translated = ParquetValueEncoding::PLAIN;
+ return Status::OK();
+ case tparquet::Encoding::RLE_DICTIONARY:
+ case tparquet::Encoding::PLAIN_DICTIONARY:
+ *translated = ParquetValueEncoding::DICTIONARY;
+ return Status::OK();
+ case tparquet::Encoding::RLE:
+ *translated = ParquetValueEncoding::RLE;
+ return Status::OK();
+ case tparquet::Encoding::BIT_PACKED:
+ *translated = ParquetValueEncoding::BIT_PACKED;
+ return Status::OK();
+ case tparquet::Encoding::DELTA_BINARY_PACKED:
+ *translated = ParquetValueEncoding::DELTA_BINARY_PACKED;
+ return Status::OK();
+ case tparquet::Encoding::DELTA_LENGTH_BYTE_ARRAY:
+ *translated = ParquetValueEncoding::DELTA_LENGTH_BYTE_ARRAY;
+ return Status::OK();
+ case tparquet::Encoding::DELTA_BYTE_ARRAY:
+ *translated = ParquetValueEncoding::DELTA_BYTE_ARRAY;
+ return Status::OK();
+ case tparquet::Encoding::BYTE_STREAM_SPLIT:
+ *translated = ParquetValueEncoding::BYTE_STREAM_SPLIT;
+ return Status::OK();
+ default:
+ return Status::NotSupported("Unsupported Parquet encoding {}",
+ tparquet::to_string(encoding));
+ }
+}
+
+template <bool HAS_FILTER>
+Status decode_selected_values(IColumn& column, const DataTypeSerDe& serde,
Decoder& decoder,
+ const ParquetDecodeContext& context,
+ ParquetMaterializationState& state,
ColumnSelectVector& select_vector,
+ int64_t* materialization_time) {
+ SCOPED_RAW_TIMER(materialization_time);
+ ColumnSelectVector::DataReadType read_type;
+ while (const size_t run_length =
select_vector.get_next_run<HAS_FILTER>(&read_type)) {
+ switch (read_type) {
+ case ColumnSelectVector::CONTENT:
+ RETURN_IF_ERROR(
+ read_native_or_serde(column, serde, decoder, context,
run_length, state));
+ break;
+ case ColumnSelectVector::NULL_DATA:
+ column.insert_many_defaults(run_length);
+ break;
+ case ColumnSelectVector::FILTERED_CONTENT:
+ RETURN_IF_ERROR(decoder.skip_values(run_length));
+ break;
+ case ColumnSelectVector::FILTERED_NULL:
+ break;
+ }
+ }
+ return Status::OK();
+}
+
+// Presents one sparse page request as an ordinary sequential source to
DataTypeSerDe. SerDe is
+// entered once per page fragment; the concrete decoder decides whether to
gather selected spans,
+// batch-decode and compact, or use the cursor-preserving range fallback.
+class SelectedDecodeSource final : public ParquetDecodeSource {
+public:
+ SelectedDecodeSource(Decoder& decoder, const ParquetSelection& selection)
+ : _decoder(decoder), _selection(selection) {}
+
+ Status decode_fixed_values(size_t num_values, ParquetFixedValueConsumer&
consumer) override {
+ DORIS_CHECK_EQ(num_values, _selection.selected_values);
+ return _decoder.decode_selected_fixed_values(_selection, consumer);
+ }
+
+ Status decode_binary_values(size_t num_values, ParquetBinaryValueConsumer&
consumer) override {
+ DORIS_CHECK_EQ(num_values, _selection.selected_values);
+ return _decoder.decode_selected_binary_values(_selection, consumer);
+ }
+
+ Status skip_values(size_t num_values) override {
+ return Status::NotSupported("Selected Parquet source cannot be
skipped, values={}",
+ num_values);
+ }
