Large lists

[Ian Kallen]

In a case like this, I don't trust a cache miss to indicate presence
or absence of state. I prefer to keep separate key spaces: one for the
positive case and one for the negative. Keep a representation of The
Truth in media with a stronger guarantee of data integrity [1]; as
states change, write through to memcached when recording The Truth and
upon cache misses, consult The Truth.


[1] This could be an RDBMS, a Berkeley DB or something else. Which one
is correct depends on your scale, performance and concurrency
requirements.

On 11/5/07, K J <sanbat at gmail.com> wrote:
>
> > However, memcached semantics don't quite give you what you want.
> > Depending on whether you can reasonably get a configuration to do what
> > you want, it might be easier to think of memcached as a bloom filter
> > than as a set in this case.  That is, if you negatively cache things
> > that *aren't* part of your list, then the presence of a key will tell
> > you for certain that a particular key is not a member, but the absence
> > of a key would mean that you don't know (or perhaps memcached *did*
> > know, but had since forgotten).
>
>
> I had thought of doing a bloom filter on it as well.  The problem here is,
> the membership list might change sometimes, and reading info on bloom
> filters, it's not really well-suited for dynamically changing lists?
>
> > You could optionally preload objects that are likely to be used if
> > you think the natural population wouldn't do it effectively (you can
> > measure this with stats).
> >
>
> Suppose I cache 10,000 recently-logged in members.  Also, suppose 50% of
> traffic actually come from these users.  Then, this cache would have a high
> hit ratio when testing for membership.
>
> However, what about the non-members?  For instance let's say 40% of the
> traffic come from non-members.  This would mean there'd need to be a full
> listing of members to check?
>
> Hmmm an interesting thought did just come across my mind.  Let me hear your
> thoughts:
>
> Cache 10,000 most recently logged-in members
> Bloom filter on the entire list
> This way, you can test for negatives (bloom filter), and if there's a
> positive, check the 10,000 most recently logged-in users.  If that still
> yields nothing, then do a database query.  In effect, only a small minority
> of checks would require a trip to the DB.
>
>
>
>
>
> >
> >
> >
> > On Nov 4, 2007, at 23:32, J A wrote:
> >
> > > I have a fairly large members list that I want to keep in memcache.
> > > What I do with this list is query it against particular user IDs to
> > > see if they are a member of that list or not.  If they are they get
> > > certain priviledges.
> > >
> > > The problem is, this list has gotten to the point of saturating the
> > > PHP's memory when fetching the MySQL query the first time.
> > >
> > > Is there a way to do this more effectively, for instance,
> > > partitioning the list into separate smaller lists, grouped by time
> > > of login?  I'm thinking of this, as users who have logged in in the
> > > past 3 months are more likely to be in the list anyway.
> >
> >
> >        It'd be easier to not think of it as a list if you're just testing
> > for membership.  All you want to know is if a particular object is an
> > element of a particular set.  You could do this by key convention if
> > you batch populate the records.
> >
> >        However, memcached semantics don't quite give you what you want.
> > Depending on whether you can reasonably get a configuration to do what
> > you want, it might be easier to think of memcached as a bloom filter
> > than as a set in this case.  That is, if you negatively cache things
> > that *aren't* part of your list, then the presence of a key will tell
> > you for certain that a particular key is not a member, but the absence
> > of a key would mean that you don't know (or perhaps memcached *did*
> > know, but had since forgotten).
> >
> >        You could, of course, record the status either way so as to tell
> the
> > difference between not knowing and knowing whether it's a member or
> > not.  This is probably best suited to your needs.
> >
> >        You could optionally preload objects that are likely to be used if
> > you think the natural population wouldn't do it effectively (you can
> > measure this with stats).
> >
> > --
> > Dustin Sallings
> >
> >
> >
> >
>
>