ddl: Fix default value filling with finer granularity

dbms/src/Storages/KVStore/Decode/RegionBlockReader.cpp

@@ -179,6 +179,7 @@ bool RegionBlockReader::readImpl(Block & block, const ReadList & data_list, bool
 {
     VersionColResolver<ReadList> version_col_resolver;
     version_col_resolver.check(block, schema_snapshot->column_defines->size());
+    // The column_ids to read according to schema_snapshot, each elem is (column_id, block_pos)
     const auto & read_column_ids = schema_snapshot->getColId2BlockPosMap();
     const auto & pk_column_ids = schema_snapshot->pk_column_ids;
     const auto & pk_pos_map = schema_snapshot->pk_pos_map;
@@ -269,6 +270,8 @@ bool RegionBlockReader::readImpl(Block & block, const ReadList & data_list, bool
             else
             {
                 // Parse column value from encoded value
+                // Decode the column_ids from `column_ids_iter` to `read_column_ids.end()`
+                // and insert into `block` at position starting from `next_column_pos`
                 if (!appendRowToBlock(
                         *value_ptr,
                         column_ids_iter,

dbms/src/Storages/KVStore/tests/gtest_region_block_reader.cpp

@@ -692,5 +692,146 @@ try
 }
 CATCH
 
+TEST_F(RegionBlockReaderTest, ReadFromRegionDefaultValue)
+try
+{
+    // With this table_info, column "c1" is "NOT NULL" and has no origin default
+    TableInfo table_info_c1_not_null_no_origin_default(
+        R"({"cols":[{"id":1,"name":{"L":"c0","O":"c0"},"offset":0,"state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":0,"Flen":4,"Tp":1}},{"id":2,"name":{"L":"handle","O":"handle"},"offset":1,"state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":515,"Flen":11,"Tp":3}},{"default":"-56083770","id":7,"name":{"L":"c1","O":"c1"},"offset":2,"state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":1,"Flen":20,"Tp":8}},{"id":4,"name":{"L":"c2","O":"c2"},"offset":3,"origin_default":"0.07954397","state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":-1,"Flag":4097,"Flen":12,"Tp":4}},{"id":5,"name":{"L":"c5","O":"c5"},"offset":4,"origin_default":"0","state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":0,"Flen":10,"Tp":246}},{"default":"247262911","id":6,"name":{"L":"c4","O":"c4"},"offset":5,"origin_default":"247262911","state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":0,"Flen":10,"Tp":246}}],"id":636,"index_info":[],"is_common_handle":false,"keyspace_id":4294967295,"name":{"L":"t0","O":"t0"},"pk_is_handle":true,"state":5,"tiflash_replica":{"Available":true,"Count":1},"update_timestamp":463845180343844895})",
+        NullspaceID);
+
+    // With this table_info, column "c1" is "NOT NULL" and has no default value
+    TableInfo table_info_c1_not_null_no_default_value(
+        R"({"cols":[{"id":1,"name":{"L":"c0","O":"c0"},"offset":0,"state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":0,"Flen":4,"Tp":1}},{"id":2,"name":{"L":"handle","O":"handle"},"offset":1,"state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":515,"Flen":11,"Tp":3}},{"default":"-56083770","id":7,"name":{"L":"c1","O":"c1"},"offset":2,"state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":4097,"Flen":20,"Tp":8}},{"id":4,"name":{"L":"c2","O":"c2"},"offset":3,"origin_default":"0.07954397","state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":-1,"Flag":4097,"Flen":12,"Tp":4}},{"id":5,"name":{"L":"c5","O":"c5"},"offset":4,"origin_default":"0","state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":0,"Flen":10,"Tp":246}},{"default":"247262911","id":6,"name":{"L":"c4","O":"c4"},"offset":5,"origin_default":"247262911","state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":0,"Flen":10,"Tp":246}}],"id":636,"index_info":[],"is_common_handle":false,"keyspace_id":4294967295,"name":{"L":"t0","O":"t0"},"pk_is_handle":true,"state":5,"tiflash_replica":{"Available":true,"Count":1},"update_timestamp":463845180343844895})",
+        NullspaceID);
+
+    // With this table_info, column "c1" has the "NOT NULL" flag and has origin default "-56083770"
+    TableInfo table_info_c1_not_null_with_origin_default(
