airborne12 commented on code in PR #67977: URL: https://github.com/apache/doris/pull/67977#discussion_r4045651039
########## fe/fe-core/src/main/java/org/apache/doris/datasource/tvf/source/IndexDiskUsageScanNode.java: ########## @@ -0,0 +1,244 @@ +// 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.doris.datasource.tvf.source; + +import org.apache.doris.analysis.TupleDescriptor; +import org.apache.doris.catalog.DiskInfo; +import org.apache.doris.catalog.Env; +import org.apache.doris.catalog.Replica; +import org.apache.doris.catalog.Tablet; +import org.apache.doris.cloud.catalog.CloudReplica; +import org.apache.doris.cloud.system.CloudSystemInfoService; +import org.apache.doris.common.Config; +import org.apache.doris.common.UserException; +import org.apache.doris.planner.OlapScanNode; +import org.apache.doris.planner.PlanNodeId; +import org.apache.doris.planner.ScanContext; +import org.apache.doris.qe.ConnectContext; +import org.apache.doris.resource.computegroup.ComputeGroup; +import org.apache.doris.system.Backend; +import org.apache.doris.system.SystemInfoService; +import org.apache.doris.tablefunction.IndexDiskUsageTableValuedFunction; +import org.apache.doris.tablefunction.IndexDiskUsageTableValuedFunction.TabletTarget; +import org.apache.doris.thrift.TMetaScanRange; +import org.apache.doris.thrift.TNetworkAddress; +import org.apache.doris.thrift.TScanRange; +import org.apache.doris.thrift.TScanRangeLocation; +import org.apache.doris.thrift.TScanRangeLocations; + +import com.google.common.base.Preconditions; +import com.google.common.collect.Lists; +import com.google.common.collect.Maps; + +import java.util.List; +import java.util.Map; +import java.util.Set; +import java.util.function.LongFunction; +import java.util.function.Predicate; +import java.util.stream.Collectors; + +/** + * Scan node of index_disk_usage. Each tablet is read by one backend that holds it, and the tablets + * of a backend are split into a few scan ranges so that its scanners can read them in parallel. + */ +public class IndexDiskUsageScanNode extends MetadataScanNode { + + /** + * Picks the backend that reads one tablet. + */ + @FunctionalInterface + interface BackendSelector { + long select(TabletTarget target) throws UserException; + } + + private final IndexDiskUsageTableValuedFunction tvf; + private final List<TScanRangeLocations> scanRanges = Lists.newArrayList(); + + public IndexDiskUsageScanNode(PlanNodeId id, TupleDescriptor desc, IndexDiskUsageTableValuedFunction tvf, + ScanContext scanContext) { + super(id, desc, tvf, scanContext); + this.tvf = tvf; + } + + @Override + public void init() throws UserException { + super.init(); + SystemInfoService systemInfo = Env.getCurrentSystemInfo(); + ConnectContext context = ConnectContext.get(); + BackendSelector selector; + if (Config.isCloudMode()) { + selector = cloudSelector(systemInfo); + } else { + selector = localSelector(systemInfo::getBackend, + queryableIn(context == null ? null : context.getComputeGroupSafely())); + } + Map<Long, List<TabletTarget>> groups = groupByBackend(tvf.getTabletTargets(), selector); + TMetaScanRange template = tvf.getMetaScanRange(Lists.newArrayList()); + int rangesPerBackend = context == null ? 1 : context.getSessionVariable().getMaxScannersConcurrency(); + scanRanges.clear(); + scanRanges.addAll(buildScanRangeLocations(template, groups, systemInfo::getBackend, rangesPerBackend)); + numNodes = scanRanges.size(); + } + + @Override + protected void initBackendPolicy() { + // Tablet replicas decide where this scan runs, so the external file backend policy, which + // also requires load-available backends, does not apply. + } + + @Override + protected void createScanRangeLocations() { + // Scan ranges are built in init(), where replica selection can report a user error. + } + + @Override + public List<TScanRangeLocations> getScanRangeLocations(long maxScanRangeLength) { + return scanRanges; + } + + @Override + public int getNumInstances() { + return scanRanges.size(); + } + + static Map<Long, List<TabletTarget>> groupByBackend(List<TabletTarget> targets, BackendSelector selector) + throws UserException { + Map<Long, List<TabletTarget>> groups = Maps.newLinkedHashMap(); + for (TabletTarget target : targets) { + groups.computeIfAbsent(selector.select(target), backendId -> Lists.newArrayList()).add(target); + } + return groups; + } + + // Applies the replica rule of OLAP scans: a queryable mix node inside the caller's compute group. + static Predicate<Backend> queryableIn(ComputeGroup computeGroup) { + boolean invalidComputeGroup = ComputeGroup.INVALID_COMPUTE_GROUP.equals(computeGroup); + boolean notCloudComputeGroup = computeGroup != null && !Config.isCloudMode(); + return backend -> backend.isQueryAvailable() && backend.isMixNode() + && !OlapScanNode.shouldFilterReplicaByResourceTag(invalidComputeGroup, notCloudComputeGroup, + computeGroup, backend.getLocationTag().value); + } + + // Spreads tablets over their eligible backends by tablet id, so one backend does not read a + // whole table while the choice stays deterministic. + static long chooseBackend(long tabletId, List<Replica> replicas, LongFunction<Backend> backendLookup, + Predicate<Backend> eligible) throws UserException { + List<Long> candidates = replicas.stream() + .map(Replica::getBackendIdWithoutException) + .filter(backendId -> { + Backend backend = backendLookup.apply(backendId); + return backend != null && eligible.test(backend); + }) + .sorted() + .collect(Collectors.toList()); + if (candidates.isEmpty()) { + throw new UserException("No queryable replica for tablet " + tabletId); + } + return candidates.get((int) Math.floorMod(tabletId, (long) candidates.size())); + } + + // Splits the tablets of each backend into at most `rangesPerBackend` consecutive ranges, the + // way MetadataScanNode splits serialized splits by scanner concurrency. + static List<TScanRangeLocations> buildScanRangeLocations(TMetaScanRange template, + Map<Long, List<TabletTarget>> groups, LongFunction<Backend> backendLookup, int rangesPerBackend) { + Map<Long, String> partitionNames = template.getIndexDiskUsageParams().getPartitionNames(); + // Drop the table-wide lists once, so each range copy only carries its own tablets and partitions. + TMetaScanRange base = template.deepCopy(); + base.getIndexDiskUsageParams().unsetTablets(); + base.getIndexDiskUsageParams().unsetPartitionNames(); + List<TScanRangeLocations> ranges = Lists.newArrayList(); + for (Map.Entry<Long, List<TabletTarget>> group : groups.entrySet()) { + Backend backend = backendLookup.apply(group.getKey()); + Preconditions.checkState(backend != null, "backend %s is not found", group.getKey()); + List<TabletTarget> tablets = group.getValue(); + int chunkSize = (int) Math.ceil((double) tablets.size() / Math.max(1, rangesPerBackend)); + for (int from = 0; from < tablets.size(); from += chunkSize) { + List<TabletTarget> chunk = tablets.subList(from, Math.min(from + chunkSize, tablets.size())); + ranges.add(buildScanRange(base, chunk, partitionNames, backend)); + } + } + return ranges; + } + + private static TScanRangeLocations buildScanRange(TMetaScanRange base, List<TabletTarget> tablets, + Map<Long, String> partitionNames, Backend backend) { + TMetaScanRange metaScanRange = base.deepCopy(); + metaScanRange.getIndexDiskUsageParams().setTablets( + tablets.stream().map(TabletTarget::toThrift).collect(Collectors.toList())); + if (partitionNames != null) { + Map<Long, String> names = Maps.newHashMap(); + for (TabletTarget tablet : tablets) { + String name = partitionNames.get(tablet.getPartitionId()); + if (name != null) { + names.put(tablet.getPartitionId(), name); + } + } + metaScanRange.getIndexDiskUsageParams().setPartitionNames(names); + } + + TScanRange scanRange = new TScanRange(); + scanRange.setMetaScanRange(metaScanRange); + TScanRangeLocation location = new TScanRangeLocation(); + location.setBackendId(backend.getId()); + location.setServer(new TNetworkAddress(backend.getHost(), backend.getBePort())); + TScanRangeLocations locations = new TScanRangeLocations(); + locations.addToLocations(location); + locations.setScanRange(scanRange); + return locations; + } + + static BackendSelector localSelector(LongFunction<Backend> backendLookup, Predicate<Backend> eligible) { + // Tablets share backends, so the alive disks of each backend are collected once per scan. + Map<Long, Set<Long>> alivePathHashes = Maps.newHashMap(); + return target -> { + Tablet tablet = target.getTablet(); + for (Replica replica : tablet.getReplicas()) { + long backendId = replica.getBackendIdWithoutException(); + if (!alivePathHashes.containsKey(backendId)) { + Backend backend = backendLookup.apply(backendId); + if (backend != null) { + alivePathHashes.put(backendId, alivePathHashes(backend)); + } + } + } + List<Replica> replicas = tablet.getQueryableReplicas(target.getVersion(), alivePathHashes, false); + return chooseBackend(target.getTabletId(), replicas, backendLookup, eligible); + }; + } + + private static BackendSelector cloudSelector(SystemInfoService systemInfo) throws UserException { + String clusterId = ((CloudSystemInfoService) systemInfo).getCurrentClusterId(); + return target -> { + for (Replica replica : target.getTablet().getReplicas()) { + long backendId = ((CloudReplica) replica).getBackendIdWithClusterId(clusterId); + Backend