Question: how does the runtime prevent netref'ed values from being garbage collected?

[sidneycadot]

Hello,

Suppose I have the following server:

import rpyc

class MyService(rpyc.Service):
    def exposed_func(self):
        return {"a": 10, "b": 20}

service = MyService()
server = rpyc.utils.server.ThreadedServer(service, port=20000)
server.start()

When a client queries func(), it receives a netref to the dictionary that is supposed to exist at the server side.

However, at the server side, the dict value has already gone out of scope after exposed_func() returns, and is thus eligible for being garbage collected.

How does the server-side ensure that the value returned by exposed_func() still exists for the netref to be valid? Who takes ownership (and manages the lifetime of) the dict returned by the exposed_func()?

[sidneycadot]

Okay I studied the source code for a bit, how this works is clever...

The Connection (aka Protocol) keeps references to objects that are returned over the link in its _local_objects attribute. These are subject to explicit reference counting, and when the remote side del's its netref, this is propagated over the line, leading to a decrease in the refcount. Once this hits zero, the object reference is removed from the _local_objects collection.

A mildly terrifying control flow for the lifetime management, but it seems to work :-)

Perhaps it is useful to mention this "live netref"-counting behavior in the docs somewhere; it demystifies what's going on.