General

• This release is focused on making the TokuMX cluster experience smoother. The major improvements are in eliminating stalls that could appear in replica sets and sharded clusters.

Highlights

The list of minor new features and bugfixes in this release is large. The most prominent of them appear here:

• Chunk migrations in sharding have been greatly simplified and optimized. They no longer take as many locks, particularly database- and server-wide locks, which could previously cause stalls and pileups. Additionally, the data transfer between shards has been simplified and optimized and is significantly faster and more reliable. (#214, #746, #823, #824)

• In TokuMX 1.3 and earlier, updates were represented in the storage layer, and in the oplog, by their pre- and post-images, which could be quite large compared to the actual change that occurred in the document (consider incrementing an integer in a large document). For replication, this ensures that secondaries use significantly less I/O for many update workloads, but costs a lot of space in the oplog and over the network.

  In TokuMX 1.4, updates to unindexed fields represent their updates on disk with the update modifier object itself, which can significantly reduce the amount of cache memory used by internal node buffers in the Fractal Tree indexing structure.

  Most updates (including to indexed fields) are represented in the oplog by the pre-image and the update object, which can reduce, by up to half, the amount of space and network bandwidth used by the oplog. (#511)

• It is now possible to define a "primary key" for a collection. A primary index for a collection is defined by the primary key, is an index which is unique and clustering, and is the index that is used by all non-clustering secondary indexes to find the full document. The primary key for each document is therefore embedded in every secondary index. The key must end in {_id: 1} to ensure uniqueness, and it is a good idea to let the _id be auto-generated.

  In some cases, if a clustering secondary index is desired, a primary key can be defined instead, which removes the need for a clustering _id index, which therefore can reduce storage costs even further for such a use case. (#684)

• This primary key concept is prominently featured in sharding. Newly sharded, nonexistent collections with a shard key other than just the _id will be created using a primary key of the shard key concatenated with {_id: 1}. In TokuMX 1.3 and earlier, to facilitate fast migrations and range queries, it was strongly encouraged to make the shard key clustering, but this had the drawback that there were automatically two clustering indexes (the _id index and the shard key). Now, sharded collections can have a clustered shard key without paying the extra storage cost, and this behavior is the default. (#740)

• In TokuMX 1.3 and earlier, the oplog was a single collection that was trimmed by a background task, to maintain a specific amount of history, specified in days (expireOplogDays). This background thread could sometimes fail to reclaim space aggressively enough, so in TokuMX 1.4, the oplog is instead partitioned into a set of N collections, one per day, and each day, the oldest partition is dropped according to expireOplogDays, which reclaims space immediately. This should make the oplog occupy a smaller disk footprint overall, and should alleviate some locking problems that used to be caused by the background trimming task.

  So-called "partitioned collections" (of which the oplog is now one) are a generally-available feature, but they currently have significant limitations, including that they do not permit secondary indexes. See the user's guide for more details. (#762)

New features and improvements

• The listDatabases command now computes a size estimate for each db, which also means that show dbs in the shell will report the size of each database (compressed and uncompressed), rather than showing "(empty)" for everything regardless of contents. (#57)

• Provided new implementations for db.getReplicationInfo(), db.printReplicationInfo(), and db.printSlaveReplicationInfo(), including compressed and uncompressed sizes. (#223)

• The compression, pageSize, and readPageSize attributes of an index can now be changed online. The db.collection.reIndex() shell function now takes a second parameter, which if specified is an object describing the options to set. These options take effect immediately for all data written after the command runs, but will not rewrite existing data. To force existing data to be rewritten with the new options, run a reIndex command without specifying options, to rewrite all the data for that index, forcing the conversion. See the user's guide for full usage details. (#389)

• The diagnostic tools db.showPendingLockRequests() and db.showLiveTransactions() now use cursors to return results, which means that in extreme cases, the results will no longer be truncated due to the maximum size of a BSON object, instead the results will be paginated. This doesn't change the fact that they are a snapshot of the state of the system. (#606)

• The count() operation has been further optimized for simple interval count operations, by up to 4x in some workloads. (#725)

• Exhaustive scans are no longer ever chosen when a useful index exists for a query. The query optimizer is a sample-based optimizer which would periodically retry all possible plans, including an exhaustive scan. For certain update workloads, this could cause a major 99th percentile latency problem, and concurrency bottleneck, so this change alleviates that problem. (#796)

• The sharding subsystem used to use database- and system-wide locking to protect metadata and sharding-only state. That usage of those locks has been refined to avoid stalls caused by long-running DML operations combined with unrelated administrative operations. (#797)

• Added the $allMatching flag to db.currentOp() to aid in finding background tasks. (#814)

