NEWS.txt

GENERAL UPGRADING ADVICE FOR ANY VERSION
========================================

Snapshotting is fast (especially if you have JNA installed) and takes
effectively zero disk space until you start compacting the live data
files again.  Thus, best practice is to ALWAYS snapshot before any
upgrade, just in case you need to roll back to the previous version.
(Cassandra version X + 1 will always be able to read data files created
by version X, but the inverse is not necessarily the case.)


1.1
===

Upgrading
---------
    - Compression is enabled by default on newly created ColumnFamilies
      (and unchanged for ColumnFamilies created prior to upgrading).
    - The KsDef.replication_factor field (deprecated since 0.8) has
      been removed.  Older clients will need to be updated to be able
      to continue to created and update keyspaces.
    - If you are running a multi datacenter setup, you should upgrade to
      the latest 1.0.x (or 0.8.x) release before upgrading.  Versions
      0.8.8 and 1.0.3-1.0.5 generate cross-dc forwarding that is incompatible
      with 1.1.
    - EACH_QUORUM ConsistencyLevel is only supported for writes and will now
      throw an InvalidRequestException when used for reads.  (Previous
      versions would silently perform a LOCAL_QUORUM read instead.)
    - ANY ConsistencyLevel is only supported for writes and will now
      throw an InvalidRequestException when used for reads.  (Previous
      versions would silently perform a ONE read for range queries;
      single-row and multiget reads already rejected ANY.)
    - The largest mutation batch accepted by the commitlog is now 128MB.  
      (In practice, batches larger than ~10MB always caused poor
      performance due to load volatility and GC promotion failures.)
      Larger batches will continue to be accepted but will not be
      durable.  Consider setting durable_writes=false if you really
      want to use such large batches.
    - Make sure that global settings: key_cache_{size_in_mb, save_period}
      and row_cache_{size_in_mb, save_period} in conf/cassandra.yaml are
      used instead of per-ColumnFamily options.
    - JMX methods no longer return custom Cassandra objects.  Any such methods
      will now return standard Maps, Lists, etc.
    - Hadoop input and output details are now separated.  If you were
      previously using methods such as getRpcPort you now need to use
      getInputRpcPort or getOutputRpcPort depending on the circumstance.
    - CQL changes:
      + Prior to 1.1, you could use KEY as the primary key name in some
        select statements, even if the PK was actually given a different
        name.  In 1.1+ you must use the defined PK name.
        

Features
--------
    - Cassandra 1.1 adds row-level isolation.  Multi-column updates to
      a single row have always been *atomic* (either all will be applied,
      or none) thanks to the CommitLog, but until 1.1 they were not *isolated*
      -- a reader may see mixed old and new values while the update happens.
    - Finer-grained control over data directories, allowing a ColumnFamily to
      be pinned to specfic media.
    - Hadoop: a new BulkOutputFormat is included which will directly write
      SSTables locally and then stream them into the cluster.
      YOU SHOULD USE BulkOutputFormat BY DEFAULT.  ColumnFamilyOutputFormat
      is still around in case for some strange reason you want results
      trickling out over Thrift, but BulkOutputFormat is significantly
      more efficient.
    - Hadoop: KeyRange.filter is now supported with ColumnFamilyInputFormat
    - Hadoop wide row mode added to ColumnFamilyInputFormat
    - The bulk loader is not longer a fat client; it can be run from an
      existing machine in a cluster.
    - A new write survey mode has been added, similar to bootstrap (enabled via
      -Dcassandra.write_survey=true), but the node will not automatically join
      the cluster.  This is useful for cases such as testing different
      compaction strategies with live traffic without affecting the cluster.
    - Key and row caches are now global, similar to the global memtable
      threshold.
    - Off-heap caches no longer require JNA.
    - Streaming is now multithreaded.
    - Compactions may now be aborted via JMX or nodetool.


1.0.8
=====

Upgrading
---------
    - Nothing specific to 1.0.8

Other
-----
    - Allow configuring socket timeout for streaming


1.0.7
=====

Upgrading
---------
    - Nothing specific to 1.0.7, please report to instruction for 1.0.6

Other
-----
    - Adds new setstreamthroughput to nodetool to configure streaming
      throttling
    - Adds JMX property to get/set rpc_timeout_in_ms at runtime
    - Allow configuring (per-CF) bloom_filter_fp_chance


1.0.6
=====

Upgrading
---------
    - This release fixes an issue related to the chunk_length_kb option for
      compressed sstables. If you use compression on some column families, it
      is recommended after the upgrade to check the value for this option on
      these column families (the default value is 64). In case the option would
      not be set correctly, you should update the column family definition,
      setting the right value and then run scrub on the column family.
    - Please report to instruction for 1.0.5 if coming from an older version.