+        R"({"cols":[{"id":1,"name":{"L":"c0","O":"c0"},"offset":0,"state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":0,"Flen":4,"Tp":1}},{"id":2,"name":{"L":"handle","O":"handle"},"offset":1,"state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":515,"Flen":11,"Tp":3}},{"origin_default":"-56083770","id":7,"name":{"L":"c1","O":"c1"},"offset":2,"state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":1,"Flen":20,"Tp":8}},{"id":4,"name":{"L":"c2","O":"c2"},"offset":3,"origin_default":"0.07954397","state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":-1,"Flag":4097,"Flen":12,"Tp":4}},{"id":5,"name":{"L":"c5","O":"c5"},"offset":4,"origin_default":"0","state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":0,"Flen":10,"Tp":246}},{"default":"247262911","id":6,"name":{"L":"c4","O":"c4"},"offset":5,"origin_default":"247262911","state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":0,"Flen":10,"Tp":246}}],"id":636,"index_info":[],"is_common_handle":false,"keyspace_id":4294967295,"name":{"L":"t0","O":"t0"},"pk_is_handle":true,"state":5,"tiflash_replica":{"Available":true,"Count":1},"update_timestamp":463845180343844895})",
+        NullspaceID);
+
+    // With this table_info, column "c1" does not have the "NOT NULL" flag
+    TableInfo table_info_c1_nullable(
+        R"({"cols":[{"id":1,"name":{"L":"c0","O":"c0"},"offset":0,"state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":0,"Flen":4,"Tp":1}},{"id":2,"name":{"L":"handle","O":"handle"},"offset":1,"state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":515,"Flen":11,"Tp":3}},{"id":7,"name":{"L":"c1","O":"c1"},"offset":2,"state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":0,"Flen":20,"Tp":8}},{"id":4,"name":{"L":"c2","O":"c2"},"offset":3,"origin_default":"0.07954397","state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":-1,"Flag":4097,"Flen":12,"Tp":4}},{"id":5,"name":{"L":"c5","O":"c5"},"offset":4,"origin_default":"0","state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":0,"Flen":10,"Tp":246}},{"default":"247262911","id":6,"name":{"L":"c4","O":"c4"},"offset":5,"origin_default":"247262911","state":5,"type":{"Charset":"binary","Collate":"binary","Decimal":0,"Flag":0,"Flen":10,"Tp":246}}],"id":636,"index_info":[],"is_common_handle":false,"keyspace_id":4294967295,"name":{"L":"t0","O":"t0"},"pk_is_handle":true,"state":5,"tiflash_replica":{"Available":true,"Count":1},"update_timestamp":463844343842340870})",
+        NullspaceID);
+
+    RegionID region_id = 4;
+    // the start_key and end_key for table_id = 68
+    String region_start_key(bytesFromHexString("7480000000000002FF7C5F720000000000FA"));
+    String region_end_key(bytesFromHexString("7480000000000002FF7D00000000000000F8"));
+    auto region = RegionBench::makeRegionForRange(region_id, region_start_key, region_end_key);
+    // the hex kv dump from RaftLog
+    std::vector<std::tuple<std::string_view, std::string_view>> kvs = {
+        {
+            "7480000000000002FFA95F728000000000FF0000010000000000FAF9901806DEF7FFDA",
+            "50A380A08892FFF9B706762C8000040000000405060708000F0016001A00BFDC4011A00000000A0080000000000A008000003CA339"
+            "1ABC85",
+        },
+        {
+            "7480000000000002FFA95F728000000000FF0000010000000000FAF9901806DEF7FFD8",
+            "50A680A08892FFF9B706762C8000040000000405060708000F0016001A00BFDC4011A00000000A008033E04D600A008000003CA339"
+            "1ABC85",
+        },
+        {
+            "7480000000000002FFA95F728000000000FF0000020000000000FAF9901806DE33FFE8",
+            "509680B08E92FFF9B706762580000300000004050608000F001600BF720CDD400000000A0080000000010A00800000393C",
+        },
+    };
+    for (const auto & [k, v] : kvs)
+    {
+        region->insertDebug("write", TiKVKey(bytesFromHexString(k)), TiKVValue(bytesFromHexString(v)));
+    }
+
+    auto data_list_read = ReadRegionCommitCache(region, true);
+    ASSERT_TRUE(data_list_read.has_value());
+
+    // Test with `table_info_c1_not_null_no_origin_default`
+    auto decoding_schema = getDecodingStorageSchemaSnapshot(table_info_c1_not_null_no_origin_default);
+    {
+        // force_decode=false can not decode because there are
