0.32

New Features

• #1079 Transactional is now implemented for
  [&Tree] and [Tree] so you can avoid the
  previous friction of using tuples, as was
  necessary previously.
• #1058 The minimum supported Rust version (MSRV)
  is now 1.39.0.
• #1037 Subscriber now implements Future (non-fused)
  so prefix watching may now be iterated over via
  while let Some(event) = (&mut subscriber).await {}

Improvements

• A large number of crash consistency issues have
  been discovered during intensive testing and fixed.
• #965 concurrency control is now dynamically enabled
  for atomic point operations, so that it may be
  avoided unless transactional functionality is
  being used in the system. This significantly
  increases performance for workloads that do not
  use transactions.
• A number of memory optimizations have been implemented.
• Disk usage has been significantly reduced for many
  workloads.
• #1016 On 64-bit systems, we can now store 1-2 trillion items.
• #993 Added DerefMut and AsMut<[u8]> for IVec where it
  works similarly to a Cow, making a private copy
  if the backing Arc's strong count is not 1.
• #1020 The sled wiki has been moved into the documentation
  itself, and is accessible through the doc module
  exported in lib.

Breaking Changes

• #975 Changed the default segment_size from 8m to 512k.
  This will result in far smaller database files due
  to better file garbage collection granularity.
• #975 deprecated several Config options that will be
  removed over time.
• #1000 rearranged some transaction-related imports, and
  moved them to the transaction module away from
  the library root to keep the top level docs clean.
• #1015 TransactionalTree::apply_batch now accepts
  its argument by reference instead of by value.
• Event has been changed to make the inner fields
  named instead of anonymous.
• #1057 read-only mode has been removed due to not having
  the resources to properly keep it tested while
  making progress on high priority issues. This may
  be correctly implemented in the future if resources
  permit.

varints, read and recovery optimizations

0.31

Improvements

• #947 dramatic read and recovery optimizations
• #921 reduced the reliance on locks while
  performing multithreaded IO on windows.
• #928 use sync_file_range on linux instead
  of a full fsync for most writes.
• #946 io_uring support changed to the rio crate
• #939 reduced memory consumption during
  zstd decompression

Breaking Changes

• #927 use SQLite-style varints for serializing
  u64. This dramatically reduces the written
  bytes for databases that store small keys and
  values.
• #943 use varints for most of the fields in
  message headers, causing an additional large
  space reduction. combined with #927, these
  changes reduce bytes written by 68% for workloads
  writing small items.

sled::open replaces sled::Db::open

New Features

• Added the open function for quickly opening a database at a path with default configuration.

new storage format and bugfixes

• much faster compile speeds after aggressive dependency pruning
• new storage format
• bugfixes in cache invalidation and combining transactions with compression

Config, Transaction, and Opening Revamp

New Features

• The Config::open method has been added to give
  Config a similar feel to std's fs::OpenOptions.
  The Config::build and Db::start methods are
  now deprecated in favor of calling Config::open
  directly
• A checksum method has been added to Tree and Db
  for use in verifying backups and migrations.
• Transactions may now involve up to 69 different
  tables. Nice.
• The TransactionError::Abort variant has had
  a generic member added that can be returned
  as a way to return information from a
  manually-aborted transaction. An abort helper
  function has been added to reduce the boiler-
  plate required to return aborted results.

Breaking Changes

• The ConfigBuilder structure has been removed
  in favor of a simplified Config structure
  with the same functionality.
• The way that sled versions are detected at
  initialization time is now independent of serde.
• The cas method is deprecated in favor of the new
  compare_and_swap method which now returns the
  proposed value that failed to be applied.
• Tree nodes now have constant prefix encoding
  lengths.
• The io_buf_size configurable renamed to
  segment_size.
• The io_buf_size configurable method has been
  removed from ConfigBuilder. This can be manually
  set by setting the attribute directly on the
  ConfigBuilder, but this is discouraged.
  Additionally, this must now be a power of 2.
• The page_consolidation_threshold method has been
  removed from ConfigBuilder, and this is now
  a constant of 10.

async flush, bugfixes, event variant rename

Breaking Changes

• Event::Set has been renamed to Event::Insert and
  Event::Del has been renamed to Event::Remove. These
  names better align with the methods of BTreeMap from
  the standard library.

Bug Fixes

• A deadlock was possible in very high write volume
  situations when the segment accountant lock was
  taken by all IO threads while a task was blocked
  trying to submit a file truncation request to the
  threadpool while holding the segment accountant lock.

New Features

• flush_async has been added to perform time-intensive
  flushing in an asynchronous manner, returning a Future.

Transactions!

0.26

New Features

• Transactions! You may now call Tree::transaction and
  perform reads, writes, and deletes within a provided
  closure with a TransactionalTree argument. This
  closure may be called multiple times if the transaction
  encounters a concurrent update in the process of its
  execution. Transactions may also be used on tuples of
  Tree objects, where the closure will then be
  parameterized on TransactionalTree instances providing
  access to each of the provided Tree instances. This
  allows you to atomically read and modify multiple
  Tree instances in a single atomic operation.
  These transactions are serializable, fully ACID,
  and optimistic.
• Tree::apply_batch allows you to apply a Batch
• TransactionalTree::apply_batch allow you to
  apply a Batch from within a transaction.

Breaking Changes

• Tree::batch has been removed. Now you can directly
  create a Batch with Batch::default() and then apply
  it to a Tree with Tree::apply_batch or during a
  transaction using TransactionalTree::apply_batch.
  This facilitates multi-Tree batches via transactions.
• Event::Merge has been removed, and Tree::merge will
  now send a complete Event::Set item to be distributed
  to all listening subscribers.

Write Batches, New Storage Architecture, Fewer Dependencies

Merge pull request #759 from spacejam/tyler_cut_0.25

cut 0.25

GC improvements, more writepath optimizations

This release brings big improvements to the underlying GC system for keeping file sizes under control, as well as some optimizations for the writepath.

API cleanup, writepath optimizations

This release includes a large number of internal and external API cleanups. Some internal CAS-like operations will retry automatically when they fail due to changes that do not impact causality, like rewriting the same page to a new location. When CAS-like operations fail, they now return the new value that was created to insert in the first place, removing several double allocations. Lots of performance improvements.