CLICKHOUSE-4127: Fix ALTER of destination table for the BUFFER engine.

[vitlibar]

I hereby agree to the terms of the CLA available at: https://yandex.ru/legal/cla/?lang=en

[vitlibar]

There was an exception at this line:

<Error> executeQuery: Code: 16, e.displayText() = DB::Exception: There is no column with name x. There are columns: y, z: (when looking at destination table test.dst), e.what() = DB::Exception (from 127.0.0.1:40774) (in query: INSERT INTO test.buffer VALUES), Stack trace:

/home/user/work/ClickHouse/build/dbms/programs/clickhouse-server(DB::Exception::Exception(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, int)+0x22) [0x2a67762]
/home/user/work/ClickHouse/build/dbms/programs/clickhouse-server(DB::ITableDeclaration::check(DB::Block const&, bool) const+0x3bc) [0x58918fc]
/home/user/work/ClickHouse/build/dbms/programs/clickhouse-server(DB::BufferBlockOutputStream::write(DB::Block const&)+0x12f) [0x5d00a9f

With my changes this line is successfully executed and causes inserting new row

y	z
400	DEFZ

to the table test.dst.

[vitlibar]

I've removed this check to allow writing blocks to StorageBuffer which doesn't match the structure of the destination table. All necessary data conversion is done by the function StorageBuffer::writeBlockToDestination() just before writing block to the destination table. This change is necessary in case the structure of the destination table is altered.

[vitlibar]

I'm not sure about this change. When reading data from a BUFFER table we should process both the BUFFER and the destination table. But after altering the destination table we cannot simply join the data from these two sources because they have different structure.

[alexey-milovidov]

We can read common subset of columns from the destination table and apply ConvertingBlockInputStream.
I think, it will be more convenient than reading from buffer only.

[vitlibar]

Using common subset can look strange, I reckon. For example:

CREATE TABLE test.dst (x UInt64, y UInt64) ENGINE = MergeTree ORDER BY tuple();
CREATE TABLE test.buffer (x UInt64, y UInt64) ENGINE = Buffer(test, dst, 1, 99999, 99999, 1, 1, 99999, 99999);
ALTER TABLE test.dst DROP COLUMN x, MODIFY COLUMN y String, ADD COLUMN z String;
SELECT x, y FROM test.buffer;  -- will produce only the column y

I mean SELECT col1, col2 FROM table always produced two-column output for any table. And Clickhouse throws an exception on trying to select non-existing column from a table.

Maybe it would be better to add missing column filled with default values to the blocks retrieved from the destination table. What do you think?

[vitlibar]

Done.

I've added a new class AddingDefaultBlockInputStream and created a new source file addMissingDefaults.cpp to keep common code for AddingDefaultBlockInputStream and AddingDefaultBlockOutputStream.