+        // missing value for column with not null flag.
+        auto reader = RegionBlockReader(decoding_schema);
+        Block res_block = createBlockSortByColumnID(decoding_schema);
+        ASSERT_FALSE(reader.read(res_block, *data_list_read, false));
+    }
+    {
+        // force_decode=true can decode the block, and filling the default value for c1
+        auto reader = RegionBlockReader(decoding_schema);
+        Block res_block = createBlockSortByColumnID(decoding_schema);
+        ASSERT_TRUE(reader.read(res_block, *data_list_read, true));
+        LOG_INFO(
+            Logger::get(),
+            "Decoded block:\n{}",
+            DB::tests::getColumnsContent(res_block.getColumnsWithTypeAndName()));
+        ASSERT_EQ(res_block.getByName("c1").type->getName(), "Int64");
+        // verify the default value is filled correctly
+        ASSERT_COLUMN_EQ( //
+            res_block.getByName("c1"),
+            createColumn<Int64>({-2051270087, -2051270087, 0}));
+    }
+
+    // Test with `table_info_c1_not_null_no_default_value`
+    decoding_schema = getDecodingStorageSchemaSnapshot(table_info_c1_not_null_no_default_value);
+    {
+        // force_decode=false can not decode because there are
+        // missing value for column with not null flag.
+        auto reader = RegionBlockReader(decoding_schema);
+        Block res_block = createBlockSortByColumnID(decoding_schema);
+        ASSERT_FALSE(reader.read(res_block, *data_list_read, false));
+    }
+    {
+        // force_decode=true can decode the block, and filling the default value for c1
+        auto reader = RegionBlockReader(decoding_schema);
+        Block res_block = createBlockSortByColumnID(decoding_schema);
+        ASSERT_TRUE(reader.read(res_block, *data_list_read, true));
+        LOG_INFO(
+            Logger::get(),
+            "Decoded block:\n{}",
+            DB::tests::getColumnsContent(res_block.getColumnsWithTypeAndName()));
+        ASSERT_EQ(res_block.getByName("c1").type->getName(), "Int64");
+        // verify the default value is filled correctly
+        ASSERT_COLUMN_EQ( //
+            res_block.getByName("c1"),
+            createColumn<Int64>({-2051270087, -2051270087, 0}));
+    }
+
+    // Test with `table_info_c1_not_null_with_origin_default`
+    decoding_schema = getDecodingStorageSchemaSnapshot(table_info_c1_not_null_with_origin_default);
+    {
+        // force_decode=false can decode because origin_default exists and NoDefaultValue flag is not set
+        // so RegionBlockReader can use origin_default to fill the missing value`
+        auto reader = RegionBlockReader(decoding_schema);
+        Block res_block = createBlockSortByColumnID(decoding_schema);
+        ASSERT_TRUE(reader.read(res_block, *data_list_read, false));
+        LOG_INFO(
+            Logger::get(),
+            "Decoded block:\n{}",
+            DB::tests::getColumnsContent(res_block.getColumnsWithTypeAndName()));
+        ASSERT_EQ(res_block.getByName("c1").type->getName(), "Int64");
+        // verify the default value is filled correctly
+        ASSERT_COLUMN_EQ( //
+            res_block.getByName("c1"),
+            createColumn<Int64>({-2051270087, -2051270087, -56083770}));
+    }
+
+    // Test with `table_info_c1_nullable`
+    decoding_schema = getDecodingStorageSchemaSnapshot(table_info_c1_nullable);
+    {
+        // force_decode=false should be able to decode because c1 is nullable
+        auto reader = RegionBlockReader(decoding_schema);
+        Block res_block = createBlockSortByColumnID(decoding_schema);
+        ASSERT_TRUE(reader.read(res_block, *data_list_read, false));
+        LOG_INFO(
+            Logger::get(),
+            "Decoded block:\n{}",
+            DB::tests::getColumnsContent(res_block.getColumnsWithTypeAndName()));
+        ASSERT_EQ(res_block.getByName("c1").type->getName(), "Nullable(Int64)");
+        // verify the default value is filled with NULL correctly at the last row
+        ASSERT_COLUMN_EQ( //
+            res_block.getByName("c1"),
+            createNullableColumn<Int64>({-2051270087, -2051270087, 0}, {0, 0, 1}));
+    }
+}
+CATCH
 
