In the member function insert(position, elem) to class Block, if function checkColumnStructure<void> throws an exception for columns with same name but unidentical structure, it will result in the mismatch between index_by_name and data. However we could not presume upon that caller will not use this block object anymore. So it seems better to guarantee consistency innerly.
Stress tests founds [1]:
==527==WARNING: MemorySanitizer: use-of-uninitialized-value
0 0x37078ffd in unsigned long std::__1::__cxx_atomic_fetch_add<unsigned long>(std::__1::__cxx_atomic_base_impl<unsigned long>*, unsigned long, std::__1::memory_order) obj-x86_64-linux-gnu/../contrib/libcxx/include/atomic:1050:12
1 0x37078ffd in std::__1::__atomic_base<unsigned long, true>::fetch_add(unsigned long, std::__1::memory_order) obj-x86_64-linux-gnu/../contrib/libcxx/include/atomic:1719:17
2 0x37078ffd in std::__1::__atomic_base<unsigned long, true>::operator++() obj-x86_64-linux-gnu/../contrib/libcxx/include/atomic:1756:57
3 0x37078ffd in SimpleIncrement::get() obj-x86_64-linux-gnu/../src/Common/SimpleIncrement.h:20:16
4 0x37078ffd in DB::MergeTreeDataWriter::writeTempPart(DB::BlockWithPartition&, std::__1::shared_ptr<DB::StorageInMemoryMetadata const> const&, std::__1::shared_ptr<DB::Context const>) obj-x86_64-linux-gnu/../src/Storages/MergeTree/MergeTreeDataWriter.cpp:276:46
5 0x373c446c in DB::MergeTreeSink::consume(DB::Chunk) obj-x86_64-linux-gnu/../src/Storages/MergeTree/MergeTreeSink.cpp:27:65
Uninitialized value was created by a heap deallocation
6 0x32d481e8 in DB::DatabaseCatalog::TableMarkedAsDropped::~TableMarkedAsDropped() obj-x86_64-linux-gnu/../src/Interpreters/DatabaseCatalog.h:248:12
7 0x32d3c134 in DB::DatabaseCatalog::dropTableDataTask() obj-x86_64-linux-gnu/../src/Interpreters/DatabaseCatalog.cpp:908:1
[1]: https://s3.amazonaws.com/clickhouse-test-reports/32534/fa6090f588dbf4cbb5f28bd2210847b070bb8218/stress_test__memory__actions_.html
The query was CREATE MATERIALIZED VIEW ... POPULATE AS SELECT ... from
00040_aggregating_materialized_view test.
This patch will:
- fix the issue when table had been tried to create multiple times for
--queries-files
- create these table for --interactive mode (before it works only if you
had some queries already, i.e. when it run interactive after
non-interactive)
This will also make ClientBase interface a little bit cleaner, by
removing one abstract method getQueryTextPrefix()
