Bring back optimization for reading subcolumns of single column in Compact parts

src/DataTypes/Serializations/SerializationObject.cpp

@@ -9,6 +9,8 @@
 #include <DataTypes/DataTypeString.h>
 #include <IO/ReadBufferFromString.h>
 
+#include <Common/logger_useful.h>
+
 namespace DB
 {
 

src/Storages/MergeTree/MergeTreeReaderCompact.cpp

@@ -148,7 +148,8 @@ void MergeTreeReaderCompact::readData(
     ColumnPtr & column,
     size_t rows_to_read,
     const InputStreamGetter & getter,
-    ISerialization::SubstreamsCache & cache)
+    ISerialization::SubstreamsCache & cache,
+    std::unordered_map<String, ColumnPtr> & columns_cache_for_subcolumns)
 {
     try
     {
@@ -171,6 +172,17 @@ void MergeTreeReaderCompact::readData(
             const auto & type_in_storage = name_and_type.getTypeInStorage();
             const auto & name_in_storage = name_and_type.getNameInStorage();
 
+            if (auto cache_for_subcolumns_it = columns_cache_for_subcolumns.find(name_in_storage); cache_for_subcolumns_it != columns_cache_for_subcolumns.end())
+            {
+                auto subcolumn = type_in_storage->getSubcolumn(name_and_type.getSubcolumnName(), cache_for_subcolumns_it->second);
+                /// TODO: Avoid extra copying.
+                if (column->empty())
+                    column = IColumn::mutate(subcolumn);
+                else
+                    column->assumeMutable()->insertRangeFrom(*subcolumn, 0, subcolumn->size());
+            }
+            else
+            {
                 auto serialization = getSerializationInPart({name_in_storage, type_in_storage});
                 ColumnPtr temp_column = type_in_storage->createColumn(*serialization);
 
@@ -182,6 +194,9 @@ void MergeTreeReaderCompact::readData(
                     column = subcolumn;
                 else
                     column->assumeMutable()->insertRangeFrom(*subcolumn, 0, subcolumn->size());
+
+                columns_cache_for_subcolumns[name_in_storage] = temp_column;
+            }
         }
         else
         {

src/Storages/MergeTree/MergeTreeReaderCompact.h

@@ -45,7 +45,8 @@ protected:
         ColumnPtr & column,
         size_t rows_to_read,
         const InputStreamGetter & getter,
-        ISerialization::SubstreamsCache & cache);
+        ISerialization::SubstreamsCache & cache,
+        std::unordered_map<String, ColumnPtr> & columns_cache_for_subcolumns);
 
     void readPrefix(
         const NameAndTypePair & name_and_type,

src/Storages/MergeTree/MergeTreeReaderCompactSingleBuffer.cpp

@@ -29,6 +29,12 @@ try
         /// Use cache to avoid reading the column with the same name twice.
         /// It may happen if there are empty array Nested in the part.
         ISerialization::SubstreamsCache cache;
+        /// If we need to read multiple subcolumns from a single column in storage,
+        /// we will read it this column only once and then reuse to extract all subcolumns.
+        /// We cannot use SubstreamsCache for it, because we may also read the full column itself
+        /// and it might me not empty inside res_columns (and SubstreamsCache contains the whole columns).
+        /// TODO: refactor the code in a way when we first read all full columns and then extract all subcolumns from them.
+        std::unordered_map<String, ColumnPtr> columns_cache_for_subcolumns;
 
         for (size_t pos = 0; pos < num_columns; ++pos)
         {
@@ -56,7 +62,7 @@ try
             };
 
             readPrefix(columns_to_read[pos], buffer_getter, buffer_getter_for_prefix, columns_for_offsets[pos]);
-            readData(columns_to_read[pos], column, rows_to_read, buffer_getter, cache);
+            readData(columns_to_read[pos], column, rows_to_read, buffer_getter, cache, columns_cache_for_subcolumns);
         }
 
         ++from_mark;