Restore DeltaMergeStore take too much time

[flowbehappy]

Enhancement

A TiFlash node takes over 15 minutes to get ready for service. Because the restore process of DeltaMergeStore is very slow when there are too many ColumnFileTinys.

[2022/11/30 22:45:41.060 +08:00] [INFO] [DeltaMergeStore.cpp:209] ["Restore DeltaMerge Store start"] [source="table_id=11"] [thread_id=44]
...
[2022/11/30 23:03:32.419 +08:00] [INFO] [DeltaMergeStore.cpp:284] ["Restore DeltaMerge Store end, ps_run_mode=2"] [source="table_id=73"] [thread_id=44]

mysql> select SEGMENT_COUNT, AVG_PACK_COUNT_IN_DELTA, TOTAL_PACK_COUNT_IN_DELTA, TIFLASH_INSTANCE from information_schema.tiflash_tables where tidb_table = 'table_name';
+---------------+-------------------------+---------------------------+------------------+
| SEGMENT_COUNT | AVG_PACK_COUNT_IN_DELTA | TOTAL_PACK_COUNT_IN_DELTA | TIFLASH_INSTANCE |
+---------------+-------------------------+---------------------------+------------------+
|         18491 |      6.6471258752646145 |                    122460 | 172.16.5.81:7940 |
|         91006 |      5.6769096009753675 |                    479591 | 172.16.5.82:7940 |
+---------------+-------------------------+---------------------------+------------------+

[JaySon-Huang]

Restoring one table with about 360GB data
profile_2022-12-01_16-12-00.zip

[hehechen]

tiflash/dbms/src/DataTypes/DataTypeFactory.cpp

Line 37 in e255112

DataTypePtr DataTypeFactory::get(const String & full_name) const

brings too much overhead when get column datatype. We can introduce a new function with datatype cache like

tiflash/dbms/src/DataTypes/DataTypeFactory.cpp

Line 67 in e255112

DataTypePtr DataTypeFactory::get(const String & family_name, const ASTPtr & parameters) const

.

We can use a unordered_map to store the cache for full_name -> DataTypePtr to avoid parsing full_name to AST repeatedly.

tiflash/dbms/src/DataTypes/DataTypeFactory.cpp

Lines 44 to 60 in a22898b

	DataTypePtr DataTypeFactory::getOrSet(const String & full_name)
	{
	auto it = fullname_types.find(full_name);
	if (it != fullname_types.end())
	{
	return it->second;
	}
	ParserIdentifierWithOptionalParameters parser;
	ASTPtr ast = parseQuery(parser, full_name.data(), full_name.data() + full_name.size(), "data type", 0);
	DataTypePtr datatype_ptr = get(ast);
	// avoid big hashmap in rare cases.
	if (fullname_types.size() < MAX_FULLNAME_TYPES)
	{
	fullname_types.emplace(full_name, datatype_ptr);
	}
	return datatype_ptr;
	}