 } // namespace DB::tests

dbms/src/TiDB/Decode/RowCodec.cpp

@@ -388,40 +388,55 @@ bool appendRowToBlock(
     }
 }
 
+// When a column is missing in the encoded row, decide whether we can append a value
+// (NULL or origin default) to `block` or we must trigger schema sync.
+//
+// Background
+// - TiDB encode semantics, see [tables.CanSkip](https://github.com/pingcap/tidb/blob/v8.5.5/pkg/table/tables/tables.go#L1463-L1489)
+// - PK handle columns may be omitted from the value part and decoded from the key.
+// - A NULL column may be omitted only when BOTH DefaultValue and OriginDefaultValue are empty.
+// - For NOT NULL columns, TiDB must encode the value in the row unless the column did not
+//   exist in the writer schema (old data); in that case OriginDefaultValue is expected.
+//
+// Policy here:
+// - If the omission is clearly valid under the current schema, we fill a value and return true.
+// - Otherwise return false (force_decode == false) to let the caller sync schema and retry.
+// - If force_decode == true, we fall back to best-effort filling.
 inline bool addDefaultValueToColumnIfPossible(
     const ColumnInfo & column_info,
     Block & block,
     size_t block_column_pos,
     bool ignore_pk_if_absent,
     bool force_decode)
 {
-    // We consider a missing column could be safely filled with NULL, unless it has not default value and is NOT NULL.
-    // This could saves lots of unnecessary schema syncs for old data with a newer schema that has newly added columns.
-
+    // 1) Primary-key columns can be decoded from the key for pk_is_handle / common-handle tables.
+    //    Skip value-part filling here if allowed.
     if (column_info.hasPriKeyFlag())
     {
-        // For clustered index or pk_is_handle, if the pk column does not exists, it can still be decoded from the key
         if (ignore_pk_if_absent)
             return true;
 
-        assert(!ignore_pk_if_absent);
+        // For non-clustered tables, a missing PK column implies schema mismatch.
         if (!force_decode)
             return false;
-        // Else non-clustered index, and not pk_is_handle, it could be a row encoded by older schema,
-        // we need to fill the column which has primary key flag with default value.
-        // fallthrough to fill default value when force_decode
+        // fallthrough for best-effort fill when force_decode == true
     }
 
