pickle problem

[krustymonkey]

I'm wondering if anyone can help with a workaround for a problem I
currently have. I'm trying to set up a prefork tcp server.
Specifically, I'm setting up a server that forks children and has them
listen on pipes created with os.pipe(). The parent process for the
group starts an inet:tcp server on a given port. In the parent, after
a "socket.accept()", I'm trying to pickle the connection object to
send over an IPC pipe (as stated previously), but I get the following
error:

File "/usr/lib/python2.4/copy_reg.py", line 76, in _reduce_ex
raise TypeError("a class that defines __slots__ without "
TypeError: a class that defines __slots__ without defining
__getstate__ cannot be pickled

Does anyone know of a workaround for this? Maybe my approach to this
is wrong? Any help would be appreciated.

Jay

 

[castironpi]

I'm wondering if anyone can help with a workaround for a problem I
currently have.  I'm trying to set up a prefork tcp server.
Specifically, I'm setting up a server that forks children and has them
listen on pipes created with os.pipe().  The parent process for the
group starts an inet:tcp server on a given port.  In the parent, after
a "socket.accept()", I'm trying to pickle the connection object to
send over an IPC pipe (as stated previously), but I get the following
error:

File "/usr/lib/python2.4/copy_reg.py", line 76, in _reduce_ex
    raise TypeError("a class that defines __slots__ without "
TypeError: a class that defines __slots__ without defining
__getstate__ cannot be pickled

Does anyone know of a workaround for this?  Maybe my approach to this
is wrong?  Any help would be appreciated.

Jay

The group currently holds that pickling socket objects is non-optimal
to the ends of making money. We can tell you good ways to accomplish
what you set out to do all the same. I find it educational; that's
fine in my book. Part of the resistence comes from a resistence to
monarchy. The bottleneck is the starting of a second Python process.

There are a number of objects which can't be pickled, though. If you
can objectively prioritize the next step in pickling sockets, you may
find out sooner when it will come in to Python. But that's the hard
part; speaker recommends against.

We accept your idea is to get a process around the socket barrier.
Can you close the server in Part N, hear the connection in Part N+1,
spawn a new server in N+2, and handle the first one in N+3? Is it
suitable to your minds?

 

[Diez B. Roggisch]

I'm wondering if anyone can help with a workaround for a problem I
currently have. I'm trying to set up a prefork tcp server.
Specifically, I'm setting up a server that forks children and has them
listen on pipes created with os.pipe(). The parent process for the
group starts an inet:tcp server on a given port. In the parent, after
a "socket.accept()", I'm trying to pickle the connection object to
send over an IPC pipe (as stated previously), but I get the following
error:

File "/usr/lib/python2.4/copy_reg.py", line 76, in _reduce_ex
raise TypeError("a class that defines __slots__ without "
TypeError: a class that defines __slots__ without defining
__getstate__ cannot be pickled

Does anyone know of a workaround for this? Maybe my approach to this
is wrong? Any help would be appreciated.

The error-message is pretty clear I'd say. You use slots - so you are
responsible yourself for implementing the pickling-protocol using
__getstate__ and __setstate__. Looking at the pickle-docs should give you
an idea.

But the really easy solution is: do not use slots. They are intended as
memory-consumption optimization technique, *not* as "I want to declare my
attributes explicitly"-mechanism. So - get rid of them and be a happy
pickler.

Diez

 

[krustymonkey]

The error-message is pretty clear I'd say. You use slots - so you are
responsible yourself for implementing the pickling-protocol using
__getstate__ and __setstate__. Looking at the pickle-docs should give you
an idea.

But the really easy solution is: do not use slots. They are intended as
memory-consumption optimization technique, *not* as "I want to declare my
attributes explicitly"-mechanism. So - get rid of them and be a happy
pickler.

Diez

The thing is, I'm not using slots by choice. I'm using the standard
lib "socket" class, which apparently uses slots.

 

[Marc 'BlackJack' Rintsch]

The thing is, I'm not using slots by choice. I'm using the standard
lib "socket" class, which apparently uses slots.

`socket` objects can't be pickled. Not just because of the `__slot__`\s
but because a substantial part of their state lives in the operating
system's space.

