Do I need a lock here?

[jasiu85]

Hey,

Please take a look at the code of the two threads below:

COMMON_DICT = {}

def thread_1():
global COMMON_DICT
local_dict = prepare_dict()
COMMON_DICT = local_dict

def thread_2():
global COMMON_DICT
local_dict = COMMON_DICT
use_dict(local_dict)

Do I need a lock to protect the COMMON_DICT dictionary? AFAIK bytecode
operations are atomic and in each thread there's only one crucial
bytecode op: STORE_NAME in the first thread and LOAD_NAME in the
second one. So I suspect that everything will work just fine. Am I
right?

Regards,

Mike

[Duncan Booth]

jasiu85 <ja*****@gmail.comwrote:

Hey,

Please take a look at the code of the two threads below:

COMMON_DICT = {}

def thread_1():
global COMMON_DICT
local_dict = prepare_dict()
COMMON_DICT = local_dict

def thread_2():
global COMMON_DICT
local_dict = COMMON_DICT
use_dict(local_dict)

Do I need a lock to protect the COMMON_DICT dictionary? AFAIK bytecode
operations are atomic and in each thread there's only one crucial
bytecode op: STORE_NAME in the first thread and LOAD_NAME in the
second one. So I suspect that everything will work just fine. Am I
right?

Possibly, but without further information it is impossible to tell.

You have one thing wrong though: bytecode operations are not all atomic.
For example STORE_NAME will remove a reference to the object previously
stored under a name and that could trigger code in a __del__ method or a
weak reference callback. That callback code will execute in the same thread
as the STORE_NAME, but while it is executing you could get a context switch
to another thread.

None of that will prevent the store working, and it may not be at all
applicable to your data structures, but in general any operation which can
release complex objects is not thread safe.

--
Duncan Booth http://kupuguy.blogspot.com

[Michael Sparks]

jasiu85 wrote:

Do I need a lock to protect the COMMON_DICT dictionary? AFAIK bytecode
operations are atomic and in each thread there's only one crucial
bytecode op: STORE_NAME in the first thread and LOAD_NAME in the
second one. So I suspect that everything will work just fine. Am I
right?

Will it crash your system? Probably not. Will you end up with correct values
in your dict? Probably not, since you can still end up with race hazards.

If you want to share data that way, you may want to look at Kamaelia's
STM[1] model, which is described on this page:

* http://www.kamaelia.org/STM

You can created stores - which are /essentially/ (key value) threadsafe
stores which enable you to figure out when things go wrong, so you can do
something about it ! :)

A short example:

from Axon.STM import Store

S = Store() # Shared store (like your COMMON_DICT)

D = S.using("account_one", "account_two", "myaccount")
D["account_one"].set(50)
D["account_two"].set(100)
D.commit() # This will fail if any of the values have changed
S.dump()

D = S.using("account_one", "account_two", "myaccount")
D["myaccount"].set(D["account_one"].value+D["account_two"].value)
D["account_one"].set(0)
D["account_two"].set(0)
D.commit() # This will fail if any of the values have changed
S.dump()

If this looks familiar, it's the same basic idiom as version control. The
nice thing of course, is in this version you don't need to worry about
locks, but just be aware that the .commit() may fail, and you may need to
retry.

Two examples - one using bare threads, one using Kamaelia components are
here:
* http://kamaelia.googlecode.com/svn/t.../Examples/STM/

Crucially whilst the above says "this will fail", the examples in that
directory handle the failures correctly :)

This code is included in the latest Kamaelia release, described here:
* http://www.kamaelia.org/GetKamaelia

However you can also use the standalone tarball attached to that page.

[1] ie a minimal software transactional memory implementation. The point of
these things really is to allow you to detect when something has gone
wrong (eg and update fails) and to redo the thing that led to the
update. It's conceptually very similar to version control for variables.
I'm tempted to replace "using" with "checkout" to mirror "commit".

Michael.
--
http://www.kamaelia.org/GetKamaelia