I am very new to Cassandra so I certainly don't have any credibility in commenting on implementation
(FYI...reading stuff from you and other committers like Eric Evans is certainly helping...cheers)

However, having hacked on top of DBMS's and custom in memory solutions I think we can probably look at the following:
1. Obviously the first is availability. Thank you CAP Theorem...duh! I feel like I need to write something deeper to maintain my credibility but instead I'll just concur with "devil in details" and move on until I understand Cassandra better to provide even a smidgen of help!
2. Build very specialized support for specific use-cases e.g. counters that can be used in applications like mine where we just want to perform an action when a threshold is exceeded. In my case we can maybe tolerate a client being slightly out of date i.e. we can handle stale reads for decision making, but also want to (eventually) capture every increment i.e. the operation has to be INCR(column, incr_value) rather than "SET(column)=new_value". Obviously this is easier since the increments can be replayed in any order. 

I noticed some similar comments and thoughts about common lock free approaches that you've probably already reviewed: http://www.mail-archive.com/cassandra-user@.../msg01102.html

My justification for wanting some of this is really:
1. I want to minimize the number of database systems I use from a single application. Naturally, I really need a lot of what Cassandra offers today to get the availability and "big data" support needed in my application so maximizing its use is a good thing as long as we avoid the "square peg, round hole".
2. As a rule of thumb, I am not a big fan of trying to solve the difficult problems at the expense of providing useful features that are easier to implement in a shorter time frame. 

Is there a place on the Cassandra wiki where the proposals/thinking on these issues has been captured in one place?
p.s. I don't want to go off on a tangent, but out of interest, given the original Dynamo article and comments in posts like this: http://www.allthingsdistributed.com/2008/12/eventually_consistent.html, what makes you think SimpleDB is not based on Dynamo? 
On Sat, Mar 6, 2010 at 9:19 AM, Jonathan Ellis-3 [via [hidden email]] <[hidden email]> wrote:
First, SimpleDB is probably not built on Dynamo.

And the devil is in the details.  I haven't seen anyone propose a
reasonable model for how Conditional Puts work (that is the tough

On Sat, Mar 6, 2010 at 8:11 AM, simon.reavely <[hidden email]> wrote:

> Werner Vogels had a recent post around Amazon's support for primitives in
> SimpleDB that can be used to build counters. Given the historical influences
> from Amazon
> s Dynamo to Cassandra I would think a similar approach might work well.
> http://www.allthingsdistributed.com/2010/02/strong_consistency_simpledb.html
> BTW...I would be VERY interested in such support.
> --
> View this message in context: http://n2.nabble.com/Incr-Decr-Counters-in-Cassandra-tp3948361p4686353.html
> Sent from the [hidden email] mailing list archive at Nabble.com.

Simon Reavely

View this message in context: Re: Incr/Decr Counters in Cassandra
Sent from the cassandra-user@incubator.apache.org mailing list archive at Nabble.com.