backend = systemInfo.getBackend(backendId); + if (backend != null && backend.isQueryAvailable()) { Review Comment: Fixed in babe79b4af0. `cloudSelector` now fails the query when a replica routes to a backend marked as a smooth upgrade source, with the same kind of message other scans use for backends that cannot run them, instead of silently returning no rows for those tablets. Covered by `testCloudSelectorRejectsSmoothUpgradeSource` and `testCloudSelectorUsesQueryableBackend`. ########## be/src/format/table/index_disk_usage_reader.cpp: ########## @@ -0,0 +1,373 @@ +// 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/table/index_disk_usage_reader.h" + +#include <boost/algorithm/string/case_conv.hpp> +#include <shared_mutex> +#include <string> +#include <string_view> +#include <utility> +#include <variant> + +#include "cloud/cloud_tablet.h" +#include "common/cast_set.h" +#include "core/block/block.h" +#include "core/column/column_nullable.h" +#include "core/column/column_string.h" +#include "core/column/column_vector.h" +#include "runtime/exec_env.h" +#include "runtime/runtime_state.h" +#include "storage/rowset/rowset.h" +#include "storage/tablet/base_tablet.h" +#include "storage/tablet/tablet_schema.h" + +namespace doris { + +using segment_v2::IndexDiskUsageLevel; +using segment_v2::IndexDiskUsageRecord; +using segment_v2::IndexDiskUsageRow; +using segment_v2::IndexDiskUsageStructure; + +namespace { + +const std::vector<std::pair<std::string_view, IndexDiskUsageReader::Column>>& column_names() { + using C = IndexDiskUsageReader::Column; + static const std::vector<std::pair<std::string_view, C>> names = { + {"PARTITION_NAME", C::kPartitionName}, + {"MATERIALIZED_INDEX_NAME", C::kMaterializedIndexName}, + {"TABLET_ID", C::kTabletId}, + {"BACKEND_ID", C::kBackendId}, + {"ROWSET_ID", C::kRowsetId}, + {"SEGMENT_ID", C::kSegmentId}, + {"INDEX_ID", C::kIndexId}, + {"INDEX_NAME", C::kIndexName}, + {"INDEX_TYPE", C::kIndexType}, + {"COLUMN_NAME", C::kColumnName}, + {"INDEX_SUFFIX", C::kIndexSuffix}, + {"STRUCTURE", C::kStructure}, + {"STORAGE_FORMAT", C::kStorageFormat}, + {"SEGMENT_COUNT", C::kSegmentCount}, + {"ROW_COUNT", C::kRowCount}, + {"TOTAL_BYTES", C::kTotalBytes}, + {"DICT_BYTES", C::kDictBytes}, + {"POSTING_BYTES", C::kPostingBytes}, + {"POSITION_BYTES", C::kPositionBytes}, + {"STATS_BYTES", C::kStatsBytes}, + {"OTHER_BYTES", C::kOtherBytes}, + {"STATS_SOURCE", C::kStatsSource}, + }; + return names; +} + +Result<IndexDiskUsageReader::Column> column_of(const std::string& slot_name) { + const std::string upper = boost::to_upper_copy(slot_name); + for (const auto& [name, column] : column_names()) { + if (upper == name) { + return column; + } + } + return ResultError(Status::InternalError("unknown index_disk_usage column {}", slot_name)); +} + +Result<IndexDiskUsageLevel> parse_level(const std::string& level) { + if (level == "tablet") { + return IndexDiskUsageLevel::kTablet; + } + if (level == "rowset") { + return IndexDiskUsageLevel::kRowset; + } + if (level == "segment") { + return IndexDiskUsageLevel::kSegment; + } + return ResultError(Status::InvalidArgument("unsupported index_disk_usage level {}", level)); +} + +std::string_view structure_name(IndexDiskUsageStructure structure) { + switch (structure) { + case IndexDiskUsageStructure::kTerm: + return "TERM"; + case IndexDiskUsageStructure::kBkd: + return "BKD"; + case IndexDiskUsageStructure::kAnn: + return "ANN"; + case IndexDiskUsageStructure::kContainer: + return "CONTAINER"; + } + return "UNKNOWN"; +} + +void insert_null(IColumn* column) { + auto& nullable = reinterpret_cast<ColumnNullable&>(*column); + nullable.get_nested_column().insert_default(); + nullable.get_null_map_data().push_back(1); +} + +IColumn* non_null_nested(IColumn* column) { + auto& nullable = reinterpret_cast<ColumnNullable&>(*column); + nullable.get_null_map_data().push_back(0); + return nullable.get_nested_column_ptr().get(); +} + +void insert_int64(IColumn* column, int64_t value) { + assert_cast<ColumnInt64*>(non_null_nested(column))->insert_value(value); +} + +void insert_int32(IColumn* column, int32_t value) { + assert_cast<ColumnInt32*>(non_null_nested(column))->insert_value(value); +} + +void insert_string(IColumn* column, std::string_view value) { + assert_cast<ColumnString*>(non_null_nested(column))->insert_data(value.data(), value.size()); +} + +} // namespace + +IndexDiskUsageReader::IndexDiskUsageReader(std::vector<SlotDescriptor*> slots, RuntimeState* state, + RuntimeProfile* /*profile*/, TMetaScanRange scan_range) + : _state(state), _slots(std::move(slots)), _scan_range(std::move(scan_range)) {} + +Status IndexDiskUsageReader::init_reader() { + if (!