1.0.5
=====

Upgrading
---------
    - 1.0.5 comes to fix two important regression of 1.0.4. So all information
      concerning 1.0.4 are valid for this release, but please avoids upgrading
      to 1.0.4.


1.0.4
=====

Upgrading
---------
    - Nothing specific to 1.0.4 but please see the 1.0 upgrading section if
      upgrading from a version prior to 1.0.0

Features
--------
    - A new upgradesstables command has been added to nodetool. It is very
      similar to scrub but without the ability to discard corrupted rows (and
      as a consequence it does not snapshot automatically before). This new
      command is to be prefered to scrub in all cases where sstables should be
      rewritten to the current format for upgrade purposes.

JMX
---
    - The path for the data, commit log and saved cache directories exposed
      through JMX
    - The in-memory bloom filter sizes are now exposed through JMX


1.0.3
=====

Upgrading
---------
    - Nothing specific to 1.0.3 but please see the 1.0 upgrading section if
      upgrading from a version prior to 1.0.0

Features
--------
    - For non compressed sstables (compressed sstable already include more
      fine grained checsums), a sha1 for the full sstable is now automatically
      created (in a fix with suffix -Digest.sha1). It can be used to check the
      sstable integrity with sha1sum.


1.0.2
=====

Upgrading
---------
    - Nothing specific to 1.0.2 but please see the 1.0 upgrading section if
      upgrading from a version prior to 1.0.0

Features
--------
    - Cassandra CLI queries now have timing information


1.0.1
=====

Upgrading
---------
    - If upgrading from a version prior to 1.0.0, please see the 1.0 Upgrading
      section
    - For running on Windows as a Service, procrun is no longer discributed
      with Cassandra, see README.txt for more information on how to download
      it if necessary.
    - The name given to snapshots directories have been improved for human
      readability. If you had scripts relying on it, you may need to update
      them.


1.0
===

Upgrading
---------
    - Upgrading from version 0.7.1+ or 0.8.2+ can be done with a rolling
      restart, one node at a time.  (0.8.0 or 0.8.1 are NOT network-compatible
      with 1.0: upgrade to the most recent 0.8 release first.)
      You do not need to bring down the whole cluster at once. 
    - After upgrading, run nodetool scrub against each node before running
      repair, moving nodes, or adding new ones.
    - CQL inserts/updates now generate microsecond resolution timestamps
      by default, instead of millisecond. THIS MEANS A ROLLING UPGRADE COULD
      MIX milliseconds and microseconds, with clients talking to servers
      generating milliseconds unable to overwrite the larger microsecond
      timestamps. If you are using CQL and this is important for your
      application, you can either perform a non-rolling upgrade to 1.0, or
      update your application first to use explicit timestamps with the "USING
      timestamp=X" syntax.
    - The BinaryMemtable bulk-load interface has been removed (use the
      sstableloader tool instead).
    - The compaction_thread_priority setting has been removed from
      cassandra.yaml (use compaction_throughput_mb_per_sec to throttle
      compaction instead).
    - CQL types bytea and date were renamed to blob and timestamp, respectively,
      to conform with SQL norms.  CQL type int is now a 4-byte int, not 8
      (which is still available as bigint).
    - Cassandra 1.0 uses arena allocation to reduce old generation
      fragmentation.  This means there is a minimum overhead of 1MB
      per ColumnFamily plus 1MB per index.
    - The SimpleAuthenticator and SimpleAuthority classes have been moved to
      the example directory (and are thus not available from the binary
      distribution). They never provided actual security and in their current
      state are only meant as examples.

