Performance tests:

merge_tree_many_partitions_2

Sped up two times.

codecs...

Various difference.

codecs...

Various difference.

It's unrelated, because compact parts don't support columns' special codecs. Need to force creation of wide parts in those tests.

compact parts don't support columns' special codecs

But why not?
It's reasonable to use codecs even for compact parts.
What if I want NONE... or if I know that Delta, LZ4 is better?

It's possible, but we have to write compressed block for every column in granule. Won't blocks be too small? Also, in that case it will be possible to count per column sizes.

Ok, I understand the reason.

There are plans to use ClickHouse with high-latency filesystem (e.g. S3). For that case the users may want to enable 100% compact parts. And if codecs won't work, it's a huge drawback.

@CurtizJ Do you want to take over this PR?

I want to finish #12183 first, and after it will be merged enable compact parts by default.

I don't understand why huge part of tests from skip list of check with enabled compact parts is skipping now in all checks. Skip list is unchanged in this PR.

It's because of code in clickhouse-test script.

00926_adaptive_index_granularity_pk: [ FAIL ] 670.21 sec. - Timeout!

In this test a table with index_granularity_bytes=40 is created, which leads to index_granularity=1. With compact parts this configuration is extremely inefficient, because of writing compressed block for every granule of column. It is not a regular case and performance tests didn't show regressions with default settings, when this behaviour was introduced and I think it's ok.

Maybe we need introduce a setting to allow set behaviour, when compressed block is written only when threshold is exceded.

The same about 00804_test_alter_compression_codecs.

Other fixed tests checked some implementation defined things like checksums or number of files or hardcoded paths to files. Also some compatibility tests, which run cluster with different versions older than 20.3, were failed, because replication protocol is incompatible with such old versions, when polymorphic parts are used.
Didn't find any real bugs.

@akuzm AST Fuzzer:
2020-09-18 08:00:36 ERROR 404: Not Found.

@CurtizJ I suspect that this PR has broke integration tests:
eb9ee72 (look at this commit in master)