_scan_range.__isset.index_disk_usage_params) { + return Status::InvalidArgument("index_disk_usage scan range has no parameters"); + } + const TIndexDiskUsageMetadataParams& params = _scan_range.index_disk_usage_params; + _level = DORIS_TRY(parse_level(params.level)); + _options.position_detail = params.position_detail; + _options.index_ids.insert(params.index_ids.begin(), params.index_ids.end()); + _options.check_cancelled = [state = _state]() { + RETURN_IF_CANCELLED(state); + return Status::OK(); + }; + _slot_columns.clear(); + for (const SlotDescriptor* slot : _slots) { + const Column column = DORIS_TRY(column_of(slot->col_name())); + _slot_columns.push_back(column); + } + return Status::OK(); +} + +Status IndexDiskUsageReader::_do_get_next_block(Block* block, size_t* read_rows, bool* eof) { + const auto& tablets = _scan_range.index_disk_usage_params.tablets; + *read_rows = 0; + while (_next_tablet < tablets.size()) { + RETURN_IF_CANCELLED(_state); + const TIndexDiskUsageTablet& target = tablets[_next_tablet++]; + std::vector<IndexDiskUsageRow> rows; + TabletSchemaSPtr current_schema; + RETURN_IF_ERROR(_collect_tablet(target, &rows, ¤t_schema)); + rows = segment_v2::aggregate_index_disk_usage(std::move(rows), _level); + if (rows.empty()) { + continue; + } + RETURN_IF_ERROR(_fill_block(block, target, *current_schema, rows)); + *read_rows = rows.size(); + *eof = false; + return Status::OK(); + } + *eof = true; + return Status::OK(); +} + +Status IndexDiskUsageReader::_collect_tablet(const TIndexDiskUsageTablet& target, + std::vector<IndexDiskUsageRow>* rows, + TabletSchemaSPtr* current_schema) const { + BaseTabletSPtr tablet = DORIS_TRY(ExecEnv::get_tablet(target.tablet_id)); + if (auto cloud_tablet = std::dynamic_pointer_cast<CloudTablet>(tablet)) { + SyncOptions options; Review Comment: Not changed. With `sync_delete_bitmap = false`, `CloudMetaMgr::sync_tablet_rowsets` still merges the new rowsets and advances the local version under the header lock, so a later MoW query on the same backend would treat its delete bitmap as current and skip syncing it, which can surface duplicate rows. No caller in `be/src` uses that combination today, and this scan follows the same sync path as an OLAP scan, so I keep the default. ########## be/src/storage/index/index_disk_usage.cpp: ########## @@ -0,0 +1,460 @@ +// 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 "storage/index/index_disk_usage.h" + +#include <algorithm> +#include <map> +#include <memory> +#include <tuple> +#include <utility> + +#include "common/cast_set.h" +#include "common/check.h" +#include "io/io_common.h" +#include "storage/index/ann/ann_index_files.h" +#include "storage/index/index_file_reader.h" +#include "storage/index/inverted/inverted_index_desc.h" +#include "storage/index/snii/format/dict_entry.h" +#include "storage/index/snii/format/format_constants.h" +#include "storage/index/snii/reader/logical_index_reader.h" +#include "storage/olap_common.h" +#include "storage/rowset/beta_rowset.h" +#include "storage/rowset/rowset.h" +#include "storage/rowset/rowset_meta.h" +#include "storage/segment/segment.h" + +namespace doris::segment_v2 { + +namespace { + +bool is_bkd_file(std::string_view name) { + return name == InvertedIndexDescriptor::get_temporary_bkd_index_data_file_name() || + name == InvertedIndexDescriptor::get_temporary_bkd_index_meta_file_name() || + name == InvertedIndexDescriptor::get_temporary_bkd_index_file_name(); +} + +bool is_ann_file(std::string_view name) { + return name == faiss_index_fila_name || name == faiss_ivfdata_file_name; +} + +bool is_wanted(const IndexDiskUsageOptions& options, int64_t index_id) { + return options.index_ids.empty() || options.index_ids.contains(index_id); +} + +Status check_cancelled(const IndexDiskUsageOptions& options) { + return options.check_cancelled ? options.check_cancelled() : Status::OK(); +} + +// A segment may have no index file on purpose, for example when every ANN index skipped a segment +// too small to train, or when a legacy table skipped writing indexes on load. Such a file holds +// no index bytes, while any other error is still reported. +bool is_absent_index_file(const Status& status) { + return status.is<ErrorCode::INVERTED_INDEX_FILE_NOT_FOUND>() || + status.is<ErrorCode::INVERTED_INDEX_BYPASS>() || status.is<ErrorCode::NOT_FOUND>(); +} + +// A V1 index file and its size persisted in the rowset meta, or -1 when the size is not recorded. +struct V1IndexFile { + const TabletIndex* index; + int64_t persisted_size; +}; + +// Lists the V1 index files of a segment. The rowset meta records every file, including the +// extracted VARIANT paths that the schema