Storages: Fix cloning delta index when there are duplicated tuples (#9000)

dbms/src/Storages/DeltaMerge/Delta/DeltaValueSpace.h

@@ -454,7 +454,8 @@ class DeltaValueReader
 
     bool shouldPlace(
         const DMContext & context,
-        DeltaIndexPtr my_delta_index,
+        size_t placed_rows,
+        size_t placed_delete_ranges,
         const RowKeyRange & segment_range,
         const RowKeyRange & relevant_range,
         UInt64 max_version);

dbms/src/Storages/DeltaMerge/Delta/Snapshot.cpp

@@ -121,15 +121,15 @@ size_t DeltaValueReader::readRows(
     //
     // So here, we should filter out those out-of-range rows.
 
-    auto mem_table_rows_offset = delta_snap->getMemTableSetRowsOffset();
-    auto total_delta_rows = delta_snap->getRows();
+    const auto mem_table_rows_offset = delta_snap->getMemTableSetRowsOffset();
+    const auto total_delta_rows = delta_snap->getRows();
 
-    auto persisted_files_start = std::min(offset, mem_table_rows_offset);
-    auto persisted_files_end = std::min(offset + limit, mem_table_rows_offset);
-    auto mem_table_start = offset <= mem_table_rows_offset
+    const auto persisted_files_start = std::min(offset, mem_table_rows_offset);
+    const auto persisted_files_end = std::min(offset + limit, mem_table_rows_offset);
+    const auto mem_table_start = offset <= mem_table_rows_offset
         ? 0
         : std::min(offset - mem_table_rows_offset, total_delta_rows - mem_table_rows_offset);
-    auto mem_table_end = offset + limit <= mem_table_rows_offset
+    const auto mem_table_end = offset + limit <= mem_table_rows_offset
         ? 0
         : std::min(offset + limit - mem_table_rows_offset, total_delta_rows - mem_table_rows_offset);
 
@@ -147,8 +147,12 @@ size_t DeltaValueReader::readRows(
     }
     if (mem_table_start < mem_table_end)
     {
-        actual_read += mem_table_reader
-                           ->readRows(output_cols, mem_table_start, mem_table_end - mem_table_start, range, row_ids);
+        actual_read += mem_table_reader->readRows( //
+            output_cols,
+            mem_table_start,
+            mem_table_end - mem_table_start,
+            range,
+            row_ids);
     }
 
     if (row_ids != nullptr)
@@ -213,14 +217,13 @@ BlockOrDeletes DeltaValueReader::getPlaceItems(
 
 bool DeltaValueReader::shouldPlace(
     const DMContext & context,
-    DeltaIndexPtr my_delta_index,
+    const size_t placed_rows,
+    const size_t placed_delete_ranges,
     const RowKeyRange & segment_range_,
     const RowKeyRange & relevant_range,
     UInt64 max_version)
 {
-    auto [placed_rows, placed_delete_ranges] = my_delta_index->getPlacedStatus();
-
-    // Already placed.
+    // The placed_rows, placed_delete_range already contains the data in delta_snap
     if (placed_rows >= delta_snap->getRows() && placed_delete_ranges == delta_snap->getDeletes())
         return false;
 

dbms/src/Storages/DeltaMerge/DeltaIndex.h

@@ -86,17 +86,17 @@ class DeltaIndex
         }
     }
 
-    DeltaIndexPtr tryCloneInner(size_t placed_deletes_limit, const Updates * updates = nullptr)
+    DeltaIndexPtr tryCloneInner(size_t rows_limit, size_t placed_deletes_limit, const Updates * updates = nullptr)
     {
         DeltaTreePtr delta_tree_copy;
         size_t placed_rows_copy = 0;
         size_t placed_deletes_copy = 0;
-        // Make sure the delta index do not place more deletes than `placed_deletes_limit`.
-        // Because delete ranges can break MVCC view.
         {
             std::scoped_lock lock(mutex);
-            // Safe to reuse the copy of the existing DeltaIndex
-            if (placed_deletes <= placed_deletes_limit)
+            // Make sure the MVCC view will not be broken by the mismatch of delta index and snapshot:
+            // - First, make sure the delta index do not place more deletes than `placed_deletes_limit`.
+            // - Second, make sure the snapshot includes all duplicated tuples in the delta index.
+            if (placed_deletes <= placed_deletes_limit && delta_tree->maxDupTupleID() < static_cast<Int64>(rows_limit))
             {
                 delta_tree_copy = delta_tree;
                 placed_rows_copy = placed_rows;
@@ -195,8 +195,9 @@ class DeltaIndex
     {
         std::scoped_lock lock(mutex);
 
-        if ((maybe_advanced.placed_rows >= placed_rows && maybe_advanced.placed_deletes >= placed_deletes)
-            && !(maybe_advanced.placed_rows == placed_rows && maybe_advanced.placed_deletes == placed_deletes))
+        if ((maybe_advanced.placed_rows >= placed_rows && maybe_advanced.placed_deletes >= placed_deletes) // advance
+            // not excatly the same
+            && (maybe_advanced.placed_rows != placed_rows || maybe_advanced.placed_deletes != placed_deletes))
         {
             delta_tree = maybe_advanced.delta_tree;
             placed_rows = maybe_advanced.placed_rows;
@@ -206,14 +207,17 @@ class DeltaIndex
         return false;
     }
 