Features
--------
    - SSTable compression is supported through the 'compression_options'
      parameter when creating/updating a column family. For instance, you can
      create a column family Cf using compression (through the Snappy library)
      in the CLI with:
        create column family Cf with compression_options={sstable_compression: SnappyCompressor}
      SSTable compression is not activated by default but can be activated or
      deactivated at any time.
    - Compressed SSTable blocks are checksummed to protect against bitrot
    - New LevelDB-inspired compaction algorithm can be enabled by setting the
      Columnfamily compaction_strategy=LeveledCompactionStrategy option.
      Leveled compaction means you only need to keep a few MB of space free for
      compaction instead of (in the worst case) 50%.
    - Ability to use multiple threads during a single compaction. See
      multithreaded_compaction in cassandra.yaml for more details.
    - Windows Service ("cassandra.bat install" to enable)
    - A dead node may be replaced in a single step by starting a new node
      with -Dcassandra.replace_token=<token>. More details can be found at
      http://wiki.apache.org/cassandra/Operations#Replacing_a_Dead_Node
    - It is now possible to repair only the first range returned by the
      partitioner for a node with `nodetool repair -pr`. It makes it
      easier/possible to repair a full cluster without any work duplication by
      running this command on every node of the cluster.

New data types
--------------
    - decimal

Other
-----
    - Hinted Handoff has two major improvements:
        - Hint replay is much more efficient thanks to a change in the data model
        - Hints are created for all replicas that do not ack a write.  (Formerly,
          only replicas known to be down when the write started were hinted.)
      This means that running with read repair completely off is much more
      viable than before, and the default read_repair_chance is reduced from 1.0
      ("always repair") to 0.1 ("repair 10% of the time").
    - The old per-ColumnFamily memtable thresholds
      (memtable_throughput_in_mb, memtable_operations_in_millions,
      memtable_flush_after_mins) are ignored, in favor of the global
      memtable_total_space_in_mb and commitlog_total_space_in_mb settings.
      This does not affect client compatibility -- the old options are
      still allowed, but have no effect. These options may be removed
      entirely in a future release.
    - Backlogged compactions will begin five minutes after startup.  The 0.8
      behavior of never starting compaction until a flush happens is usually
      not what is desired, but a short grace period is useful to allow caches
      to warm up first.
    - The deletion of compacted data files is not performed during Garbage
      Collection anymore. This means compacted files will now be deleted
      without delay.


0.8.5
=====

Features
--------
    - SSTables copied to a data directory can be loaded by a live node through
      nodetool refresh (may be handy to load snapshots).
    - The configured compaction throughput is exposed through JMX.

Other
-----
    - The sstableloader is now bundled with the debian package.
    - Repair detects when a participating node is dead and fails instead of
      hanging forever.


0.8.4
=====

Upgrading
---------
    - Nothing specific to 0.8.4

Other
-----
    - This release comes to fix a bug in counter that could lead to
      (important) over-count.
    - It also fixes a slight upgrade regression from 0.8.3. It is thus advised
      to jump directly to 0.8.4 if upgrading from before 0.8.3.


0.8.3
=====

Upgrading
---------
    - Token removal has been revamped.  Removing tokens in a mixed cluster with
      0.8.3 will not work, so the entire cluster will need to be running 0.8.3
      first, except for the dead node.

Features
--------
    - It is now possible to use thrift asynchronous and
      half-synchronous/half-asynchronous servers (see cassandra.yaml for more
      details).
    - It is now possible to access counter columns through Hadoop.

Other
-----
    - This release fix a regression of 0.8 that can make commit log segment to
      be deleted even though not all data it contains has been flushed.
      Upgrades from 0.8.* is very much encouraged.


0.8.2
=====

Upgrading
---------
    - 0.8.0 and 0.8.1 shipped with a bug that was setting the
      replicate_on_write option for counter column families to false (this
      option has no effect on non-counter column family). This is an unsafe
      default and 0.8.2 correct this, the default for replicate_on_write is
      now true. It is advised to update your counter column family definitions
      if replicate_on_write was uncorrectly set to false (before or after
      upgrade).


0.8.1
=====

Upgrading
---------
    - 0.8.1 is backwards compatible with 0.8, upgrade can be achieved by a
      simple rolling restart.
    - If upgrading for earlier version (0.7), please refer to the 0.8 section
      for instructions.

Features
--------
    - Numerous additions/improvements to CQL (support for counters, TTL, batch
      inserts/deletes, index dropping, ...).
    - Add two new AbstractTypes (comparator) to support compound keys
      (CompositeType and DynamicCompositeType), as well as a ReverseType to
      reverse the order of any existing comparator.
    - New option to bypass the commit log on some keyspaces (for advanced
      users).

Tools
-----
    - Add new data bulk loading utility (sstableloader).