does not list; rowsets written without that record fall +// back to the schema indexes. `owned` keeps the indexes built from the rowset meta. +std::vector<V1IndexFile> list_v1_index_files(const TabletSchema& schema, + const InvertedIndexFileInfo& file_info, + std::vector<TabletIndex>* owned) { + std::vector<V1IndexFile> files; + if (file_info.index_info_size() == 0) { + for (const TabletIndex* index : schema.inverted_indexes()) { + files.push_back({.index = index, .persisted_size = -1}); + } + return files; + } + owned->reserve(file_info.index_info_size()); + for (const auto& index_info : file_info.index_info()) { + TabletIndexPB index_pb; + index_pb.set_index_type(IndexType::INVERTED); + index_pb.set_index_id(index_info.index_id()); + index_pb.set_index_suffix_name(index_info.index_suffix()); + owned->emplace_back().init_from_pb(index_pb); + files.push_back({.index = &owned->back(), + .persisted_size = index_info.index_file_size() > 0 + ? index_info.index_file_size() + : -1}); + } + return files; +} + +// Classifies every sub-file of a CLucene directory into `record` and adds their total length to +// `files_bytes`. +Status add_directory_files(const lucene::store::Directory& dir, IndexDiskUsageRecord* record, + int64_t* files_bytes) { + try { + std::vector<std::string> names; + if (!dir.list(&names)) { + return Status::Error<ErrorCode::INVERTED_INDEX_CLUCENE_ERROR>( + "failed to list inverted index sub-files"); + } + for (const auto& name : names) { + const int64_t length = dir.fileLength(name.c_str()); + classify_clucene_file(name, length, record); + *files_bytes += length; + } + } catch (CLuceneError& e) { + return Status::Error<ErrorCode::INVERTED_INDEX_CLUCENE_ERROR>( + "failed to read inverted index sub-files: {}", e.what()); + } + return Status::OK(); +} + +// Sums the position bytes of the dictionary entries whose postings live in the posting region. +// Inline postings stay in the dictionary region and are not counted. +Status sum_snii_position_bytes(const IndexFileReader& reader, uint64_t index_id, + std::string_view suffix, const IndexDiskUsageOptions& options, + int64_t* position_bytes) { + // A full dictionary scan should not evict blocks that queries keep in the file cache. + io::IOContext io_ctx; + io_ctx.is_disposable = true; + io_ctx.is_inverted_index = true; + auto logical = DORIS_TRY(reader.open_snii_logical_index( + index_id, suffix, &io_ctx, snii::reader::LogicalIndexOpenMode::kCompaction)); + snii_doris::DorisSniiFileReader::ScopedIOContext io_context_scope(&io_ctx); + std::vector<snii::format::DictEntry> entries; + for (uint32_t block = 0; block < logical->n_dict_blocks(); ++block) { + RETURN_IF_ERROR(check_cancelled(options)); + uint64_t frq_base = 0; + uint64_t prx_base = 0; + RETURN_IF_ERROR(logical->decode_dict_block(block, &entries, &frq_base, &prx_base)); + for (const auto& entry : entries) { + if (entry.kind == snii::format::DictEntryKind::kPodRef) { + *position_bytes += cast_set<int64_t>(entry.prx_len); + } + } + } + return Status::OK(); +} + +int64_t merge_component(int64_t lhs, int64_t rhs) { + return lhs < 0 || rhs < 0 ? -1 : lhs + rhs; +} + +} // namespace + +std::vector<IndexDiskUsageRow> aggregate_index_disk_usage(std::vector<IndexDiskUsageRow> rows, + IndexDiskUsageLevel level) { + if (level == IndexDiskUsageLevel::kSegment) { + return rows; + } + using Key = std::tuple<std::string, int64_t, std::string, int, int>; + std::map<Key, size_t> positions; + std::vector<IndexDiskUsageRow> merged; + for (auto& row : rows) { + row.segment_id = -1; + if (level == IndexDiskUsageLevel::kTablet) { + row.rowset_id.clear(); + } + Key key {row.rowset_id, row.record.index_id, row.record.index_suffix, + static_cast<int>(row.format), static_cast<int>(row.record.structure)}; + auto [it, inserted] = positions.emplace(std::move(key), merged.size()); + if (inserted) { + merged.push_back(std::move(row)); + continue; + } + IndexDiskUsageRow& target = merged[it->second]; + target.segment_count += row.segment_count; + target.row_count += row.row_count; + IndexDiskUsageRecord& dst = target.record; + const IndexDiskUsageRecord& src = row.record; + dst.total_bytes += src.total_bytes; + dst.dict_bytes = merge_component(dst.dict_bytes, src.dict_bytes); + dst.posting_bytes = merge_component(dst.posting_bytes, src.posting_bytes); + dst.position_bytes = merge_component(dst.position_bytes, src.position_bytes); + dst.stats_bytes = merge_component(dst.stats_bytes, src.stats_bytes); + dst.other_bytes = merge_component(dst.other_bytes, src.other_bytes); + } + return merged; +} + +void classify_clucene_file(std::string_view name, int64_t length, IndexDiskUsageRecord* record) { + record->total_bytes += length; + if (is_bkd_file(name)) { + record->structure = IndexDiskUsageStructure::kBkd; + return; + } + if (is_ann_file(name)) { + record->structure = IndexDiskUsageStructure::kAnn; + return; + } + const size_t dot = name.rfind('.'); + const std::string_view extension = + dot == std::string_view::npos ? std::string_view() : name.substr(dot + 1); + if (extension == "tis" || extension == "tii") { + record->dict_bytes += length; + } else if (extension == "frq") { + record->posting_bytes += length; + } else if (extension == "prx") { + record->position_bytes += length; + } else if (extension == "nrm") { + record->stats_bytes += length; + } else { + record->other_bytes += length; + } +} + +IndexDiskUsageCollector::IndexDiskUsageCollector(io::FileSystemSPtr fs, + std::string index_path_prefix, + TabletSchemaSPtr schema, + InvertedIndexStorageFormatPB format, + int64_t tablet_id, + InvertedIndexFileInfo index_file_info) + : _fs(std::move(fs)), + _index_path_prefix(std::move(index_path_prefix)), + _schema(std::move(schema)), + _format(format), + _tablet_id(tablet_id), + _index_file_info(std::move(index_file_info)) {} + +Status IndexDiskUsageCollector::collect(const IndexDiskUsageOptions& options, + std::vector<IndexDiskUsageRecord>* out) { + DORIS_CHECK(out != nullptr); + // A rowset returns no file system when its tablet or storage resource cannot be resolved. + if (_fs == nullptr) { + return Status::Error<ErrorCode::INIT_FAILED>("no file system for inverted index files {}", + _index_path_prefix); + } + switch (_format) { + case InvertedIndexStorageFormatPB::V1: + return _collect_v1(options, out); + case InvertedIndexStorageFormatPB::V2: + case InvertedIndexStorageFormatPB::V3: + return _collect_compound(options, out); + case InvertedIndexStorageFormatPB::SNII: + return _collect_snii(options, out); + default: + return Status::NotSupported("index disk usage does not support inverted index format {}", + InvertedIndexStorageFormatPB_Name(_format)); + } +} + +Status IndexDiskUsageCollector::_collect_v1(const IndexDiskUsageOptions& options, + std::vector<IndexDiskUsageRecord>* out) { + IndexFileReader reader(_fs, _index_path_prefix, _format, _index_file_info, _tablet_id); + RETURN_IF_ERROR(reader.init()); + std::vector<TabletIndex> file_indexes; + for (const V1IndexFile& file : list_v1_index_files(*_schema, _index_file_info, &file_indexes)) { + const TabletIndex& index = *file.index; + if (!is_wanted(options, index.index_id())) { + continue; + } + RETURN_IF_ERROR(check_cancelled(options)); + int64_t file_size = file.persisted_size; + if (file_size < 0) { + const std::string path = InvertedIndexDescriptor::get_index_file_path_v1( + _index_path_prefix, index.index_id(), index.get_index_suffix()); + const Status size_status = _fs->file_size(path, &file_size); + if (is_absent_index_file(size_status)) { + continue; + } + RETURN_IF_ERROR(size_status); + } + auto directory = reader.open(&index); + if (!directory.has_value()) { + if (is_absent_index_file(directory.error())) { + continue; + } + return directory.error(); + } + + IndexDiskUsageRecord record; + record.index_id = index.index_id(); + record.index_suffix = index.get_index_suffix(); + int64_t files_bytes = 0; + RETURN_IF_ERROR(add_directory_files(*directory.value(), &record, &files_bytes)); + // Each V1 index owns its file, so its compound header is part of the index. + record.other_bytes += file_size - files_bytes; + record.total_bytes += file_size - files_bytes; + out->push_back(std::move(record)); + } + return Status::OK(); +} + +Status IndexDiskUsageCollector::_collect_compound(const IndexDiskUsageOptions& options, + std::vector<IndexDiskUsageRecord>* out) { + IndexFileReader reader(_fs, _index_path_prefix, _format, _index_file_info, _tablet_id); + if (const Status st = reader.init(); !st.ok()) { + return is_absent_index_file(st) ? Status::OK() : st; + } + auto directories = DORIS_TRY(reader.get_all_directories()); + + // Every index counts toward the attributed bytes, even the filtered ones, so the container + // record only holds the shared header. + int64_t attributed_bytes = 0; + for (const auto& [key, directory] : directories) { + IndexDiskUsageRecord record; + record.index_id = key.first; + record.index_suffix = key.second; + RETURN_IF_ERROR(add_directory_files(*directory, &record, &attributed_bytes)); + if (is_wanted(options, record.index_id)) { + out->push_back(std::move(record)); + } + } + if (options.index_ids.empty()) { + IndexDiskUsageRecord container; + container.structure = IndexDiskUsageStructure::kContainer; + container.total_bytes = reader.get_inverted_file_size() - attributed_bytes; + container.other_bytes = container.total_bytes; + out->push_back(std::move(container)); + } + return Status::OK(); +} + +Status IndexDiskUsageCollector::_collect_snii(const IndexDiskUsageOptions& options, + std::vector<IndexDiskUsageRecord>* out) { + IndexFileReader reader(_fs, _index_path_prefix, _format, _index_file_info, _tablet_id); + if (const Status st = reader.init(); !st.ok()) { + return is_absent_index_file(st) ? Status::OK() : st; + } + const auto entries = DORIS_TRY(reader.snii_logical_indexes()); + + int64_t attributed_bytes = 0; + for (const auto& entry : entries) { + const auto index_id = cast_set<int64_t>(entry.index_id); + // The container record is only reported without a filter, so filtered-out indexes + // need no metadata read. + if (!is_wanted(options, index_id)) { + continue; + } + RETURN_IF_ERROR(check_cancelled(options)); + IndexDiskUsageRecord record; + record.index_id = index_id; + record.index_suffix = entry.index_suffix; + if (entry.kind == snii::format::LogicalIndexKind::kInverted) { + snii::format::CoreMetadata core; + RETURN_IF_ERROR(reader.snii_core_metadata(entry.index_id, entry.index_suffix, &core)); + const auto& refs = core.section_refs; + record.structure = IndexDiskUsageStructure::kTerm; + record.dict_bytes = + cast_set<int64_t>(refs.dict_region.length + entry.sampled_term_index.length + + entry.dict_block_directory.length + refs.bsbf.length); + record.posting_bytes = cast_set<int64_t>(refs.posting_region.length); + record.stats_bytes = cast_set<int64_t>(entry.core_metadata.length + refs.norms.length); + record.other_bytes = cast_set<int64_t>(refs.null_bitmap.length); + if (!snii::format::has_positions(core.index_config)) { + record.position_bytes = 0; + } else if (options.position_detail) { + int64_t position_bytes = 0; + RETURN_IF_ERROR(sum_snii_position_bytes(reader, entry.index_id, entry.index_suffix, + options, &position_bytes)); + record.position_bytes = position_bytes; + record.posting_bytes -= position_bytes; + } else { + record.position_bytes = -1; + } + record.total_bytes = record.dict_bytes + record.posting_bytes + + std::max<int64_t>(record.position_bytes, 0) + record.stats_bytes + + record.other_bytes; + } else { + record.structure = entry.kind == snii::format::LogicalIndexKind::kBkd + ? IndexDiskUsageStructure::kBkd + : IndexDiskUsageStructure::kAnn; + for (const auto& file : entry.files) { + record.total_bytes += cast_set<int64_t>(file.length); Review Comment: Not changed in this PR. Overlapping blob extents in a CRC-valid directory are a corrupt container case, and the same directory would mislead the queries that open those blobs, so the extent validation belongs in `SniiSegmentReader` for every reader rather than in this accounting path. I keep it with the other corrupt-header findings for a separate reader-side hardening change. ########## be/src/format/table/index_disk_usage_reader.cpp: ########## @@ -0,0 +1,373 @@ +// 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/table/index_disk_usage_reader.h" + +#include <boost/algorithm/string/case_conv.hpp> +#include <shared_mutex> +#include <string> +#include <string_view> +#include <utility> +#include <variant> + +#include "cloud/cloud_tablet.h" +#include "common/cast_set.h" +#include "core/block/block.h" +#include "core/column/column_nullable.h" +#include "core/column/column_string.h" +#include "core/column/column_vector.h" +#include "runtime/exec_env.h" +#include "runtime/runtime_state.h" +#include "storage/rowset/rowset.h" +#include "storage/tablet/base_tablet.h" +#include "storage/tablet/tablet_schema.h" + +namespace doris { + +using segment_v2::IndexDiskUsageLevel; +using segment_v2::IndexDiskUsageRecord; +using segment_v2::IndexDiskUsageRow; +using segment_v2::IndexDiskUsageStructure; + +namespace { + +const std::vector<std::pair<std::string_view, IndexDiskUsageReader::Column>>& column_names() { + using C = IndexDiskUsageReader::Column; + static const std::vector<std::pair<std::string_view, C>> names = { + {"PARTITION_NAME", C::kPartitionName}, + {"MATERIALIZED_INDEX_NAME", C::kMaterializedIndexName}, + {"TABLET_ID", C::kTabletId}, + {"BACKEND_ID", C::kBackendId}, + {"ROWSET_ID", C::kRowsetId}, + {"SEGMENT_ID", C::kSegmentId}, + {"INDEX_ID", C::kIndexId}, + {"INDEX_NAME", C::kIndexName}, + {"INDEX_TYPE", C::kIndexType}, + {"COLUMN_NAME", C::kColumnName}, + {"INDEX_SUFFIX", C::kIndexSuffix}, + {"STRUCTURE", C::kStructure}, + {"STORAGE_FORMAT", C::kStorageFormat}, + {"SEGMENT_COUNT", C::kSegmentCount}, + {"ROW_COUNT", C::kRowCount}, + {"TOTAL_BYTES", C::kTotalBytes}, + {"DICT_BYTES", C::kDictBytes}, + {"POSTING_BYTES", C::kPostingBytes}, + {"POSITION_BYTES", C::kPositionBytes}, + {"STATS_BYTES", C::kStatsBytes}, + {"OTHER_BYTES", C::kOtherBytes}, + {"STATS_SOURCE", C::kStatsSource}, + }; + return