Wiredtiger

The WiredTiger 2.3.1 release contains mostly performance enhancements and bug fixes.

Changes to the WiredTiger API (see upgrading documentation for details):
- Fix a bug in WT_CURSOR::set_value that could lead to undefined behavior with some value formats.
- Make the asynchronous API generally available [1139]
- Add log cursors for replay and verification. Make generated log record and operation types public. [1106]
- Allow eviction worker threads to be started and stopped dynamically. [1143, 1158]

Other significant changes:
- Improve performance and reduce latency during checkpoints and LSM merges. Remove uses of the checkpoint lock other than serializing checkpoints: compact holds the schema lock, so it doesn't need to hold the checkpoint lock, the new WT_BTREE handle close lock prevents checkpoints from colliding with handle close, so LSM doesn't need the checkpoint lock either.
- Some minor cleanups, setting the internal session's name in a few places. [1073]
- Grab the live lock when loading a checkpoint in diagnostic mode: that could race with a read. [1102]
- Instead of keeping a list of file URIs for checkpoint to flush, open a handle and stash it. [1114]
- Start and stop eviction worker threads based on demand. [1116]
- Add a new OS-layer function __wt_fsync_async to flush a file without waiting for the results, call it from the Btree flush-leaves code so pages start flushing while we're working the rest of the checkpoint. [1136, 1152]
- Wait for the handle flush lock when writing the leaf pages instead of returning EBUSY. [1136]
- Add a wtperf page to the documentation, describe how to simulate workloads and view statistics. [1147]
- Flag new structures not listed in PREDEFINE. [1148]
- Return EBUSY if no async handles available and fix ex_async to look for it. [1153]
- Fix some problems with navigation in the reference guide.
- Bump the number of slots for internal sessions: we have a lot more than 2 now. Add a test for `session_max` settings, make sure we add enough to account for at least the default internal sessions.
- Remove tcbench: we're no longer maintaining it.

The WiredTiger 2.3.0 release contains significant new features, performance
enhancements and bug fixes. Documentation for this release is here:

http://source.wiredtiger.com/2.3.0/index.html

Significant changes are described below.

Changes to the WiredTiger API (see upgrading documentation for details):
- Add a LevelDB API implementation for WiredTiger. This includes support for
stock LevelDB as well as Basho, HyperLevelDB and RocksDB versions of the API.
To build the LevelDB API include --enable-leveldb in the configure command,
to specify compatability with an alternative LevelDB API use
--enable-leveldb=[basho,hyper,rocksdb]. [1028]
- Add ability to build some common extensions into the WiredTiger library.
This means that the libraries for those extensions don't need to be
dynamically loaded at runtime. Currently supported extensions are Snappy
compression and zlib compression. The option can be enabled by passing
--with-builtins=[snappy,zlib] to the configure command line.
- Add a new configuration to wiredtiger_open: statistics_log=(on_close=true),
that causes a set of statistics to be logged on WT_CONNECTION::close. [1086]
- Add a new configuration to wiredtiger_open: exclusive, that causes the open
to fail if the database already exists.

Other significant changes:
- Performance improvement for high throughput workloads using multiple
eviction threads. Performance of some workloads improves by over 15% [1087]
- Significant performance optimizations for queries, giving up to 20%
throughput improvement for in-memory query workloads.
https://github.com/wiredtiger/wiredtiger/wiki/Query-throughput
- Fix an off-by-one bug that could lead to ENOMEM during commit with logging.
[1104][1121]
- Allow bulk loads to multiple files to complete in parallel. [1114][1126]