flush cache before segment merge (#4955)

dbms/src/Storages/DeltaMerge/DeltaMergeStore.cpp

@@ -1275,6 +1275,29 @@ void DeltaMergeStore::segmentMerge(DMContext & dm_context, const SegmentPtr & le
 {
     LOG_DEBUG(log, "Merge Segment [" << left->info() << "] and [" << right->info() << "], safe point:" << dm_context.min_version);
 
+    /// This segment may contain some rows that not belong to this segment range which is left by previous split operation.
+    /// And only saved data in this segment will be filtered by the segment range in the merge process,
+    /// unsaved data will be directly copied to the new segment.
+    /// So we flush here to make sure that all potential data left by previous split operation is saved.
+    while (!left->flushCache(dm_context))
+    {
+        // keep flush until success if not abandoned
+        if (left->hasAbandoned())
+        {
+            LOG_DEBUG(log, "Give up merge segments left [" << left->segmentId() << "], right [" << right->segmentId() << "]");
+            return;
+        }
+    }
+    while (!right->flushCache(dm_context))
+    {
+        // keep flush until success if not abandoned
+        if (right->hasAbandoned())
+        {
+            LOG_DEBUG(log, "Give up merge segments left [" << left->segmentId() << "], right [" << right->segmentId() << "]");
+            return;
+        }
+    }
+
     SegmentSnapshotPtr left_snap;
     SegmentSnapshotPtr right_snap;
     ColumnDefinesPtr   schema_snap;

dbms/src/Storages/DeltaMerge/Segment.cpp

@@ -240,7 +240,7 @@ bool Segment::writeToCache(DMContext & dm_context, const Block & block, size_t o
     return delta->appendToCache(dm_context, block, offset, limit);
 }
 
-bool Segment::write(DMContext & dm_context, const Block & block)
+bool Segment::write(DMContext & dm_context, const Block & block, bool flush_cache)
 {
     LOG_TRACE(log, "Segment [" << segment_id << "] write to disk rows: " << block.rows());
     WriteBatches wbs(dm_context.storage_pool);
@@ -249,7 +249,14 @@ bool Segment::write(DMContext & dm_context, const Block & block)
 
     if (delta->appendToDisk(dm_context, pack))
     {
-        flushCache(dm_context);
+        if (flush_cache)
+        {
+            while (!flushCache(dm_context))
+            {
+                if (hasAbandoned())
+                    return false;
+            }
+        }
         return true;
     }
     else
@@ -957,6 +964,28 @@ SegmentPair Segment::applySplit(DMContext &                dm_context, //
 SegmentPtr Segment::merge(DMContext & dm_context, const ColumnDefinesPtr & schema_snap, const SegmentPtr & left, const SegmentPtr & right)
 {
     WriteBatches wbs(dm_context.storage_pool);
+    /// This segment may contain some rows that not belong to this segment range which is left by previous split operation.
+    /// And only saved data in this segment will be filtered by the segment range in the merge process,
+    /// unsaved data will be directly copied to the new segment.
+    /// So we flush here to make sure that all potential data left by previous split operation is saved.
+    while (!left->flushCache(dm_context))
+    {
+        // keep flush until success if not abandoned
+        if (left->hasAbandoned())
+        {
+            LOG_FMT_DEBUG(left->log, "Give up merge segments left [{}], right [{}]", left->segmentId(), right->segmentId());
+            return {};
+        }
+    }
+    while (!right->flushCache(dm_context))
+    {
+        // keep flush until success if not abandoned
+        if (right->hasAbandoned())
+        {
+            LOG_FMT_DEBUG(right->log, "Give up merge segments left [{}], right [{}]", left->segmentId(), right->segmentId());
+            return {};
+        }
+    }
 
     auto left_snap  = left->createSnapshot(dm_context, true);
     auto right_snap = right->createSnapshot(dm_context, true);
@@ -977,6 +1006,10 @@ SegmentPtr Segment::merge(DMContext & dm_context, const ColumnDefinesPtr & schem
     return merged;
 }
 
+/// Segments may contain some rows that not belong to its range which is left by previous split operation.
+/// And only saved data in the segment will be filtered by the segment range in the merge process,
+/// unsaved data will be directly copied to the new segment.
+/// So remember to do a flush for the segments before merge.
 StableValueSpacePtr Segment::prepareMerge(DMContext &                dm_context, //
                                           const ColumnDefinesPtr &   schema_snap,
                                           const SegmentPtr &         left,

dbms/src/Storages/DeltaMerge/Segment.h

@@ -103,7 +103,7 @@ class Segment : private boost::noncopyable
 
     bool writeToDisk(DMContext & dm_context, const PackPtr & pack);
     bool writeToCache(DMContext & dm_context, const Block & block, size_t offset, size_t limit);
-    bool write(DMContext & dm_context, const Block & block); // For test only
+    bool write(DMContext & dm_context, const Block & block, bool flush_cache = true); // For test only
     bool write(DMContext & dm_context, const HandleRange & delete_range);
 
     SegmentSnapshotPtr createSnapshot(const DMContext & dm_context, bool for_update = false) const;

dbms/src/Storages/DeltaMerge/tests/gtest_dm_segment.cpp

@@ -636,11 +636,17 @@ CATCH
 TEST_F(Segment_test, Split)
 try
 {
-    const size_t num_rows_write = 100;
+    const size_t num_rows_write_per_batch = 100;
+    const size_t num_rows_write = num_rows_write_per_batch * 2;
     {
-        // write to segment
-        Block block = DMTestEnv::prepareSimpleWriteBlock(0, num_rows_write, false);
-        segment->write(dmContext(), std::move(block));
+        // write to segment and flush
+        Block block = DMTestEnv::prepareSimpleWriteBlock(0, num_rows_write_per_batch, false);
+        segment->write(dmContext(), std::move(block), true);
+    }
+    {
+        // write to segment and don't flush
+        Block block = DMTestEnv::prepareSimpleWriteBlock(num_rows_write_per_batch, 2 * num_rows_write_per_batch, false);
+        segment->write(dmContext(), std::move(block), false);
     }
 
     {
@@ -685,7 +691,7 @@ try
             in->readSuffix();
         }
         {
-            auto in = segment->getInputStream(dmContext(), *tableColumns());
+            auto in = new_segment->getInputStream(dmContext(), *tableColumns());
             in->readPrefix();
             while (Block block = in->read())
             {
@@ -696,9 +702,13 @@ try
         ASSERT_EQ(num_rows_seg1 + num_rows_seg2, num_rows_write);
     }
 
+    // delete rows in the right segment
+    {
+        new_segment->write(dmContext(), /*delete_range*/ new_segment->getRowKeyRange());
+        new_segment->flushCache(dmContext());
+    }
+
     // merge segments
-    // TODO: enable merge test!
-    if (false)
     {
         segment = Segment::merge(dmContext(), tableColumns(), segment, new_segment);
         {
@@ -717,7 +727,7 @@ try
                 num_rows_read += block.rows();
             }
             in->readSuffix();
-            EXPECT_EQ(num_rows_read, num_rows_write);
+            EXPECT_EQ(num_rows_read, num_rows_seg1);
         }
     }
 }