+
+ bool has_dictionary() const override { return _decoder.has_dictionary(); }
+ uint64_t dictionary_generation() const override { return
_decoder.dictionary_generation(); }
+ size_t dictionary_size() const override { return
_decoder.dictionary_size(); }
+
+ Status decode_dictionary(ParquetFixedValueConsumer& fixed_consumer,
+ ParquetBinaryValueConsumer& binary_consumer)
override {
+ return _decoder.decode_dictionary(fixed_consumer, binary_consumer);
+ }
+
+ Status decode_dictionary_indices(size_t num_values, std::vector<uint32_t>*
indices) override {
+ DORIS_CHECK_EQ(num_values, _selection.selected_values);
+ return _decoder.decode_selected_dictionary_indices(_selection,
indices);
+ }
+
+private:
+ Decoder& _decoder;
+ const ParquetSelection& _selection;
+};
+
+Status decode_selected_non_null_values(IColumn& column, const DataTypeSerDe&
serde,
+ Decoder& decoder, const
ParquetDecodeContext& context,
+ ParquetMaterializationState& state,
+ ColumnSelectVector& select_vector,
+ int64_t* materialization_time) {
+ auto& selection = state.selection;
+ selection.ranges.clear();
+ selection.total_values = select_vector.num_values();
+ selection.selected_values = 0;
+
+ size_t cursor = 0;
+ ColumnSelectVector::DataReadType read_type;
+ while (const size_t run_length =
select_vector.get_next_run<true>(&read_type)) {
+ DORIS_CHECK(read_type == ColumnSelectVector::CONTENT ||
+ read_type == ColumnSelectVector::FILTERED_CONTENT);
+ if (read_type == ColumnSelectVector::CONTENT) {
+ selection.ranges.push_back({.first = cursor, .count = run_length});
+ selection.selected_values += run_length;
+ }
+ cursor += run_length;
+ }
+ DORIS_CHECK_EQ(cursor, selection.total_values);
+ if (selection.selected_values == 0) {
+ return decoder.skip_values(selection.total_values);
+ }
+
+ SCOPED_RAW_TIMER(materialization_time);
+ SelectedDecodeSource selected_source(decoder, selection);
+ return read_native_or_serde(column, serde, selected_source, context,
selection.selected_values,
+ state);
+}
+
+} // namespace
+
+template <bool IN_COLLECTION, bool OFFSET_INDEX>
+ColumnChunkReader<IN_COLLECTION, OFFSET_INDEX>::ColumnChunkReader(
+ io::BufferedStreamReader* reader, tparquet::ColumnChunk* column_chunk,
+ NativeFieldSchema* field_schema, const tparquet::OffsetIndex*
offset_index,
+ size_t total_rows, io::IOContext* io_ctx, const
ParquetPageReadContext& page_read_ctx,
+ const ColumnChunkRange* chunk_range)
+ : _field_schema(field_schema),
+ _max_rep_level(field_schema->repetition_level),
+ _max_def_level(field_schema->definition_level),
+ _stream_reader(reader),
+ _metadata(column_chunk->meta_data),
+ _offset_index(offset_index),
+ _total_rows(total_rows),
+ _io_ctx(io_ctx),
+ _page_read_ctx(page_read_ctx) {
+ if (chunk_range != nullptr) {
+ _chunk_range = *chunk_range;
+ _has_validated_chunk_range = true;
+ }
+}
+
+template <bool IN_COLLECTION, bool OFFSET_INDEX>
+Status ColumnChunkReader<IN_COLLECTION, OFFSET_INDEX>::init() {
+ size_t start_offset = _has_validated_chunk_range
+ ? _chunk_range.offset
+ : (has_dict_page(_metadata) ?
