Bug report

Bug description:

This is perhaps an annoyance rather than a bug, but it's nice to avoid creating reference cycles if possible. Practically, this means that the ResourceWarning from an sqlite connection object going out of scope fires from a random irrelevant line - where the cyclic GC happens to run - rather than predictably where the object goes out of scope.

The cycle is between the connection object and a functools_lru_cache_wrapper object, as can be seen like this:

importgcimportsqlite3conn=sqlite3.connect(":memory:")referrers=gc.get_referrers(conn)forobj1ingc.get_referents(conn):forobj2inreferrers:ifobj1isobj2:print(obj1)

If I'm following the relevant C code correctly, it's similar to doing this in Python:

# In Connection.__init__self.statement_cache=lru_cache(maxsize)(self)# In Cursor get_statement_from_cacheself.connection.statement_cache(sql)

I believe the reference cycle could be avoided by doing something like this:

# A function (/staticmethod etc.) not bound to the Connection instancedefcompile_stmt(wr_conn,sql):conn=wr_conn()assertconnisnotNonereturnconn(sql)# In Connection.__init__self.statement_cache=lru_cache(maxsize)(compile_stmt)# In Cursor get_statement_from_cachewr_conn=weakref.ref(self.connection)self.connection.statement_cache(wr_conn,sql)

Or functools.partial could be used to wrap the function so that the weakref is only created once. The significant bit is the connection's cache having only weakrefs back to the connection object. I'm not great at writing C code, but I think I can see that everything required for this is available as C functions.

Thanks!

CPython versions tested on:

3.14

Operating systems tested on:

Linux

Linked PRs

• gh-144345: Clarify subtle behaviors of cmd.Cmd in documentation #144383