Ciao,
Marc 'BlackJack' Rintsch

 

[Hrvoje Niksic]

`socket` objects can't be pickled. Not just because of the
`__slot__`\s but because a substantial part of their state lives in
the operating system's space.

Of course, if it makes sense to pickle sockets in the application, one
is can do so by defining __getstate__ and __setstate__:

class Connection(object):
def __init__(self, host, port):
self.host = host
self.port = port
self.init_sock()

def init_sock(self):
self.sock = socket.socket()
self.sock.connect((host, port))
... init communication ...

def __getstate__(self):
# pickle self as a (host, port) pair
return self.host, self.port

def __setstate__(self, state):
# reinstate self by setting host and port and
# recreating the socket
self.host, self.port = state
self.init_sock()

 

[castironpi]

Of course, if it makes sense to pickle sockets in the application, one
is can do so by defining __getstate__ and __setstate__:

class Connection(object):
    def __init__(self, host, port):
        self.host = host
        self.port = port
        self.init_sock()

    def init_sock(self):
        self.sock = socket.socket()
        self.sock.connect((host, port))
        ... init communication ...

    def __getstate__(self):
        # pickle self as a (host, port) pair
        return self.host, self.port

    def __setstate__(self, state):
        # reinstate self by setting host and port and
        # recreating the socket
        self.host, self.port = state
        self.init_sock()

I, local, am mystified that you'd want to pickle a socket. It's a
live connection, which flies on a pocket dollar. You don't want it on
disk, do you?

If you're running a net buoy through a cluster somewhere, do you want
to drop a server and reconnect? Is Amazon's EC2 up and running?

Certainly no one was talking on the internet. Were you?

I don't think you hit anything but banks surfing the web, and the
American dollar is notoriously stuffy. Pump a wi-fi to a diner,
though, and you're just talking more. Where's Usenet?

 

[Marc 'BlackJack' Rintsch]

Of course, if it makes sense to pickle sockets in the application, one
is can do so by defining __getstate__ and __setstate__:

When does it make sense!?

class Connection(object):
def __init__(self, host, port):
self.host = host
self.port = port
self.init_sock()

def init_sock(self):
self.sock = socket.socket()
self.sock.connect((host, port))
... init communication ...

But if you unpickle it while the original connection is still open it
can't connect. If it's not open anymore, there's no one answering at
host/port anymore or some program that has no idea what this connection
is all about.

Ciao,
Marc 'BlackJack' Rintsch

 

[Hrvoje Niksic]

When does it make sense!?

When recreating the object from on-disk state requires reestablishing
the communication, as the example shows. I'm not saying that doing
that is always a good idea, only that it can be done if/when needed.

But if you unpickle it while the original connection is still open it
can't connect.

Why not? I was thinking "client socket", not "server socket".

 

[krustymonkey]

I, local, am mystified that you'd want to pickle a socket. It's a
live connection, which flies on a pocket dollar. You don't want it on
disk, do you?

If you're running a net buoy through a cluster somewhere, do you want
to drop a server and reconnect? Is Amazon's EC2 up and running?

Certainly no one was talking on the internet. Were you?

I don't think you hit anything but banks surfing the web, and the
American dollar is notoriously stuffy. Pump a wi-fi to a diner,
though, and you're just talking more. Where's Usenet?

The idea is a pre-fork socket server. What I want to do is fork off
some child processes from the parent and use IPC to send the
connection object (via socket.accept()) over a pipe to one of the
preexisting child processes and have said child handle the
transaction. Think Apache, if you want an example of what I'm trying
to do here. This alleviates the process startup cost that occurs when
you fork. If the socket object can't be serialized, is there another
way to go about handling this in python?

 

[castironpi]

The idea is a pre-fork socket server.  What I want to do is fork off
some child processes from the parent and use IPC to send the
connection object (via socket.accept()) over a pipe to one of the
preexisting child processes and have said child handle the
transaction.  Think Apache, if you want an example of what I'm trying
to do here.  This alleviates the process startup cost that occurs when
you fork.  If the socket object can't be serialized, is there another
way to go about handling this in python?- Hide quoted text -

- Show quoted text -

That's out of my expertise. You would have to either hack it on a
router you are running (masquerade), or choose a different family of
sockets. For HTTP, you could issue a 'reroute' return. Just multi-
thread your server process.