_metadata.dictionary_page_offset
+ :
_metadata.data_page_offset);
+ size_t chunk_size =
+ _has_validated_chunk_range ? _chunk_range.length :
_metadata.total_compressed_size;
+ // create page reader
+ _page_reader = create_page_reader<IN_COLLECTION, OFFSET_INDEX>(
+ _stream_reader, _io_ctx, start_offset, chunk_size, _total_rows,
_metadata,
+ _page_read_ctx, _offset_index);
+ // get the block compression codec
+ RETURN_IF_ERROR(get_block_compression_codec(_metadata.codec,
&_block_compress_codec));
+ _state = INITIALIZED;
+ RETURN_IF_ERROR(_parse_first_page_header());
+ return Status::OK();
+}
+
+template <bool IN_COLLECTION, bool OFFSET_INDEX>
+Status ColumnChunkReader<IN_COLLECTION, OFFSET_INDEX>::skip_nested_values(
+ const std::vector<level_t>& def_levels, size_t start_index) {
+ size_t no_value_cnt = 0;
+ size_t value_cnt = 0;
+
+ DORIS_CHECK(start_index <= def_levels.size());
+ for (size_t idx = start_index; idx < def_levels.size(); idx++) {
+ level_t def_level = def_levels[idx];
+ if (IN_COLLECTION && def_level <
_field_schema->repeated_parent_def_level) {
+ no_value_cnt++;
+ } else if (def_level < _field_schema->definition_level) {
+ no_value_cnt++;
+ } else {
+ value_cnt++;
+ }
+ }
+
+ RETURN_IF_ERROR(skip_values(value_cnt, true));
+ RETURN_IF_ERROR(skip_values(no_value_cnt, false));
+ return Status::OK();
+}
+
+template <bool IN_COLLECTION, bool OFFSET_INDEX>
+Status ColumnChunkReader<IN_COLLECTION, OFFSET_INDEX>::read_levels(
+ size_t num_values, std::vector<level_t>* rep_levels,
std::vector<level_t>* def_levels) {
+ DORIS_CHECK(rep_levels != nullptr);
+ DORIS_CHECK(def_levels != nullptr);
+ if (_remaining_num_values < num_values || _remaining_rep_nums < num_values
||
+ _remaining_def_nums < num_values) {
+ return Status::Corruption(
+ "Parquet level reader requested {} slots with only {}/{}/{}
remaining", num_values,
+ _remaining_num_values, _remaining_rep_nums,
_remaining_def_nums);
+ }
+
+ const size_t start_index = def_levels->size();
+ rep_levels->resize(rep_levels->size() + num_values, 0);
+ def_levels->resize(def_levels->size() + num_values, 0);
+ if (_max_rep_level > 0) {
+ const size_t decoded = _rep_level_decoder.get_levels(
+ rep_levels->data() + rep_levels->size() - num_values,
num_values);
+ if (decoded != num_values) {
+ return Status::Corruption("Parquet repetition level stream ended
after {} of {} slots",
+ decoded, num_values);
+ }
+ }
+ if (_max_def_level > 0) {
+ const size_t decoded = _def_level_decoder.get_levels(
+ def_levels->data() + def_levels->size() - num_values,
num_values);
+ if (decoded != num_values) {
+ return Status::Corruption("Parquet definition level stream ended
after {} of {} slots",
+ decoded, num_values);
+ }
+ }
+ _remaining_rep_nums -= num_values;
+ _remaining_def_nums -= num_values;
+ return skip_nested_values(*def_levels, start_index);
+}
+
+template <bool IN_COLLECTION, bool OFFSET_INDEX>
+Status ColumnChunkReader<IN_COLLECTION,
OFFSET_INDEX>::_parse_first_page_header() {
+ while (true) {
+ RETURN_IF_ERROR(_page_reader->parse_page_header());
+ const tparquet::PageHeader* header = nullptr;
+ RETURN_IF_ERROR(_page_reader->get_page_header(&header));
+ if (header->type == tparquet::PageType::DATA_PAGE ||
+ header->type == tparquet::PageType::DATA_PAGE_V2) {
+ _state = INITIALIZED;
+ return parse_page_header();
+ }
+ if (header->type != tparquet::PageType::DICTIONARY_PAGE) {
+ RETURN_IF_ERROR(_page_reader->skip_auxiliary_page());
+ _state = INITIALIZED;
+ continue;
+ }
+ // the first page maybe directory page even if
_metadata.__isset.dictionary_page_offset == false,
+ // so we should parse the directory page in next_page()
+ RETURN_IF_ERROR(_decode_dict_page());
+ // parse the real first data page
+ RETURN_IF_ERROR(_page_reader->dict_next_page());
+ _state = INITIALIZED;
+ // A dictionary is the only non-data page with decoder state. Any
following index or
+ // extension pages are skipped by the same pre-data loop.
