On Thu, Sep 06, 2018 at 05:19:15PM -0300, Fabrízio de Royes Mello wrote: > I also run some similar tests as Jimmy pointed and using PLpgSQL to execute > DDLs and the new consistent behavior is ok. Also I run one session using > DROP SCHEMA at end and after COMMIT the session 2 report 'ERROR: schema > "testschema" does not exist', so avoiding concerns about lock overhead > seems the proposed patch is ok.
Thanks Fabrízio for the review. I think so too, patching the low-level API is I think a proper way to go particularly for back-branches because referencing objects which do not exist at catalog level is a consistency problem. Double-checking for the callers of QualifiedNameGetCreationNamespace, CREATE STATISTICS could be called with CREATE TABLE LIKE, where it would not get called but the table reference blocks the schema drop. I am thinking about adding more tests to cover all the callers of QualifiedNameGetCreationNamespace with: - range type - domain - enum type - statictics - text search - composite type This way if any refactoring is done with this routine, then we don't break schema lock logic. Andres, Tom and others, any objections? -- Michael
signature.asc
Description: PGP signature