-    if (column_info.hasNoDefaultValueFlag() && column_info.hasNotNullFlag())
+    // 2) NOT NULL columns:
+    //    - If the column has NO DEFAULT, TiDB should always encode a value. Missing datum implies mismatch.
+    //    - If the column has NO origin default, missing datum may come from a newer schema where the column
+    //      became NULLABLE and was skipped; require schema sync.
+    //    - If origin default exists, missing datum can be from older rows before the column was added; safe to fill.
+    if (!force_decode && column_info.hasNotNullFlag())
     {
-        if (!force_decode)
+        if (column_info.hasNoDefaultValueFlag() || !column_info.hasOriginDefaultValue())
             return false;
-        // Else the row does not contain this "not null" / "no default value" column,
-        // it could be a row encoded by older schema.
-        // fallthrough to fill default value when force_decode
     }
-    // not null or has no default value, tidb will fill with specific value.
-    auto * raw_column = const_cast<IColumn *>((block.getByPosition(block_column_pos)).column.get());
+
+    // 3) Fill using origin default or NULL.
+    //    Note: defaultValueToField() uses origin_default_value/origin_default_bit_value,
+    //    and falls back to NULL (nullable) or GenDefaultField (NOT NULL) when they are empty.
+    auto * raw_column = const_cast<IColumn *>(block.getByPosition(block_column_pos).column.get());
     raw_column->insert(column_info.defaultValueToField());
     return true;
 }
@@ -460,7 +475,9 @@ bool appendRowV2ToBlockImpl(
         num_not_null_columns,
         value_offsets);
     size_t values_start_pos = cursor;
+    // how many not null columns have been processed
     size_t idx_not_null = 0;
+    // how many null columns have been processed
     size_t idx_null = 0;
     // Merge ordered not null/null columns to keep order.
     while (idx_not_null < not_null_column_ids.size() || idx_null < null_column_ids.size())
@@ -481,20 +498,21 @@ bool appendRowV2ToBlockImpl(
         const auto next_column_id = column_ids_iter->first;
         if (next_column_id > next_datum_column_id)
         {
-            // The next column id to read is bigger than the column id of next datum in encoded row.
+            // The next_column_id to read is bigger than the next_datum_column_id in encoded row.
             // It means this is the datum of extra column. May happen when reading after dropping
             // a column.
+            // For `force_decode == false`, we should return false to let upper layer trigger schema sync.
             if (!force_decode)
                 return false;
-            // Ignore the extra column and continue to parse other datum
+            // For `force_decode == true`, we just skip this extra column and continue to parse other datum.
             if (is_null)
                 idx_null++;
             else
                 idx_not_null++;
         }
         else if (next_column_id < next_datum_column_id)
         {
-            // The next column id to read is less than the column id of next datum in encoded row.
+            // The next_column_id to read is less than the next_datum_column_id in encoded row.
             // It means this is the datum of missing column. May happen when reading after adding
             // a column.
             // Fill with default value and continue to read data for next column id.
@@ -505,7 +523,10 @@ bool appendRowV2ToBlockImpl(
                     block_column_pos,
                     ignore_pk_if_absent,
                     force_decode))
+            {
+                // If failed to fill default value, return false to let upper layer trigger schema sync.
                 return false;
+            }
             column_ids_iter++;
             block_column_pos++;
         }
@@ -570,8 +591,12 @@ bool appendRowV2ToBlockImpl(
             block_column_pos++;
         }
     }
+
+    // There are more columns to read other than the datum encoded in the row.
     while (column_ids_iter != column_ids_iter_end)
     {
+        // Skip if the column_id is the same as `pk_handle_id`. The value of column
+        // `pk_handle_id` will be filled in upper layer but not in this function.
         if (column_ids_iter->first != pk_handle_id)
         {
             const auto & column_info = column_infos[column_ids_iter->second];
@@ -581,7 +606,10 @@ bool appendRowV2ToBlockImpl(
                     block_column_pos,
                     ignore_pk_if_absent,
                     force_decode))
+            {
+                // If failed to fill default value, return false to let upper layer trigger schema sync.
                 return false;
+            }
         }
         column_ids_iter++;
         block_column_pos++;