Change Local Memory Cache to Use LRU #28977
Hi all-- Long time user, first time poster, here. Thank you all for Django! The current local memory cache (locmem) in Django uses a pseudo-random culling strategy. Rather than random, the OrderedDict data type can be used to implement an LRU eviction policy. A prototype implementation is already used by functools.lru_cache and Python 3 now supports OrderedDict.move_to_end and OrderedDict.popitem to ease the implementation. I have created an example set of changes at https://github.com/grantjenks/django/tree/ticket_28977 in commit https://github.com/grantjenks/django/commit/b06574f6713d4b7d367d7a11e0268fb62f5fd1d1 Is there a consensus as to the value of these changes? Sincerely, Grant Jenks -- You received this message because you are subscribed to the Google Groups "Django developers (Contributions to Django itself)" group. To unsubscribe from this group and stop receiving emails from it, send an email to django-developers+unsubscr...@googlegroups.com. To post to this group, send email to django-developers@googlegroups.com. Visit this group at https://groups.google.com/group/django-developers. To view this discussion on the web visit https://groups.google.com/d/msgid/django-developers/1f9d2225-18e8-46f3-9311-b07177c4baca%40googlegroups.com. For more options, visit https://groups.google.com/d/optout.
Re: Change Local Memory Cache to Use LRU #28977
Josh, it's nice to meet you here. I cited your django-lrucache-backend project in the original post of the Trac ticket. I'm also the author of DiskCache http://www.grantjenks.com/docs/diskcache/ which your project refs for benchmarks :) I added some benchmark data to the Trac ticket which may interest you. Thank you, Josh and Adam, for the +1's. I've summarized the feedback and updated the ticket at https://code.djangoproject.com/ticket/28977 It's now marked as Accepted and Has patch. I've created a pull request at https://github.com/django/django/pull/9555 Please review when able. Grant Ps. As always the docs are excellent. Even the docs for contributing are extremely helpful! On Friday, January 5, 2018 at 1:46:38 PM UTC-8, Adam Johnson wrote: > > I'm +1 for moving to LRU too, the eviction algorithm has always looked > weird to me. And Josh's library shows there are valid uses of local memory > caching in applications - perhaps moreso these days than when Django added > caching and memcached was the latest thing. > > >> You can also get a very nice bump in throughput if you eliminate the >> `validate_key` check > > > +1 to adding an option to disable the check as well. If you're using a > LocMemCache in production, you probably don't care about compatibility with > memcached, because you'll be using it for different types of data. > > On 5 January 2018 at 02:53, Josh Smeaton > wrote: > >> To lend some weight to this, I've implemented an LRU loc mem cache and >> have done some benchmarking. There are some graphs in the readme: >> https://github.com/kogan/django-lrucache-backend - which I've written a >> little about >> https://devblog.kogan.com/blog/a-smarter-local-memory-django-cache-backend >> (I don't think my particular implementation is useful to core, as it relies >> on a 3rd party C lib, but using OrderedDict is cool!). >> >> You can also get a very nice bump in throughput if you eliminate the >> `validate_key` check, which does a character by character check of the key >> to avoid issues with particular characters in memcache. >> >> We don't want to be promoting locmemcache too much, but that said, if we >> can provide a better default then we should in my opinion. >> >> On Friday, 5 January 2018 10:12:39 UTC+11, Grant Jenks wrote: >>> >>> Hi all-- >>> >>> Long time user, first time poster, here. Thank you all for Django! >>> >>> The current local memory cache (locmem) in Django uses a pseudo-random >>> culling strategy. Rather than random, the OrderedDict data type can be used >>> to implement an LRU eviction policy. A prototype implementation is already >>> used by functools.lru_cache and Python 3 now supports >>> OrderedDict.move_to_end and OrderedDict.popitem to ease the implementation. >>> >>> I have created an example set of changes at >>> https://github.com/grantjenks/django/tree/ticket_28977 in commit >>> https://github.com/grantjenks/django/commit/b06574f6713d4b7d367d7a11e0268fb62f5fd1d1 >>> >>> Is there a consensus as to the value of these changes? >>> >>> Sincerely, >>> Grant Jenks >>> >>> -- >> You received this message because you are subscribed to the Google Groups >> "Django developers (Contributions to Django itself)" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to django-develop...@googlegroups.com . >> To post to this group, send email to django-d...@googlegroups.com >> . >> Visit this group at https://groups.google.com/group/django-developers. >> To view this discussion on the web visit >> https://groups.google.com/d/msgid/django-developers/fdb91901-c378-4258-9201-d24a9f5f103e%40googlegroups.com >> >> <https://groups.google.com/d/msgid/django-developers/fdb91901-c378-4258-9201-d24a9f5f103e%40googlegroups.com?utm_medium=email&utm_source=footer> >> . >> >> For more options, visit https://groups.google.com/d/optout. >> > > > > -- > Adam > -- You received this message because you are subscribed to the Google Groups "Django developers (Contributions to Django itself)" group. To unsubscribe from this group and stop receiving emails from it, send an email to django-developers+unsubscr...@googlegroups.com. To post to this group, send email to django-developers@googlegroups.com. Visit this group at https://groups.google.com/group/django-developers. To view this discussion on the web visit https://groups.google.com/d/msgid/django-developers/460cee35-9eb6-426a-8847-27ee398fcaea%40googlegroups.com. For more options, visit https://groups.google.com/d/optout.
