The problem I'm working on is very similar to this. I'm working on a
reputation system and we keep a fixed number of day buckets for the
scores. So when new data comes in you need to find out what bucket is
supposed to be used, remove the data in it if you've moved to a new
bucket as the data there would be at least n + 1 days old where n is the
number of days you are keeping, and then store the value. In our case we
are willing to accept that we might occasionally lose a bit of data, as
the data tends to trend towards a "good enough" value quite quickly.
Still it would be cool to know a way to make sure that we really know
that we are safe to "nuke" a bucket shy if always insisting on blocking
writes on a "all" read. This can be very painful if you are replicating
data across datacentres.
On 04/16/2010 11:50 AM, Chris Shorrock wrote:
I'm attempting to come up with a technique for limiting the number of
columns a single key (or super column - doesn't matter too much for
the context of this conversation) may contain at any one time. My
actual use-case is a little too meaty to try to describe so an
alternate use-case of this mechanism could be:
/Construct a twitter-esque feed which maintains a list N tweets.
Tweets (in this system - and in reality I suppose) occur at such a
rate that you want to limit a given users "feed" to N items. You
do not have the ability to store an infinite number of tweets due
to the physical constraints of your hardware./
The "/my first idea/" answer is when a tweet is inserted into the the
feed of a given person, that you then do a count and delete of any
outstanding tweets. In reality you could first count, then (if count
>= N) do a batch mutate for the insertion of the new entry and the
removal of the old. My issue with this approach is that after a
certain point every new entry into the system will incur the removal
of an old entry. The count, once a feed has reached N will always be
>= N on any subsequent queries. Depending on how you index the tweets
you may need to actually do a read instead of count to get the row
identifiers.
My second approach was to utilize a "slot" system where you have a
record stored somewhere that indicates the next slot for insertion.
This can be thought of as a fixed length array where you store the
next insertion point in some other column family. When a new tweet
occurs you retrieve the current "slot" meta-data, insert into that
index, then update the meta-data for the next insertion. My concerns
with this relate around synchronization and losing entries due to
concurrent operations. I'd rather not have to something like ZooKeeper
to synchronize in the application cluster.
I have some other ideas but I'm mostly just spit-balling at this
point. So I thought I'd reach out the collective intelligence of the
group to see if anyone has implemented something similar. Thanks in
advance.