0.8
===

Upgrading
---------
    - Upgrading from version 0.7.1 or later can be done with a rolling
      restart, one node at a time.  You do not need to bring down the
      whole cluster at once. 
    - After upgrading, run nodetool scrub against each node before running
      repair, moving nodes, or adding new ones.
    - Running nodetool drain before shutting down the 0.7 node is
      recommended but not required. (Skipping this will result in
      replay of entire commitlog, so it will take longer to restart but
      is otherwise harmless.)
    - 0.8 is fully API-compatible with 0.7.  You can continue
      to use your 0.7 clients.
    - Avro record classes used in map/reduce and Hadoop streaming code have
      been removed. Map/reduce can be switched to Thrift by changing
      org.apache.cassandra.avro in import statements to 
      org.apache.cassandra.thrift (no class names change). Streaming support 
      has been removed for the time being.
    - The loadbalance command has been removed from nodetool.  For similar
      behavior, decommission then rebootstrap with empty initial_token.
    - Thrift unframed mode has been removed.
    - The addition of key_validation_class means the cli will assume keys
      are bytes, instead of strings, in the absence of other information.
      See http://wiki.apache.org/cassandra/FAQ#cli_keys for more details.


Features
--------
    - added CQL client API and JDBC/DBAPI2-compliant drivers for Java and
      Python, respectively (see: drivers/ subdirectory and doc/cql)
    - added distributed Counters feature; 
      see http://wiki.apache.org/cassandra/Counters
    - optional intranode encryption; see comments around 'encryption_options'
      in cassandra.yaml
    - compaction multithreading and rate-limiting; see 
      'concurrent_compactors' and 'compaction_throughput_mb_per_sec' in
      cassandra.yaml
    - cassandra will limit total memtable memory usage to 1/3 of the heap
      by default.  This can be ajusted or disabled with the 
      memtable_total_space_in_mb option.  The old per-ColumnFamily
      throughput, operations, and age settings are still respected but
      will be removed in a future major release once we are satisfied that
      memtable_total_space_in_mb works adequately.

Tools
-----
    - stress and py_stress moved from contrib/ to tools/
    - clustertool was removed (see 
      https://issues.apache.org/jira/browse/CASSANDRA-2607 for examples
      of how to script nodetool across the cluster instead)

Other
-----
    - In the past, sstable2json would write column names and values as
      hex strings, and now creates human readable values based on the
      comparator/validator.  As a result, JSON dumps created with
      older versions of sstable2json are no longer compatible with
      json2sstable, and imports must be made with a configuration that
      is identical to the export.
    - manually-forced compactions ("nodetool compact") will do nothing
      if only a single SSTable remains for a ColumnFamily. To force it
      to compact that anyway (which will free up space if there are
      a lot of expired tombstones), use the new forceUserDefinedCompaction
      JMX method on CompactionManager.
    - most of contrib/ (which was not part of the binary releases)
      has been moved either to examples/ or tools/. We plan to move the
      rest for 0.8.1.

JMX
---
    - By default, JMX now listens on port 7199.


0.7.6
=====

Upgrading
---------
    - Nothing specific to 0.7.6, but see 0.7.3 Upgrading if upgrading
      from earlier than 0.7.1.


0.7.5
=====

Upgrading
---------
    - Nothing specific to 0.7.5, but see 0.7.3 Upgrading if upgrading
      from earlier than 0.7.1.

Changes
-------
    - system_update_column_family no longer snapshots before applying
      the schema change. (_update_keyspace never did.  _drop_keyspace
      and _drop_column_family continue to snapshot.)
    - added memtable_flush_queue_size option to cassandra.yaml to
      avoid blocking writes when multiple column families (or a colum
      family with indexes) are flushed at the same time.
    - allow overriding initial_token, storage_port and rpc_port using
      system properties


0.7.4
=====

Upgrading
---------
    - Nothing specific to 0.7.4, but see 0.7.3 Upgrading if upgrading
      from earlier than 0.7.1.

Features
--------
    - Output to Pig is now supported as well as input


0.7.3
=====

Upgrading
---------
    - 0.7.1 and 0.7.2 shipped with a bug that caused incorrect row-level
      bloom filters to be generated when compacting sstables generated
      with earlier versions.  This would manifest in IOExceptions during
      column name-based queries.  0.7.3 provides "nodetool scrub" to 
      rebuild sstables with correct bloom filters, with no data lost.
      (If your cluster was never on 0.7.0 or earlier, you don't have to
      worry about this.)  Note that nodetool scrub will snapshot your
      data files before rebuilding, just in case.


