Skip materialized view checking if source table does not exist

[MikhailBurdukov]

Changelog category (leave one):

• Improvement

Changelog entry (a user-readable short description of the changes that goes to CHANGELOG.md):

Fixed UNKNOWN_TABLE exception when attaching to a materialized view that has dependent tables that are not available.
This might be useful when trying to restore state from a backup.

Steps to reproduce:

CREATE DATABASE test_db;

CREATE TABLE test_db.table (n Int32, s String) ENGINE MergeTree PARTITION BY n ORDER BY n;

CREATE TABLE test_db.mview_backend (n Int32, n2 Int64) ENGINE MergeTree PARTITION BY n ORDER BY n;

CREATE MATERIALIZED VIEW test_db.mview TO test_db.mview_backend AS SELECT n, n * n AS "n2" FROM test_db.table;

DROP TABLE test_db.table;

DETACH TABLE test_db.mview;

ATTACH TABLE test_db.mview;

Received exception from server (version 22.11.5):
Code: 60. DB::Exception: Received from localhost:9001. DB::Exception: Table test_db.table doesn't exist. (UNKNOWN_TABLE)

Expected behavior: attaching materialized view succeeds regardless of whether it's valid and references existing table, or not. It's important for backup-restore scenarios.

With ordinary view works as expected.

CREATE DATABASE test_db;

CREATE TABLE test_db.table (n Int32, s String) ENGINE MergeTree PARTITION BY n ORDER BY n;

CREATE VIEW test_db.view AS SELECT n, n * n AS "n2" FROM test_db.table;

DROP TABLE test_db.table;

DETACH TABLE test_db.view;

ATTACH TABLE test_db.view;

Ok.

[CLAassistant]

All committers have signed the CLA.

[Alex-Burmak]

Not sure that an additional setting is required. For me this PR a bug fix that makes behaviour of ATTACH query consistent for the both ordinary and materialized views.

[kssenii]

Agree that setting is probably redundant, ok to remove. (Though if not removing it - in my opinion its name should end with _on_attach )

[tavplubix]

We should always skip the check on ATTACH query, the check makes sense for CREATE only

[MikhailBurdukov]

We should always skip the check on ATTACH query, the check makes sense for CREATE only

If I understood you correctly, we should check type compatibility only in cases of CREATE. I have reverted changes and added a simple check that the query is not attached.

[kssenii]

Agree that setting is probably redundant, ok to remove. (Though if not removing it - in my opinion its name should end with _on_attach )

[tavplubix]

A minor suggestion: consider using "early return": https://dev.to/jpswade/return-early-12o5

[kssenii]

AST fuzzer (ubsan) — Received signal 11

#48094