+ }
+}
+
+template <bool IN_COLLECTION, bool OFFSET_INDEX>
+Status ColumnChunkReader<IN_COLLECTION, OFFSET_INDEX>::parse_page_header() {
+ if (_state == HEADER_PARSED || _state == DATA_LOADED) {
+ return Status::OK();
+ }
+ const tparquet::PageHeader* header = nullptr;
+ while (true) {
+ RETURN_IF_ERROR(_page_reader->parse_page_header());
+ RETURN_IF_ERROR(_page_reader->get_page_header(&header));
+ if (header->type == tparquet::PageType::DATA_PAGE ||
+ header->type == tparquet::PageType::DATA_PAGE_V2) {
+ break;
+ }
+ if (header->type == tparquet::PageType::DICTIONARY_PAGE) {
+ return Status::Corruption("Parquet dictionary page appears after
data pages");
+ }
+ RETURN_IF_ERROR(_page_reader->skip_auxiliary_page());
+ }
+ int32_t page_num_values = _page_reader->is_header_v2() ?
header->data_page_header_v2.num_values
+ :
header->data_page_header.num_values;
+ if (page_num_values < 0 || page_num_values > _metadata.num_values ||
+ (!OFFSET_INDEX &&
+ static_cast<uint64_t>(page_num_values) >
+ static_cast<uint64_t>(_metadata.num_values) -
_chunk_parsed_values)) {
+ // Page counts are untrusted and feed both level decoders and scratch
sizing. Bound each
+ // page by the column metadata before converting to unsigned counters.
+ return Status::Corruption("Parquet data page value count {} exceeds
column total {}",
+ page_num_values, _metadata.num_values);
+ }
+ if constexpr (!IN_COLLECTION) {
+ const size_t page_start_row = _page_reader->start_row();
+ const size_t page_end_row = _page_reader->end_row();
+ if (UNLIKELY(page_end_row < page_start_row ||
+ static_cast<size_t>(page_num_values) != page_end_row -
page_start_row)) {
+ // Flat columns have exactly one physical value slot per logical
row. Rejecting a
+ // divergent header/OffsetIndex span prevents every later page
from shifting rows.
+ return Status::Corruption(
+ "Parquet flat data page has {} values for logical row
range [{}, {})",
+ page_num_values, page_start_row, page_end_row);
+ }
+ }
+ _remaining_rep_nums = page_num_values;
+ _remaining_def_nums = page_num_values;
+ _remaining_num_values = page_num_values;
+
+ // no offset will parse all header.
+ if constexpr (OFFSET_INDEX == false) {
+ _chunk_parsed_values += _remaining_num_values;
+ }
+ _state = HEADER_PARSED;
+ return Status::OK();
+}
+
+template <bool IN_COLLECTION, bool OFFSET_INDEX>
+Status ColumnChunkReader<IN_COLLECTION, OFFSET_INDEX>::next_page() {
+ _state = INITIALIZED;
+ RETURN_IF_ERROR(_page_reader->next_page());
+ return Status::OK();
+}
+
+template <bool IN_COLLECTION, bool OFFSET_INDEX>
+Status ColumnChunkReader<IN_COLLECTION,
OFFSET_INDEX>::_get_uncompressed_levels(
+ const tparquet::DataPageHeaderV2& page_v2, Slice& page_data) {
+ const size_t rl = page_v2.repetition_levels_byte_length;
+ const size_t dl = page_v2.definition_levels_byte_length;
+ if (UNLIKELY(rl > page_data.size || dl > page_data.size - rl)) {
+ // Validate the physical slice again because a cached entry may itself
be truncated.