0.7.1
=====

Upgrading
---------
    - 0.7.1 is completely backwards compatible with 0.7.0.  Just restart
      each node with the new version, one at a time.  (The cluster does
      not all need to be upgraded simultaneously.)

Features
--------
    - added flush_largest_memtables_at and reduce_cache_sizes_at options
      to cassandra.yaml as an escape valve for memory pressure
    - added option to specify -Dcassandra.join_ring=false on startup
      to allow "warm spare" nodes or performing JMX maintenance before
      joining the ring

Performance
-----------
    - Disk writes and sequential scans avoid polluting page cache
      (requires JNA to be enabled)
    - Cassandra performs writes efficiently across datacenters by
      sending a single copy of the mutation and having the recipient
      forward that to other replicas in its datacenter.
    - Improved network buffering
    - Reduced lock contention on memtable flush
    - Optimized supercolumn deserialization
    - Zero-copy reads from mmapped sstable files
    - Explicitly set higher JVM new generation size
    - Reduced i/o contention during saving of caches


0.7.0
=====

Features
--------
    - Secondary indexes (indexes on column values) are now supported
    - Row size limit increased from 2GB to 2 billion columns.  rows
      are no longer read into memory during compaction.
    - Keyspace and ColumnFamily definitions may be added and modified live
    - Streaming data for repair or node movement no longer requires 
      anticompaction step first
    - NetworkTopologyStrategy (formerly DatacenterShardStrategy) is ready for 
      use, enabling ConsistencyLevel.DCQUORUM and DCQUORUMSYNC.  See comments 
      in `cassandra.yaml.`
    - Optional per-Column time-to-live field allows expiring data without
      have to issue explicit remove commands
    - `truncate` thrift method allows clearing an entire ColumnFamily at once
    - Hadoop OutputFormat and Streaming [non-jvm map/reduce via stdin/out]
      support
    - Up to 8x faster reads from row cache
    - A new ByteOrderedPartitioner supports bytes keys with arbitrary content,
      and orders keys by their byte value.  This should be used in new
      deployments instead of OrderPreservingPartitioner.
    - Optional round-robin scheduling between keyspaces for multitenant
      clusters
    - Dynamic endpoint snitch mitigates the impact of impaired nodes
    - New `IntegerType`, faster than LongType and allows integers of 
      both less and more bits than Long's 64
    - A revamped authentication system that decouples authorization and 
      allows finer-grained control of resources.

Upgrading
---------
    The Thrift API has changed in incompatible ways; see below, and refer
    to http://wiki.apache.org/cassandra/ClientOptions for a list of
    higher-level clients that have been updated to support the 0.7 API.

    The Cassandra inter-node protocol is incompatible with 0.6.x
    releases (and with 0.7 beta1), meaning you will have to bring your
    cluster down prior to upgrading: you cannot mix 0.6 and 0.7 nodes.
    
    The hints schema was changed from 0.6 to 0.7. Cassandra automatically
    snapshots and then truncates the hints column family as part of 
    starting up 0.7 for the first time.

    Keyspace and ColumnFamily definitions are stored in the system
    keyspace, rather than the configuration file.

    The process to upgrade is:
    1) run "nodetool drain" on _each_ 0.6 node.  When drain finishes (log
       message "Node is drained" appears), stop the process.
    2) Convert your storage-conf.xml to the new cassandra.yaml using 
       "bin/config-converter".  
    3) Rename any of your keyspace or column family names that do not adhere
       to the '^\w+' regex convention.
    4) Start up your cluster with the 0.7 version.
    5) Initialize your Keyspace and ColumnFamily definitions using 
       "bin/schematool <host> <jmxport> import".  _You only need to do 
       this to one node_.

Thrift API
----------
    - The Cassandra server now defaults to framed mode, rather than
      unframed.  Unframed is obsolete and will be removed in the next
      major release.
    - The Cassandra Thrift interface file has been updated for Thrift 0.5.
      If you are compiling your own client code from the interface, you
      will need to upgrade the Thrift compiler to match.
    - Row keys are now bytes: keys stored by versions prior to 0.7.0 will be
      returned as UTF-8 encoded bytes. OrderPreservingPartitioner and
      CollatingOrderPreservingPartitioner continue to expect that keys contain
      UTF-8 encoded strings, but RandomPartitioner now works on any key data.
    - keyspace parameters have been replaced with the per-connection
      set_keyspace method.
    - The return type for login() is now AccessLevel.
    - The get_string_property() method has been removed.
    - The get_string_list_property() method has been removed.

