[MutableFire]

Add view_log extension that logs all references to database views during
query analysis. It hooks into post_parse_analyze_hook to inspect the range
table and subqueries, writing a CSV entry with timestamp, user, session ID,
and schema-qualified view name to pg_log/view.log on the master node.
Logging is presence-based (before view expansion) and does not guarantee
data was actually fetched.

[andr-sokolov]

What about using a standard walker function from PostgreSQL core?

[MutableFire]

query_tree_walker crashes with unrecognized node type: 705 when it encounters specific fields inside the Query struct. The core walker in nodeFuncs.c does not know how to descend into these specific node types and aborts the query with an error, which breaks CREATE VIEW and every subsequent test. Therefore we cannot use query_tree_walker here. We have to enumerate the relevant standard fields manually (targetList, jointree, havingQual, limitCount, limitOffset, cteList, rtable) and walk each of them with expression_tree_walker, which only touches standard expression nodes and never hits the specific ones.

[Alena0704]

You should also check cteList, JoinExpr conditions, limitCount, and other fields. For example both queries below execute without logging the reference to v:

WITH c AS (SELECT * FROM v) SELECT * FROM c;
SELECT 1 LIMIT (SELECT id FROM v);

parse_sub_analyze() does not invoke the hook separately, maybe you should use query_tree_walker() to cover the relevant fields instead of enumerating them manually?

[MutableFire]

query_tree_walker crashes with unrecognized node type: 705 when it encounters specific fields inside the Query struct. The core walker in nodeFuncs.c does not know how to descend into these specific node types and aborts the query with an error, which breaks CREATE VIEW and every subsequent test. Therefore we cannot use query_tree_walker here. We have to enumerate the relevant standard fields manually (targetList, jointree, havingQual, limitCount, limitOffset, cteList, rtable) and walk each of them with expression_tree_walker, which only touches standard expression nodes and never hits the specific ones.

[Alena0704]

what about this case?
SELECT (SELECT id FROM v) IN (SELECT 1);
SubLink handling skips its testexpr. The walker returns immediately after traversing subselect, skipping the left-hand expression of IN/ANY/ALL.

[MutableFire]

Done.

[Alena0704]

What about this case?

postgres=# CREATE OR REPLACE VIEW "comma,view" AS SELECT 1;
CREATE VIEW
postgres=# SELECT * FROM "comma,view";
 ?column? 
----------
        1
(1 row)

A valid view name such as "comma,view" produces ...,public.comma,view, resulting in five fields instead of four.

[MutableFire]

Done.

[Vlasdislav]

Why will happen if write isn't success?

https://pubs.opengroup.org/onlinepubs/9799919799/functions/write.html

[Vlasdislav]

Did you consider all variants subqueries?

For example what happen if subqueries in functions?

[andr-sokolov]

Can we remove ignore?

Suggested change

	-- start_ignore
	SELECT pg_file_unlink('pg_log/view.log');
	-- end_ignore
	SELECT pg_file_unlink('pg_log/view.log');

[Vlasdislav]

Do we need this code?

Suggested change

	-- start_ignore
	SELECT pg_file_unlink('pg_log/view.log');
	-- end_ignore

[andr-sokolov]

Yes, we do. At the end of a test we should bring DB in the same state as before the test. So the file should be deleted to avoid impact on other tests

[Vlasdislav]

The file deleted in last block used: https://github.com/open-gpdb/gpdb/pull/434/changes/BASE..83f75054d7f12293468c49bc0b7a8161e18e97a3#diff-40354da305a822f20aa639b54810286931f86085e0d444f147d867e65fe2d441R108

And the file deleted each run test: https://github.com/open-gpdb/gpdb/pull/434/changes/BASE..83f75054d7f12293468c49bc0b7a8161e18e97a3#diff-40354da305a822f20aa639b54810286931f86085e0d444f147d867e65fe2d441R4