• The storage format for integers with absolute value greater than 2^52 has been optimized to reduce the CPU cost of operations on such values. This is of particular importance to users of hashed keys (e.g. hashed shard keys), because hash values are pseudorandom bit patterns and are therefore almost always larger than 2^52. In some tests, up to a 20% speed improvement has been measured for CPU-bound workloads. (#820)

• Added the compression, pageSize, and readPageSize parameters to the db.createCollection() shell function, so the following is now valid:

  db.createCollection('foo',
                      {compression: 'lzma'
                       pageSize: '8MB'})
  

  Also, shell functions were added to wrap the getParameter and setParameter commands. See db.help() for more. (#854)

• Added a server option loaderCompressTmp that, if on, causes the bulk loader (used by mongorestore and mongoimport and non-background index builds) to compress intermediate files before writing them to disk, thus using less disk space during the index build. This can also be set with the setParameter command. (#858)

• TokuMX is now built with CMake. Therefore, the binary tarball names have changed slightly, though they will unpack to a similar location to previous tarballs. (#869)

• When a document-level lock request times out (the lockTimeout parameter), some debugging information is printed to the server's log, including a transaction identifier for the operation which was holding the lock that the current operation failed to acquire. Now, that blocking transaction's information (the same information printed in db.currentOp() is also printed to the log, which should help in debugging lock timeouts in applications. (#876)

• Improved the latency of short range query operations. (#882)

• Improved the performance of workloads with high concurrency. (#904)

• Added information about the following to db.serverStatus():

  • Cache misses of the "partial fetch" type.
  • Cache evictions.
  • Number of pending document-level lock requests.

  See the user's guide for a full description. (#905, #906, #907)

• Added a server parameter for mongod, migrateUniqueChecks. If this setting is on, then during a migration, TokuMX will perform unique checks for all data in the chunk being migrated. That is the default.

  These unique checks exist to prevent users from inserting different documents that initially are inserted to different shards, but which have the same _id, from causing issues if those documents are migrated to the same server.

  With this setting turned on, such a migration would fail, but the documents would be preserved.

  If this setting is turned off, such a migration would succeed, and silently overwrite the document that was there before, just in the _id index. This would effectively overwrite that document, and for any secondary indexes, the newly migrated document would appear twice. This can only happen when using an _id index as the primary key (the old behavior, see above), but when also using a shard key other than the _id, and simultaneously generating inappropriate (non-unique) values for the _id in the application.

  For the best performance, use the default behavior of shardCollection which will create a primary key containing the shard key, allow TokuMX to generate _id values on your application's behalf (use a different field for an application-level ID if needed), and turn off migrateUniqueChecks. If these recommendations are followed, this setting is safe and will not lead to corruption or data loss. However, to be careful, the default is to do the unique checks, which is slower. (#941)

Bug fixes

• The fsync command, which has always been useless in TokuMX (its functionality is available with other commands: getLastError with j or fsync set to 1 flushes the transaction log, and checkpoint forces data pages to be written), has finally been deprecated, and is now completely a no-op. (#180)

• Updates with upsert can no longer allow bad values for the _id field (Array, RegEx, or undefined). (#736)

• Inserts into capped collections within a multi-statement transaction have been disabled, because they could not properly maintain the size of the collection. (#852)

• Replication initial sync no longer fails when a profile collection exists. (#862)

• All relevant bugfixes from MongoDB 2.4.9 have been incorporated. (#868)

  This includes:

  • SERVER-5625
  • SERVER-7246
  • SERVER-9238
  • SERVER-10538
  • SERVER-10919
  • SERVER-11099
  • SERVER-11194
  • SERVER-11502
  • SERVER-11731
  • SERVER-11971
  • SERVER-11977
  • SERVER-12041
  • SERVER-12092
  • SERVER-12094
  • SERVER-12146

• An index can no longer become a "multi-key" index inside a multi-statement transaction. An index becomes multi-key when a document is inserted such that the key generated for that index contains an array in the document, and therefore generates multiple keys (this is useful information for the query optimizer), but this only happens when the first such document is inserted. This transition cannot be safely maintained if done in a multi-statement transaction, so it will throw an error if done.

  If you intend to use documents with array-valued key fields, as well as multi-statement transactions, you may wish to prepare the collection by inserting one such document so that the index becomes multi-key, then removing that document, all outside of a multi-statement transaction. (#872)

• Fixed a bug where a dropped collection could leave behind a piece of zombie metadata that would prevent future collections from being created with the same name. If this has happened already, upgrading to 1.4 will clean out any such zombie metadata. (#879, #937)

• Fixed two bugs related to setting the size of a capped collection:

  • When setting a capped collection size to an integer larger than 2GB, no longer reject all insertions until the server is restarted.
  • When setting a capped collection size to a string to be interpreted with a SI suffix (K, M, G), no longer reject all insertions after the server is restarted.

  (#934)

• Prevented update messages from growing leaf nodes pathologically. (Tokutek/ft-index#168)