+ return Status::Corruption("Parquet data page v2 level bytes exceed
available payload");
+ }
+ _v2_rep_levels = Slice(page_data.data, rl);
+ _v2_def_levels = Slice(page_data.data + rl, dl);
+ page_data.data += dl + rl;
+ page_data.size -= dl + rl;
+ return Status::OK();
+}
+
+template <bool IN_COLLECTION, bool OFFSET_INDEX>
+Status ColumnChunkReader<IN_COLLECTION, OFFSET_INDEX>::load_page_data() {
+ if (_state == DATA_LOADED) {
+ return Status::OK();
+ }
+ if (UNLIKELY(_state != HEADER_PARSED)) {
+ return Status::Corruption("Should parse page header");
+ }
+
+ const tparquet::PageHeader* header = nullptr;
+ RETURN_IF_ERROR(_page_reader->get_page_header(&header));
+ int32_t uncompressed_size = header->uncompressed_page_size;
+ bool page_loaded = false;
+
+ // First, try to reuse a cache handle previously discovered by PageReader
+ // (header-only lookup) to avoid a second lookup here.
+ if (_page_read_ctx.enable_parquet_file_page_cache &&
!config::disable_storage_page_cache &&
+ StoragePageCache::instance() != nullptr) {
+ if (_page_reader->has_page_cache_handle()) {
+ const PageCacheHandle& handle = _page_reader->page_cache_handle();
+ Slice cached = handle.data();
+ size_t header_size = _page_reader->header_bytes().size();
+ size_t levels_size = 0;
+ if (header->__isset.data_page_header_v2) {
+ const tparquet::DataPageHeaderV2& header_v2 =
header->data_page_header_v2;
+ size_t rl = header_v2.repetition_levels_byte_length;
+ size_t dl = header_v2.definition_levels_byte_length;
+ levels_size = rl + dl;
+ if (UNLIKELY(header_size > cached.size ||
+ levels_size > cached.size - header_size)) {
+ return Status::Corruption("Cached Parquet page is shorter
than its v2 levels");
+ }
+ _v2_rep_levels =
+ Slice(reinterpret_cast<const uint8_t*>(cached.data) +
header_size, rl);
+ _v2_def_levels =
+ Slice(reinterpret_cast<const uint8_t*>(cached.data) +
header_size + rl, dl);
+ }
+ // payload_slice points to the bytes after header and levels
+ if (UNLIKELY(header_size + levels_size > cached.size)) {
+ return Status::Corruption("Cached Parquet page is shorter than
its header");
+ }
+ Slice payload_slice(cached.data + header_size + levels_size,
+ cached.size - header_size - levels_size);
+
+ bool cache_payload_is_decompressed =
_page_reader->is_cache_payload_decompressed();
+ const size_t expected_payload_size =
+ cache_payload_is_decompressed
+ ?
static_cast<size_t>(header->uncompressed_page_size) - levels_size
+ :
static_cast<size_t>(header->compressed_page_size) - levels_size;
+ if (UNLIKELY(payload_slice.size != expected_payload_size)) {
+ return Status::Corruption("Cached Parquet page payload has
size {}, expected {}",
+ payload_slice.size,
expected_payload_size);
+ }
+
+ if (cache_payload_is_decompressed) {
+ // Cached payload is already uncompressed
+ _page_data = payload_slice;
+ } else {
+ CHECK(_block_compress_codec);
+ // Decompress cached payload into _decompress_buf for decoding
+ size_t uncompressed_payload_size =
+ header->__isset.data_page_header_v2
+ ?
static_cast<size_t>(header->uncompressed_page_size) - levels_size
+ :
static_cast<size_t>(header->uncompressed_page_size);
+ _reserve_decompress_buf(uncompressed_payload_size);
+ _page_data = Slice(_decompress_buf.get(),
uncompressed_payload_size);
+ SCOPED_RAW_TIMER(&_chunk_statistics.decompress_time);
+ _chunk_statistics.decompress_cnt++;
+
RETURN_IF_ERROR(_block_compress_codec->decompress(payload_slice, &_page_data));
+ if (UNLIKELY(_page_data.size != uncompressed_payload_size)) {
+ return Status::Corruption("Parquet page decompressed to {}
bytes, expected {}",
+ _page_data.size,
uncompressed_payload_size);
+ }
+ }
+ // page cache counters were incremented when PageReader did the
header-only
+ // cache lookup. Do not increment again to avoid double-counting.
