General

• This release is focused on the addition of a new collection type, partitioned collections, which provide excellent performance and flexibility for time-series workloads.

Highlights

• Partitioned collections are a new feature with a similar use case to capped collections in basic MongoDB, but with better concurrency and flexibility.

  Briefly, a partitioned collection is similar to a partitioned table in some RDBMSes. The collection is divided physically by primary key into multiple underlying partitions, allowing for the instantaneous and total deletion of entire partitions, but it is treated as one logical collection from the client's perspective.

  A typical use case for this would be to partition on a {date: 1, _id: 1} primary key, and periodically add new partitions and drop old ones to work with the latest data in a time-series data set. (#1063)

New features and improvements

• Further reduced the locking done during migrations, to reduce stalls. (#1041, #1135)

• Initial sync now reports its progress in terms of the # of dbs, collections within a db, and bytes copied, then indexes built for a collection, in the server log, in db.currentOp(), and in rs.status(). (#1050)

• Added a new option to allow operations that require a metadata write lock (such as adding an index) to interrupt operations that are holding a read lock. This option can be used to avoid those write lock requests stalling behind a long read locked operation (such as a large multi: true update), and in turn stalling all other operations, including readers, behind itself.

  To use this option, your application must be ready to handle failed operations and retry them (transactional atomicity makes this somewhat simpler), then you can turn it on for all clients with --setParameter forceWriteLocks=true or with the setParameter command. This makes all future connections interruptible while they have read locks, but this state can be controlled for each connection with the setWriteLockYielding command. See https://github.com/Tokutek/mongo/issues/1132 for more details. (#1132)

• During initial sync, there is a step that checks that all collections that existed at the beginning of the sync's snapshot transaction still exist, to maintain the snapshot's validity. This step has been optimized to avoid slowness in some rare situations. (#1162)

• Serial insertions have been made significantly faster with improvements to the storage engine. (Tokutek/ft-index#158)

• Update workloads on extremely large data sets (where a large portion of even internal nodes do not fit in main memory) have been optimized somewhat. (Tokutek/ft-index#226)

• Added per-connection control of document-level lock wait timeouts with a new command, {setClientLockTimeout: <timeoutms>}.

  This controls the same behavior as --lockTimeout which can also be controlled server-wide with setParameter (which now controls the default for all new connections). The value should be specified in milliseconds. (#1168)

Bug fixes

• Cleaned up the way remote errors are handled in the distributed locking framework, which will reduce the impact of some rare errors on config servers. (#857, #1140)
• While a shard is draining, it will no longer be chosen to receive new chunks due to the sequential insert optimization, nor will it be chosen as the primary shard for new databases. (#1069)
• The mongo2toku tool now correctly uses the admin database for authentication to the target cluster by default. (#1125)
• Fixed the way packaging scripts check for transparent hugepages, and changed them to allow madvise as an acceptable setting. (#1130)
• Connections used to copy data for initial sync no longer time out improperly if some operations (such as checking for snapshot invalidation) take too long. (#1147)