Configuraton
------------
    - Configuration file renamed to cassandra.yaml and log4j.properties to
      log4j-server.properties
    - PropertyFileSnitch configuration file renamed to 
      cassandra-topology.properties
    - The ThriftAddress and ThriftPort directives have been renamed to
      RPCAddress and RPCPort respectively.
    - EndPointSnitch was renamed to RackInferringSnitch.  A new SimpleSnitch
      has been added.
    - RackUnawareStrategy and RackAwareStrategy have been renamed to
      SimpleStrategy and OldNetworkTopologyStrategy, respectively.
    - RowWarningThresholdInMB replaced with in_memory_compaction_limit_in_mb
    - GCGraceSeconds is now per-ColumnFamily instead of global
	- Keyspace and column family names that do not confirm to a '^\w+' regex
      are considered illegal.
    - Keyspace and column family definitions will need to be loaded via
      "bin/schematool <host> <jmxport> import".  _You only need to do this to
      one node_.
    - In addition to an authenticator, an authority must be configured as
      well. Users of SimpleAuthenticator should use SimpleAuthority for this
      value (the default is AllowAllAuthority, which corresponds with 
      AllowAllAuthenticator).
    - The format of access.properties has changed, see the sample configuration
      conf/access.properties for documentation on the new format.


JMX
---
    - StreamingService moved from o.a.c.streaming to o.a.c.service
    - GMFD renamed to GOSSIP_STAGE
    - {Min,Mean,Max}RowCompactedSize renamed to {Min,Mean,Max}RowSize
      since it no longer has to wait til compaction to be computed

Other
-----
    - If extending AbstractType, make sure you follow the singleton pattern
      followed by Cassandra core AbstractType classes: provide a public
      static final variable called 'instance'.


0.6.6
=====

Upgrading
---------
    - As part of the cache-saving feature, a third directory
      (along with data and commitlog) has been added to the config
      file.  You will need to set and create this directory
      when restarting your node into 0.6.6.


0.6.1
=====

Upgrading
---------
    - We try to keep minor versions 100% compatible (data format,
      commitlog format, network format) within the major series, but
      we introduced a network-level incompatibility in 0.6.1.
      Thus, if you are upgrading from 0.6.0 to any higher version
      (0.6.1, 0.6.2, etc.) then you will need to restart your entire
      cluster with the new version, instead of being able to do a
      rolling restart.


0.6.0
=====

Features
--------
    - row caching: configure with the RowsCached attribute in
      ColumnFamily definition
    - Hadoop map/reduce support: see contrib/word_count for an example
    - experimental authentication support, described under
      Authenticator in storage.conf

Configuraton
------------
    - MemtableSizeInMB has been replaced by MemtableThroughputInMB which
      triggers a memtable flush when the specified amount of data has 
      been written, including overwrites.
    - MemtableObjectCountInMillions has been replaced by the
      MemtableOperationsInMillions directive which causes a memtable flush
      to occur after the specified number of operations.
    - Like MemtableSizeInMB, BinaryMemtableSizeInMB has been replaced by
      BinaryMemtableThroughputInMB.
    - Replication factor is now per-keyspace, rather than global.
    - KeysCachedFraction is deprecated in favor of KeysCached
    - RowWarningThresholdInMB added, to warn before very large rows
      get big enough to threaten node stability

Thrift API
----------
    - removed deprecated get_key_range method
    - added batch_mutate meethod
    - deprecated multiget and batch_insert methods in favor of
      multiget_slice and batch_mutate, respectively
    - added ConsistencyLevel.ANY, for when you want write
      availability even when it may not be readable immediately.
      Unlike CL.ZERO, though, it will throw an exception if
      it cannot be written *somewhere*.

