0.4 dev
 * On-disk data format has changed to allow billions of keys/rows per
   node instead of only millions
 * Multi-keyspace support
 * Added option to fsync before acknowledging writes
 * Scan all sstables for all queries to avoid situations where
   different types of operation on the same ColumnFamily could
   disagree on what data was present
 * Configurable LRU cache for key lookups
 * Thrift API has changed a _lot_:
    - removed time-sorted CFs; instead, user-defined comparators
      may be defined on the column names, which are now byte arrays.
      Default comparators are provided for UTF8, Bytes, Ascii, Long (i64),
      and UUID types.
    - removed colon-delimited strings in thrift api in favor of explicit
      structs such as ColumnPath, ColumnParent, etc.  Also normalized
      thrift struct and argument naming.
    - Added columnFamily argument to get_key_range.
    - Change signature of get_slice and get_slice_super to accept
      starting and ending columns as well as an offset.  (This allows use
      of indexes.)  Added "ascending" flag to allow reasonably-efficient
      reverse scans as well.  Removed get_slice_by_range as redundant.
    - Similarly, changed signature of get_slice_super.
    - get_key_range operates on one CF at a time
    - changed `block` boolean on insert methods to ConsistencyLevel enum,
      with options of NONE, ONE, QUORUM, and ALL.
    - added similar consistency_level parameter to read methods
    - column-name-set slice with no names given now returns zero columns
      instead of all of them.  ("all" can run your server out of memory.
      use a range-based slice with a high max column count instead.)
 * Removed the web interface. Node information can now be obtained by 
   using the newly introduced nodeprobe utility.
 * More JMX stats
 * Remove magic values from internals (e.g. special key to indicate
   when to flush memtables)
 * Rename configuration "table" to "keyspace"
 * Moved to crash-only design; no more shutdown (just kill the process)
 * Lots of bug fixes

Full list of issues resolved in 0.4 is at https://issues.apache.org/jira/secure/IssueNavigator.jspa?reset=true&&pid=12310865&fixfor=12313862&resolution=1&sorter/field=issuekey&sorter/order=DESC


0.3.0 RC3
 * Fix potential deadlock under load in TCPConnection.
   (CASSANDRA-220)


0.3.0 RC2
 * Fix possible data loss when server is stopped after replaying
   log but before new inserts force memtable flush.
   (CASSANDRA-204)
 * Added BUGS file


0.3.0 RC1
 * Range queries on keys, including user-defined key collation
 * Remove support
 * Workarounds for a weird bug in JDK select/register that seems
   particularly common on VM environments. Cassandra should deploy
   fine on EC2 now
 * Much improved infrastructure: the beginnings of a decent test suite
   ("ant test" for unit tests; "nosetests" for system tests), code
   coverage reporting, etc.
 * Expanded node status reporting via JMX
 * Improved error reporting/logging on both server and client
 * Reduced memory footprint in default configuration
 * Combined blocking and non-blocking versions of insert APIs
 * Added FlushPeriodInMinutes configuration parameter to force
   flushing of infrequently-updated ColumnFamilies