names; +} + +Result<IndexDiskUsageReader::Column> column_of(const std::string& slot_name) { + const std::string upper = boost::to_upper_copy(slot_name); + for (const auto& [name, column] : column_names()) { + if (upper == name) { + return column; + } + } + return ResultError(Status::InternalError("unknown index_disk_usage column {}", slot_name)); +} + +Result<IndexDiskUsageLevel> parse_level(const std::string& level) { + if (level == "tablet") { + return IndexDiskUsageLevel::kTablet; + } + if (level == "rowset") { + return IndexDiskUsageLevel::kRowset; + } + if (level == "segment") { + return IndexDiskUsageLevel::kSegment; + } + return ResultError(Status::InvalidArgument("unsupported index_disk_usage level {}", level)); +} + +std::string_view structure_name(IndexDiskUsageStructure structure) { + switch (structure) { + case IndexDiskUsageStructure::kTerm: + return "TERM"; + case IndexDiskUsageStructure::kBkd: + return "BKD"; + case IndexDiskUsageStructure::kAnn: + return "ANN"; + case IndexDiskUsageStructure::kContainer: + return "CONTAINER"; + } + return "UNKNOWN"; +} + +void insert_null(IColumn* column) { + auto& nullable = reinterpret_cast<ColumnNullable&>(*column); + nullable.get_nested_column().insert_default(); + nullable.get_null_map_data().push_back(1); +} + +IColumn* non_null_nested(IColumn* column) { + auto& nullable = reinterpret_cast<ColumnNullable&>(*column); + nullable.get_null_map_data().push_back(0); + return nullable.get_nested_column_ptr().get(); +} + +void insert_int64(IColumn* column, int64_t value) { + assert_cast<ColumnInt64*>(non_null_nested(column))->insert_value(value); +} + +void insert_int32(IColumn* column, int32_t value) { + assert_cast<ColumnInt32*>(non_null_nested(column))->insert_value(value); +} + +void insert_string(IColumn* column, std::string_view value) { + assert_cast<ColumnString*>(non_null_nested(column))->insert_data(value.data(), value.size()); +} + +} // namespace + +IndexDiskUsageReader::IndexDiskUsageReader(std::vector<SlotDescriptor*> slots, RuntimeState* state, + RuntimeProfile* /*profile*/, TMetaScanRange scan_range) + : _state(state), _slots(std::move(slots)), _scan_range(std::move(scan_range)) {} + +Status IndexDiskUsageReader::init_reader() { + if (!_scan_range.__isset.index_disk_usage_params) { + return Status::InvalidArgument("index_disk_usage scan range has no parameters"); + } + const TIndexDiskUsageMetadataParams& params = _scan_range.index_disk_usage_params; + _level = DORIS_TRY(parse_level(params.level)); + _options.position_detail = params.position_detail; + _options.index_ids.insert(params.index_ids.begin(), params.index_ids.end()); + _options.check_cancelled = [state = _state]() { + RETURN_IF_CANCELLED(state); + return Status::OK(); + }; + _slot_columns.clear(); + for (const SlotDescriptor* slot : _slots) { + const Column column = DORIS_TRY(column_of(slot->col_name())); + _slot_columns.push_back(column); + } + return Status::OK(); +} + +Status IndexDiskUsageReader::_do_get_next_block(Block* block, size_t* read_rows, bool* eof) { + const auto& tablets = _scan_range.index_disk_usage_params.tablets; + *read_rows = 0; + while (_next_tablet < tablets.size()) { + RETURN_IF_CANCELLED(_state); + const TIndexDiskUsageTablet& target = tablets[_next_tablet++]; + std::vector<IndexDiskUsageRow> rows; + TabletSchemaSPtr current_schema; + RETURN_IF_ERROR(_collect_tablet(target, &rows, ¤t_schema)); + rows = segment_v2::aggregate_index_disk_usage(std::move(rows), _level); + if (rows.empty()) { + continue; + } + RETURN_IF_ERROR(_fill_block(block, target, *current_schema, rows)); + *read_rows = rows.size(); + *eof = false; + return Status::OK(); + } + *eof = true; + return Status::OK(); +} + +Status IndexDiskUsageReader::_collect_tablet(const TIndexDiskUsageTablet& target, + std::vector<IndexDiskUsageRow>* rows, + TabletSchemaSPtr* current_schema) const { + BaseTabletSPtr tablet = DORIS_TRY(ExecEnv::get_tablet(target.tablet_id)); + if (auto cloud_tablet = std::dynamic_pointer_cast<CloudTablet>(tablet)) { + SyncOptions options; + options.query_version = target.version; + RETURN_IF_ERROR(cloud_tablet->sync_rowsets(options)); Review Comment: Not changed. The cold `get_tablet` lookup and the versioned `sync_rowsets` are the same calls an OLAP scan makes in `OlapScanLocalState` (`bthread_fork_join` over its scan ranges), and neither receives the query cancellation signal there either. Making `CloudMetaMgr::sync_tablet_rowsets_unlocked` cancellable changes the shared cloud sync path and is out of scope for this TVF. -- 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]