JMX metrics
-----------
    - read and write statistics are reported as lifetime totals,
      instead of averages over the last minute.  average-since-last
      requested are also available for convenience.
    - cache hit rate statistics are now available from JMX under
      org.apache.cassandra.db.Caches
    - compaction JMX metrics are moved to
      org.apache.cassandra.db.CompactionManager.  PendingTasks is now
      a much better estimate of compactions remaining, and the
      progress of the current compaction has been added.
    - commitlog JMX metrics are moved to org.apache.cassandra.db.Commitlog
    - progress of data streaming during bootstrap, loadbalance, or other
      data migration, is available under 
      org.apache.cassandra.streaming.StreamingService.
      See http://wiki.apache.org/cassandra/Streaming for details.

Installation/Upgrade
--------------------
    - 0.6 network traffic is not compatible with earlier versions.  You
      will need to shut down all your nodes at once, upgrade, then restart.



0.5.0
=====

0. The commitlog format has changed (but sstable format has not). 
   When upgrading from 0.4, empty the commitlog either by running 
   bin/nodeprobe flush on each machine and waiting for the flush to finish,
   or simply remove the commitlog directory if you only have test data.
   (If more writes come in after the flush command, starting 0.5 will error
   out; if that happens, just go back to 0.4 and flush again.)
   The format changed twice: from 0.4 to beta1, and from beta2 to RC1.

.5 The gossip protocol has changed, meaning 0.5 nodes cannot coexist
   in a cluster of 0.4 nodes or vice versa; you must upgrade your
   whole cluster at the same time.

1. Bootstrap, move, load balancing, and active repair have been added.
   See http://wiki.apache.org/cassandra/Operations.  When upgrading
   from 0.4, leave autobootstrap set to false for the first restart
   of your old nodes.

2. Performance improvements across the board, especially on the write
   path (over 100% improvement in stress.py throughput).

3. Configuration:
     - Added "comment" field to ColumnFamily definition.
     - Added MemtableFlushAfterMinutes, a global replacement for the 
       old per-CF FlushPeriodInMinutes setting
     - Key cache settings

4. Thrift:
     - Added get_range_slice, deprecating get_key_range



0.4.2
=====

1. Improve default garbage collector options significantly --
   throughput will be 30% higher or more.



0.4.1
=====

1. SnapshotBeforeCompaction configuration option allows snapshotting
   before each compaction, which allows rolling back to any version
   of the data.



0.4.0
=====

1. On-disk data format has changed to allow billions of keys/rows per
   node instead of only millions.  The new format is incompatible with 0.3;
   see 0.3 notes below for how to import data from a 0.3 install.

2. Cassandra now supports multiple keyspaces.  Typically you will have
   one keyspace per application, allowing applications to be able to
   create and modify ColumnFamilies at will without worrying about
   collisions with others in the same cluster.

3. Many Thrift API changes and documentation.  See 
   http://wiki.apache.org/cassandra/API

4. Removed the web interface in favor of JMX and bin/nodeprobe, which
   has significantly enhanced functionality.

5. Renamed configuration "<Table>" to "<Keyspace>".

6. Added commitlog fsync; see "<CommitLogSync>" in configuration.



0.3.0
=====

1. With enough and large enough keys in a ColumnFamily, Cassandra will
   run out of memory trying to perform compactions (data file merges).
   The size of what is stored in memory is (S + 16) * (N + M) where S
   is the size of the key (usually 2 bytes per character), N is the
   number of keys and M, is the map overhead (which can be guestimated
   at around 32 bytes per key).
   So, if you have 10-character keys and 1GB of headroom in your heap
   space for compaction, you can expect to store about 17M keys
   before running into problems.
   See https://issues.apache.org/jira/browse/CASSANDRA-208

2. Because fixing #1 requires a data file format change, 0.4 will not
   be binary-compatible with 0.3 data files.  A client-side upgrade
   can be done relatively easily with the following algorithm:
     for key in old_client.get_key_range(everything):
         columns = old_client.get_slice or get_slice_super(key, all columns)
     new_client.batch_insert or batch_insert_super(key, columns)
   The inner loop can be trivially parallelized for speed.

3. Commitlog does not fsync before reporting a write successful.
   Using blocking writes mitigates this to some degree, since all
   nodes that were part of the write quorum would have to fail
   before sync for data to be lost.
   See https://issues.apache.org/jira/browse/CASSANDRA-182

Additionally, row size (that is, all the data associated with a single
key in a given ColumnFamily) is limited by available memory, because
compaction deserializes each row before merging.

See https://issues.apache.org/jira/browse/CASSANDRA-16