+ page_loaded = true;
+ }
+ }
+
+ if (!page_loaded) {
+ const bool prepare_cache_payload = can_prepare_page_cache_payload(
+ _page_read_ctx.enable_parquet_file_page_cache,
config::disable_storage_page_cache,
+ StoragePageCache::instance() != nullptr,
!_page_reader->header_bytes().empty());
+ if (_block_compress_codec != nullptr) {
+ Slice compressed_data;
+ RETURN_IF_ERROR(_page_reader->get_page_data(compressed_data));
+ std::vector<uint8_t> level_bytes;
+ if (header->__isset.data_page_header_v2) {
+ const tparquet::DataPageHeaderV2& header_v2 =
header->data_page_header_v2;
+ // uncompressed_size = rl + dl + uncompressed_data_size
+ // compressed_size = rl + dl + compressed_data_size
+ uncompressed_size -= header_v2.repetition_levels_byte_length +
+ header_v2.definition_levels_byte_length;
+ // copy level bytes (rl + dl) so that we can cache header +
levels + uncompressed payload
+ size_t rl = header_v2.repetition_levels_byte_length;
+ size_t dl = header_v2.definition_levels_byte_length;
+ size_t level_sz = rl + dl;
+ if (prepare_cache_payload && level_sz > 0) {
+ level_bytes.resize(level_sz);
+ memcpy(level_bytes.data(), compressed_data.data, level_sz);
+ }
+ // now remove levels from compressed_data for decompression
+ RETURN_IF_ERROR(_get_uncompressed_levels(header_v2,
compressed_data));
+ }
+ bool is_v2_compressed = header->__isset.data_page_header_v2 &&
+ (header->data_page_header_v2.is_compressed
||
+
_page_read_ctx.data_page_v2_always_compressed);
+ bool page_has_compression = header->__isset.data_page_header ||
is_v2_compressed;
+
+ if (page_has_compression) {
+ // Decompress payload for immediate decoding
+ _reserve_decompress_buf(uncompressed_size);
+ _page_data = Slice(_decompress_buf.get(), uncompressed_size);
+ SCOPED_RAW_TIMER(&_chunk_statistics.decompress_time);
+ _chunk_statistics.decompress_cnt++;
+
RETURN_IF_ERROR(_block_compress_codec->decompress(compressed_data,
&_page_data));
+ if (UNLIKELY(_page_data.size !=
static_cast<size_t>(uncompressed_size))) {
+ return Status::Corruption("Parquet page decompressed to {}
bytes, expected {}",
+ _page_data.size,
uncompressed_size);
+ }
+
+ // Decide whether to cache decompressed payload or compressed
payload based on threshold
+ bool cache_payload_decompressed = should_cache_decompressed(
+ header, _metadata,
_page_read_ctx.data_page_v2_always_compressed);
+
+ if (prepare_cache_payload) {
+ if (cache_payload_decompressed) {
+ _insert_page_into_cache(level_bytes, _page_data);
+
_chunk_statistics.page_cache_decompressed_write_counter += 1;
+ } else {
+ if (config::enable_parquet_cache_compressed_pages) {
+ // cache the compressed payload as-is (header |
levels | compressed_payload)
+ _insert_page_into_cache(
+ level_bytes, Slice(compressed_data.data,
compressed_data.size));
+
_chunk_statistics.page_cache_compressed_write_counter += 1;
+ }
+ }
+ }
+ } else {
+ // no compression on this page, use the data directly
+ _page_data = Slice(compressed_data.data, compressed_data.size);
Review Comment:
[P2] Reject contradictory sizes on uncompressed data pages
This branch treats the physical `compressed_page_size` slice as already
decompressed without checking that it matches `uncompressed_page_size`; the
codec-UNCOMPRESSED branch below does the same. For example, a one-value PLAIN
INT32 page with 8 physical bytes but an advertised uncompressed size of 4 is
accepted cold because the decoder consumes the first value and ignores the
padding. If page caching is enabled, that cold read stores all 8 bytes, while a
warm hit computes an expected decompressed payload of 4 bytes at lines 638-644
and rejects its own cache entry. Require equal sizes whenever the payload is
declared uncompressed (after the legacy V2 compatibility override), and cover
cold and warm V1/V2 mismatches.
--
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]