Re: Change Local Memory Cache to Use LRU #28977
I was able to run the more extensive benchmarks under no-contention and high-contention scenarios with measurements at the 50th, 90th, 99th, and 100th percentiles. I updated the ticket at https://code.djangoproject.com/ticket/28977 with the results. Under high-contention scenarios, the RWLock did serve a purpose, albeit a rather limited one. There's a small tradeoff to be had but I think it's easy to accept. Note that the benchmark does not include a miss-rate penalty which the LRU-eviction policy will almost certainly improve compared with the current random-eviction policy. Grant On Mon, Jan 8, 2018 at 2:25 PM, Josh Smeaton wrote: > Nice to meet you too. Your benchmarking code was extremely handy when I > was profiling lru-cache-backend, so thank you! > > Are you able to run the same benchmarks using this version of the cache to > see how it performs in low/medium/high eviction scenarios? I think those > benchmarks will be nicer than the simpler cache.get() ones you have on the > ticket. > > I'll get to reviewing that change some time this week regardless. > > > > > On Tuesday, 9 January 2018 08:58:55 UTC+11, Grant Jenks wrote: >> >> Josh, it's nice to meet you here. I cited your django-lrucache-backend >> project in the original post of the Trac ticket. I'm also the author of >> DiskCache http://www.grantjenks.com/docs/diskcache/ which your project >> refs for benchmarks :) I added some benchmark data to the Trac ticket which >> may interest you. >> >> Thank you, Josh and Adam, for the +1's. I've summarized the feedback and >> updated the ticket at https://code.djangoproject.com/ticket/28977 It's >> now marked as Accepted and Has patch. I've created a pull request at >> https://github.com/django/django/pull/9555 Please review when able. >> >> Grant >> >> Ps. As always the docs are excellent. Even the docs for contributing are >> extremely helpful! >> >> >> On Friday, January 5, 2018 at 1:46:38 PM UTC-8, Adam Johnson wrote: >>> >>> I'm +1 for moving to LRU too, the eviction algorithm has always looked >>> weird to me. And Josh's library shows there are valid uses of local memory >>> caching in applications - perhaps moreso these days than when Django added >>> caching and memcached was the latest thing. >>> >>> >>>> You can also get a very nice bump in throughput if you eliminate the >>>> `validate_key` check >>> >>> >>> +1 to adding an option to disable the check as well. If you're using a >>> LocMemCache in production, you probably don't care about compatibility with >>> memcached, because you'll be using it for different types of data. >>> >>> On 5 January 2018 at 02:53, Josh Smeaton wrote: >>> >>>> To lend some weight to this, I've implemented an LRU loc mem cache and >>>> have done some benchmarking. There are some graphs in the readme: >>>> https://github.com/kogan/django-lrucache-backend - which I've written >>>> a little about https://devblog.kogan.com/blog >>>> /a-smarter-local-memory-django-cache-backend (I don't think my >>>> particular implementation is useful to core, as it relies on a 3rd party C >>>> lib, but using OrderedDict is cool!). >>>> >>>> You can also get a very nice bump in throughput if you eliminate the >>>> `validate_key` check, which does a character by character check of the key >>>> to avoid issues with particular characters in memcache. >>>> >>>> We don't want to be promoting locmemcache too much, but that said, if >>>> we can provide a better default then we should in my opinion. >>>> >>>> On Friday, 5 January 2018 10:12:39 UTC+11, Grant Jenks wrote: >>>>> >>>>> Hi all-- >>>>> >>>>> Long time user, first time poster, here. Thank you all for Django! >>>>> >>>>> The current local memory cache (locmem) in Django uses a pseudo-random >>>>> culling strategy. Rather than random, the OrderedDict data type can be >>>>> used >>>>> to implement an LRU eviction policy. A prototype implementation is already >>>>> used by functools.lru_cache and Python 3 now supports >>>>> OrderedDict.move_to_end and OrderedDict.popitem to ease the >>>>> implementation. >>>>> >>>>> I have created an example set of changes at >>>>> https://github.com/grantjenks/django/tree/ticket_2
