Cache backends and thread locals

Previous Topic Next Topic
 
classic Classic list List threaded Threaded
3 messages Options
Reply | Threaded
Open this post in threaded view
|

Cache backends and thread locals

Nicolas Le Manchet
Hello,

Thanks to commit f02dbbe1ae[0], the Memcached cache backend using pylibmc can now keep connections open between requests. Establishing a new TCP connection is rather expensive and each saved round trips to the cache server allows to shave a few ms of response time.

It appears that in a multithreaded environment we could improve the situation even more by letting threads share the same `PyLibMCCache` instance.

Currently in Django, each thread asking for a cache backend gets its own personal Backend object[1], thus each thread also get its own connection pool to memcached. After a few requests the process ends up opening as many connections to memcached as there are threads.

If instead we allowed the connection pool to be shared between threads, connections would only be opened when necessary (other threads using all the pooled connections).

Now the important questions. Why do we have thread locals in the first place? Can we share Backend instances between threads?

After looking at the code of all cache backends I feel that nothing prevents dropping threadlocal altogether. Did I miss something?

[0] https://github.com/django/django/commit/f02dbbe1ae02c3258fced7b7a75d35d7745cc02a
[1] https://github.com/django/django/blob/master/django/core/cache/__init__.py#L64
--
Nicolas Le Manchet

--
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 [hidden email].
To post to this group, send email to [hidden email].
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/1520452437.1399418.1295099912.51D6677F%40webmail.messagingengine.com.
For more options, visit https://groups.google.com/d/optout.
Reply | Threaded
Open this post in threaded view
|

Re: Cache backends and thread locals

Curtis Maloney-2
I believe I'm (at least partly) responsible for this change, and the short answer is basically "historically memcache client libs have been awful with threading".

This may have improved, but in order to avoid a whole slew of potential problems (with memcache and other backbends) I copied the pattern used by database connections.

That said, I like your idea of sharing the connection pool between threads. It would mean at worst we'd wind up where we are.

--
Curtis

On 8 March 2018 06:53:57 GMT+11:00, Nicolas Le Manchet <[hidden email]> wrote:
Hello,

Thanks to commit f02dbbe1ae[0], the Memcached cache backend using pylibmc can now keep connections open between requests. Establishing a new TCP connection is rather expensive and each saved round trips to the cache server allows to shave a few ms of response time.

It appears that in a multithreaded environment we could improve the situation even more by letting threads share the same `PyLibMCCache` instance.

Currently in Django, each thread asking for a cache backend gets its own personal Backend object[1], thus each thread also get its own connection pool to memcached. After a few requests the process ends up opening as many connections to memcached as there are threads.

If instead we allowed the connection pool to be shared between threads, connections would only be opened when necessary (other threads using all the pooled connections).

Now the important questions. Why do we have thread locals in the first place? Can we share Backend instances between threads?

After looking at the code of all cache backends I feel that nothing prevents dropping threadlocal altogether. Did I miss something?

[0] https://github.com/django/django/commit/f02dbbe1ae02c3258fced7b7a75d35d7745cc02a
[1] https://github.com/django/django/blob/master/django/core/cache/__init__.py#L64

--
Sent from my Android device with K-9 Mail. Please excuse my brevity.

--
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 [hidden email].
To post to this group, send email to [hidden email].
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/7D6EADB0-F6D3-410F-905D-DECDAC8D975D%40tinbrain.net.
For more options, visit https://groups.google.com/d/optout.
Reply | Threaded
Open this post in threaded view
|

Re: Cache backends and thread locals

Nicolas Le Manchet
I've been checking this a bit more and my conclusion is that it is definitely possible to provide thread-safety for libraries that are not safe out of the box while still allowing thread-safe ones to make use of connection pools.

Now the bad news is that none of the two blessed memcached libraries (python-memcached and pylibmc) provide a good connection pool. This change would have marginal performance impact on them.

Nevertheless I believe that moving the  thread-safety responsibility from Django to the specific cache Backend class is sound. It would allow third-party backends to easily leverage connection pools, as right now they have to hack it: https://github.com/niwinz/django-redis/blob/6d7bfec1e29480593eb9afe84b4c94ae5586c30f/django_redis/pool.py#L9

If you guys think that this is moving is the right direction I can propose a patch.


On Thu, Mar 8, 2018, at 00:06, Curtis Maloney wrote:

> I believe I'm (at least partly) responsible for this change, and the
> short answer is basically "historically memcache client libs have been
> awful with threading".
>
> This may have improved, but in order to avoid a whole slew of potential
> problems (with memcache and other backbends) I copied the pattern used
> by database connections.
>
> That said, I like your idea of sharing the connection pool between
> threads. It would mean at worst we'd wind up where we are.
>
> --
> Curtis
>
> On 8 March 2018 06:53:57 GMT+11:00, Nicolas Le Manchet
> <[hidden email]> wrote:
> >Hello,
> >
> >Thanks to commit f02dbbe1ae[0], the Memcached cache backend using
> >pylibmc can now keep connections open between requests. Establishing a
> >new TCP connection is rather expensive and each saved round trips to
> >the cache server allows to shave a few ms of response time.
> >
> >It appears that in a multithreaded environment we could improve the
> >situation even more by letting threads share the same `PyLibMCCache`
> >instance.
> >
> >Currently in Django, each thread asking for a cache backend gets its
> >own personal Backend object[1], thus each thread also get its own
> >connection pool to memcached. After a few requests the process ends up
> >opening as many connections to memcached as there are threads.
> >
> >If instead we allowed the connection pool to be shared between threads,
> >connections would only be opened when necessary (other threads using
> >all the pooled connections).
> >
> >Now the important questions. Why do we have thread locals in the first
> >place? Can we share Backend instances between threads?
> >
> >After looking at the code of all cache backends I feel that nothing
> >prevents dropping threadlocal altogether. Did I miss something?
> >
> >[0]
> >https://github.com/django/django/commit/f02dbbe1ae02c3258fced7b7a75d35d7745cc02a
> >[1]
> >https://github.com/django/django/blob/master/django/core/cache/__init__.py#L64
> >--
> >Nicolas Le Manchet
> >
> >--
> >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 [hidden email].
> >To post to this group, send email to
> >[hidden email].
> >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/1520452437.1399418.1295099912.51D6677F%40webmail.messagingengine.com.
> >For more options, visit https://groups.google.com/d/optout.
>
> --
> Sent from my Android device with K-9 Mail. Please excuse my brevity.
>
> --
> 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 [hidden email].
> To post to this group, send email to [hidden email].
> 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/7D6EADB0-F6D3-410F-905D-DECDAC8D975D%40tinbrain.net.
> For more options, visit https://groups.google.com/d/optout.


--
Nicolas Le Manchet

--
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 [hidden email].
To post to this group, send email to [hidden email].
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/1520619348.2592203.1297633040.208917EB%40webmail.messagingengine.com.
For more options, visit https://groups.google.com/d/optout.