checksum: do not check inverted index files

[save-my-heart]

MergeTreeDataPartChecksums::checkSize will check file size while loading parts, but inverted index files only have default MergeTreeDataPartChecksum(file_size = 0), these parts will fail to load.

Changelog category (leave one):

• Not for changelog (changelog entry is not required)

Information about CI checks: https://clickhouse.com/docs/en/development/continuous-integration/

[rschu1ze]

Could you please move l. 50+51 down to l. 55? (the exception for inverted indexes is the least important one of the three if's)

[rschu1ze]

Kindly add a reason to the comment in l. 49, e.g. /// Skip inverted index files, these have a default MergeTreeDataPartChecksum with file_size == 0

[rschu1ze]

If you don't mind, I would like to understand how the database can end up with a checksum with file_size == 0 for inverted indexes (e.g. which methods are called in which order). Asking because I never encountered this error.

[rschu1ze]

What would be great if you could add a simple end-to-end stateless test (tests/queries/0_stateless/) which makes sure the situation doesn't occur again.

[save-my-heart]

@rschu1ze I just built clickhouse on latest master branch(commit id: 8e0a2a3), then:

1. create a table with inverted index.
2. insert some data.
3. stop clickhouse server and then start it again.

The data can not be loaded, and there are some error in logs as ...gin_dict has unexpected size: 44 instead of 0. (BAD_SIZE_OF_FILE_IN_DATA_PART)...

[save-my-heart]

@rschu1ze I should put the if at the top of three, because when I:

1. create a table with inverted index.
2. insert some data.
3. detach the previous parts
4. stop clickhouse server and then start again

There are some errors in logs as ...gin_dict doesn't exist. (FILE_DOESNT_EXIST)....
Because detach part will generate an empty part, but without .gin_dict, .gin_post, .gin_seg or .gin_sid files. (so we can also create these files while we generating an empty part?)

[rschu1ze]

Alright, that makes sense. Thanks a ton! I can confirm after local testing that the test indeed turns green with your fix.

[cangyin]

inverted index did introduce some special treatmemts.

• A better alternative to #46344 #46921
• skip hardlinking inverted index files in mutation #47663

[cangyin]

inverted index did introduce some special treatmemts.

• A better alternative to #46344 #46921
• skip hardlinking inverted index files in mutation #47663

potentially another:

• Inverted index files may not be generated but demanded during replicated part send #52560