Best approach for expiring objects

[Caunter, Stefan]

On 2011-01-19, at 2:50 AM, "Tyler Love" <t at tylr.org> wrote:

> I know this is a commonly misunderstood aspect of Varnish, but I think I have the hang of it.
> 
> I am using varnish on a very high traffic website and have yet to find a satisfactory approach to expiring objects from the cache.

You need a reasonable default that balances performance and freshness. 

> 
> My goal is to increase ttls (1 day, maybe even a week), to maximize hit rates, and expire objects when content is changed. This has proven to be more difficult than I anticipated.
> 
> 


Indeed. Admin wants long cache, business wants instant everything. 


> 
> The first approach was to use the "purge" function that varnish gives you. The caveat that defated this approach was that purge is basically a memory leak, that also slows down requests.
> 
> Please correct me if I have a misunderstanding of purge but, when you call purge you add either a url or a regex to a list stored within varnish, and then it checks all incoming requests against this url and makes sure to not serve them from the cache.

True, it is a ban list. We do not get to manage the cache directly, we simply steer the ship. 

> 
> This make sense with an understanding of how objects are stored in varnish, but the name "purge" is less than ideal and even misleading (I read this is being changed in trunk?).
> 

English words convey extra meaning, which unfortunately does not resonate with machines. 

> Second, banned items are not served in grace mode, which is behavior I do not want.
> 
> 

So purge if backend is healthy? Should be possible in vcl.


> 
> The alternative is to set the ttl on an object to 0 (zero). The first caveat was being fully aware of every single resource/url that needs to be called for a given purge/expiration. One big problem area is caching on paginated urls.

Sounds like your app needs to set a header which you can use in recv. 
> 
> Second was the sheer volume of purges I need to send. At the scale I am at, I have to purge approximately 400-800 urls per second (I have multiple instances of varnish running by the way).
> 

Your app needs to communicate with your vcl better. 

> Lastly, purging the objects with different compression (Accept-Encoding: gzip, deflate, etc) in the cache. I now have to multiply the amount of urls for every possible encoding I may or may not use.
> 

Normalize these where possible or again it will get complex. 

> 
> 
> So there you have it, removing objects from varnish is tedious, and full of pitfalls. I understand the architecture of varnish just well enough to know why it works this way, but I still wish it was better.

Again, if you are trying to manage the objects you aren't getting what you should from varnish. Your app and your vcl should get you a nice hit rate. 

> 
> If I am approaching this problem the wrong way, I am more than happy to hear some thoughts.
> 
> Tyler
> ______________________________


Stefan Caunter
Operations
TorstarDigital
416.561.4871