[openstack-dev] [oslo] memoizer aka cache

Joshua Harlow harlowja at yahoo-inc.com
Thu Jan 23 22:54:12 UTC 2014

Sure, no cancelling cases of conscious usage, but we need to be careful
here and make sure its really appropriate. Caching and invalidation
techniques are right up there in terms of problems that appear easy and
simple to initially do/use, but doing it correctly is really really hard
(especially at any type of scale).


On 1/23/14, 1:35 PM, "Renat Akhmerov" <rakhmerov at mirantis.com> wrote:

>On 23 Jan 2014, at 08:41, Joshua Harlow <harlowja at yahoo-inc.com> wrote:
>> So to me memoizing is typically a premature optimization in a lot of
>>cases. And doing it incorrectly leads to overfilling the python
>>processes memory (your global dict will have objects in it that can't be
>>garbage collected, and with enough keys+values being stored will act
>>just like a memory leak; basically it acts as a new GC root object in a
>>way) or more cache invalidation races/inconsistencies than just
>>recomputing the initial valueŠ
>I agree with your concerns here. At the same time, I think this thinking
>shouldn¹t cancel cases of conscious usage of caching technics. A decent
>cache implementation would help to solve lots of performance problems
>(which eventually becomes a concern for any project).
>> Overall though there are a few caching libraries I've seen being used,
>>any of which could be used for memoization.
>> - 
>> - 
>I looked at the code. I have lots of question to the implementation (like
>cache eviction policies, whether or not it works well with green threads,
>but I think it¹s a subject for a separate discussion though). Could you
>please share your experience of using it? Were there specific problems
>that you could point to? May be they are already described somewhere?
>> - dogpile cache @ https://pypi.python.org/pypi/dogpile.cache
>This one looks really interesting in terms of claimed feature set. It
>seems to be compatible with Python 2.7, not sure about 2.6. As above, it
>would be cool you told about your experience with it.
>> I am personally weary of using them for memoization, what expensive
>>method calls do u see the complexity of this being useful? I didn't
>>think that many method calls being done in openstack warranted the
>>complexity added by doing this (premature optimization is the root of
>>all evil...). Do u have data showing where it would be
>I believe there¹s a great deal of use cases like caching db objects or
>more generally caching any heavy objects involving interprocess
>communication. For instance, API clients may be caching objects that are
>known to be immutable on the server side.
>> Sent from my really tiny device...
>>> On Jan 23, 2014, at 8:19 AM, "Shawn Hartsock" <hartsock at acm.org> wrote:
>>> I would like to have us adopt a memoizing caching library of some kind
>>> for use with OpenStack projects. I have no strong preference at this
>>> time and I would like suggestions on what to use.
>>> I have seen a number of patches where people have begun to implement
>>> their own caches in dictionaries. This typically confuses the code and
>>> mixes issues of correctness and performance in code.
>>> Here's an example:
>>> We start with:
>>> def my_thing_method(some_args):
>>>   # do expensive work
>>>   return value
>>> ... but a performance problem is detected... maybe the method is
>>> called 15 times in 10 seconds but then not again for 5 minutes and the
>>> return value can only logically change every minute or two... so we
>>> end up with ...
>>> def my_thing_method(some_args):
>>>   key = key_from(some_args)
>>>    if key in _GLOBAL_THING_CACHE:
>>>        return _GLOBAL_THING_CACHE[key]
>>>    else:
>>>         # do expensive work
>>>         _GLOBAL_THING_CACHE[key] = value
>>>         return value
>>> ... which is all well and good... but now as a maintenance programmer
>>> I need to comprehend the cache mechanism, when cached values are
>>> invalidated, and if I need to debug the "do expensive work" part I
>>> need to tease out some test that prevents the cache from being hit.
>>> Plus I've introduced a new global variable. We love globals right?
>>> I would like us to be able to say:
>>> @memoize(seconds=10)
>>> def my_thing_method(some_args):
>>>   # do expensive work
>>>   return value
>>> ... where we're clearly addressing the performance issue by
>>> introducing a cache and limiting it's possible impact to 10 seconds
>>> which allows for the idea that "do expensive work" has network calls
>>> to systems that may change state outside of this Python process.
>>> I'd like to see this done because I would like to have a place to
>>> point developers to during reviews... to say: use "common/memoizer" or
>>> use "Bob's awesome memoizer" because Bob has worked out all the cache
>>> problems already and you can just use it instead of worrying about
>>> introducing new bugs by building your own cache.
>>> Does this make sense? I'd love to contribute something... but I wanted
>>> to understand why this state of affairs has persisted for a number of
>>> years... is there something I'm missing?
>>> -- 
>>> # Shawn.Hartsock - twitter: @hartsock - plus.google.com/+ShawnHartsock
>>> _______________________________________________
>>> OpenStack-dev mailing list
>>> OpenStack-dev at lists.openstack.org
>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>> _______________________________________________
>> OpenStack-dev mailing list
>> OpenStack-dev at lists.openstack.org
>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>OpenStack-dev mailing list
>OpenStack-dev at lists.openstack.org

More information about the OpenStack-dev mailing list