ad-hoc subclasses of event-bound classes don't get garbage collected

[Mike Bayer]

but does in 0.8

from sqlalchemy import create_engine
from sqlalchemy.orm import sessionmaker, Session
import weakref
import gc

the_weak_ref = None
engine_is_removed = False

def engine_removed(ref):
    global engine_is_removed
    engine_is_removed = True
    print "Engine was GCed!"

def create_sn(url):
    global the_weak_ref
    engine = create_engine(url)
    the_weak_ref = weakref.ref(engine, engine_removed)
    # this works
    #return Session(bind=engine, expire_on_commit=False)
    return sessionmaker(bind=engine) #(expire_on_commit=False)

sn = create_sn("sqlite://")
sn = sn()
sn.execute("select 1")
sn.close()

print "about to delete sn..."
del sn

print "about to gc.collect()..."
gc.collect()

assert engine_is_removed

[Mike Bayer]

because 0.7 creates a Session subclass, once an event fires off for that subclass it gets placed into the dispatch, which makes it permanent.

[Mike Bayer]

Ah and 0.8 also has this permanent Session subclass, but the Engine isn't stuck in it. So we have a potential memory leak here, if someone is creating and removing lots of sessionmakers.

[Mike Bayer]

the attached patch moves to weakrefs to resolve for 0.8 and 0.7. We get a method hit on two of the low-method perf tests. But this is the correct approach, as just like __subclasses__() we need to use a weakref approach.

backport to 0.7.

[Mike Bayer]

576224edd01f96bb505f9162282cfa9ba6efcb46

973fd4a28a1fb6a11a27997874d57234ae6f384d

[Mike Bayer]

Removing milestone: 0.8.0final (automated comment)