[asterisk-dev] Review Request: Sorcery Caching

[Joshua Colp]

Scott Griepentrog wrote:
> On the subject of maxiumum cache size - I like being able to limit the
> memory usage, but there is apparently no way to prioritize often used
> entries.  I'm thinking of an example where my cache is full, having gone
> through every entry, but then a small subset of entries is frequently
> read which may not have been the first into the cache.  If I understand
> the wording of the wiki article correctly, these would then not benefit
> from caching, since they would not be allowed in the cache?

Correct. That approach was based around the concept where you know 
you'll never have more than 'n' number of objects. Set the limit, allow 
it to populate, it'll just work.

>
> Also, I'm unclear on the differences between the three object
> expire/lifetime values.  Can you add more descriptive details to each of
> these values, possibly give an example showing why it is beneficial?

Will do!

> My thoughts on cache needs are this:
>
> 1) I want to be able to set the maximum memory dedicated to caching,
> either in MB or in # of entries, but have the caching algorithm take
> into account the most recently read entries when determining what to
> hold onto and what to throw away.  Something I've just looked up is more
> likely to used again than something that has been sitting in the cache
> for a while.  To accomplish this each entry needs a timestamp of the
> last cache hit (when it was last used/read).

The con of that is you have to protect the object when on retrieval, 
introducing a slight contention point for that specific object.

> 2) I want to be able to set two maximum time in cache values: The first
> is the seconds after which the entry is stale, and the second where it
> is guarunteed to be removed from the cache and no longer used.  Both
> values are from the time the entry was last obtained from the source
> (db), requiring a timestamp on entry creation or when it was last
> refreshed from the source.  Before the entry is stale, it can be re-used
> without any further action needed.  When a stale entry is accessed (the
> first time) it should trigger a background refresh to ensure that
> changes have not been made, but continue to return a hit from the cache
> until the second timeout.  Background refreshes to be initiated by a
> separate thread only when there are no primary requests active.  The
> idea here is I can set 300 seconds for stale, and 600 for lifetime --
> for a frequently used entry I'm only refreshing from the database every
> 2.5 minutes and 100% of the requests come immediately from the cache
> with no database lookup delay to the requesting thread.  For any entry
> that hasn't been read in 5 minutes, it's deleted so the memory
> utilization goes away also.

Hrm, without sorcery architectural changes it would be rather difficult 
to implement. I'll look at it.

>
> 3) Flushing all entries on a periodic interval doesn't sound like an
> option I'd be interested in -- assuming that a maximum time in cache is
> enforced properly, this seems like an odd duplication.  My worry is that
> dumping all the entries will occasionally occur in the middle of a burst
> of reads, and thus significantly slow things down.

Understood.

> 4) The external notification should be able to forcibly empty the cache,
> or optionally instead just force all objects to be stale without
> removing them.  This would require adding a stale flag to the entry.
>
> 5) The rebalancing option sounds like a good idea, but since it is such
> an implementation specific value, it would be useful to run some tests
> and document some statistics on what conditions are needed for it to go
> from negligable to noticeable improvement, so that people have an idea
> of when and how to use it.

I have some experience with that actually when I added support for it to 
the res_sorcery_config module. You can reduce lookup times quite a lot 
depending on the number of objects. Could definitely be experimented 
with, though.

-- 
Joshua Colp
Digium, Inc. | Senior Software Developer
445 Jan Davis Drive NW - Huntsville, AL 35806 - US
Check us out at: www.digium.com & www.asterisk.org