-    DeltaIndexPtr tryClone(size_t /*rows*/, size_t deletes) { return tryCloneInner(deletes); }
+    /**
+     * Try to get a clone of current instance.
+     * Return an empty DeltaIndex if `deletes < this->placed_deletes` because the advanced delta-index will break
+     * the MVCC view.
+     */
+    DeltaIndexPtr tryClone(size_t rows, size_t deletes) { return tryCloneInner(rows, deletes); }
 
     DeltaIndexPtr cloneWithUpdates(const Updates & updates)
     {
-        if (unlikely(updates.empty()))
-            throw Exception("Unexpected empty updates");
-
-        return tryCloneInner(updates.front().delete_ranges_offset, &updates);
+        RUNTIME_CHECK_MSG(!updates.empty(), "Unexpected empty updates");
+        return tryCloneInner(updates.front().rows_offset, updates.front().delete_ranges_offset, &updates);
     }
 
     const std::optional<Remote::RNDeltaIndexCache::CacheKey> & getRNCacheKey() const { return rn_cache_key; }

dbms/src/Storages/DeltaMerge/DeltaPlace.h

@@ -262,7 +262,10 @@ bool placeInsert(
             tuple_id = delta_value_space_offset + (offset + i);
 
         if (dup)
+        {
             delta_tree.addDelete(rid);
+            delta_tree.setMaxDupTupleID(tuple_id);
+        }
         delta_tree.addInsert(rid, tuple_id);
     }
 

dbms/src/Storages/DeltaMerge/DeltaTree.h

@@ -826,6 +826,7 @@ class DeltaTree
     size_t num_inserts = 0;
     size_t num_deletes = 0;
     size_t num_entries = 0;
+    Int64 max_dup_tuple_id = -1;
 
     std::unique_ptr<Allocator> allocator;
     size_t bytes = 0;
@@ -1040,6 +1041,8 @@ class DeltaTree
     size_t numEntries() const { return num_entries; }
     size_t numInserts() const { return num_inserts; }
     size_t numDeletes() const { return num_deletes; }
+    Int64 maxDupTupleID() const { return max_dup_tuple_id; }
+    void setMaxDupTupleID(Int64 tuple_id) { max_dup_tuple_id = std::max(tuple_id, max_dup_tuple_id); }
 
     void addDelete(UInt64 rid);
     void addInsert(UInt64 rid, UInt64 tuple_id);
@@ -1056,6 +1059,7 @@ DT_CLASS::DeltaTree(const DT_CLASS::Self & o)
     , num_inserts(o.num_inserts)
     , num_deletes(o.num_deletes)
     , num_entries(o.num_entries)
+    , max_dup_tuple_id(o.max_dup_tuple_id)
     , allocator(std::make_unique<Allocator>())
 {
     // If exception is thrown before clear copying_nodes, all nodes will be destroyed.

dbms/src/Storages/DeltaMerge/ReadThread/UnorderedInputStream.h

@@ -22,6 +22,10 @@
 
 namespace DB::DM
 {
+namespace tests
+{
+class DeltaMergeStoreRWTest;
+}
 class UnorderedInputStream : public IProfilingBlockInputStream
 {
     static constexpr auto NAME = "UnorderedInputStream";
@@ -151,5 +155,7 @@ class UnorderedInputStream : public IProfilingBlockInputStream
     // runtime filter
     std::vector<RuntimeFilterPtr> runtime_filter_list;
     int max_wait_time_ms;
+
+    friend class tests::DeltaMergeStoreRWTest;
 };
 } // namespace DB::DM

dbms/src/Storages/DeltaMerge/Segment.cpp

@@ -2485,7 +2485,8 @@ std::pair<DeltaIndexPtr, bool> Segment::ensurePlace(
 {
     const auto & stable_snap = segment_snap->stable;
     auto delta_snap = delta_reader->getDeltaSnap();
-    // Clone a new delta index.
+    // Try to clone from the sahred delta index, if it fails to reuse the shared delta index,
+    // it will return an empty delta index and we should place it in the following branch.
     auto my_delta_index = delta_snap->getSharedDeltaIndex()->tryClone(delta_snap->getRows(), delta_snap->getDeletes());
     auto my_delta_tree = my_delta_index->getDeltaTree();
 
@@ -2501,9 +2502,17 @@ std::pair<DeltaIndexPtr, bool> Segment::ensurePlace(
     auto [my_placed_rows, my_placed_deletes] = my_delta_index->getPlacedStatus();
 
     // Let's do a fast check, determine whether we need to do place or not.
-    if (!delta_reader->shouldPlace(dm_context, my_delta_index, rowkey_range, relevant_range, max_version))
+    if (!delta_reader->shouldPlace( //
+            dm_context,
+            my_placed_rows,
+            my_placed_deletes,
+            rowkey_range,
+            relevant_range,
+            max_version))
+    {
+        // We can reuse the shared-delta-index
         return {my_delta_index, false};
-
+    }
     CurrentMetrics::Increment cur_dm_segments{CurrentMetrics::DT_PlaceIndexUpdate};
     GET_METRIC(tiflash_storage_subtask_count, type_place_index_update).Increment();
     Stopwatch watch;