= New Features

* An index_caching extension has been added, which makes
  Database#indexes use a cache similar to Database#schema, and also
  offers methods for saving and loading the cache from a file, similar
  to the schema_caching extension.
  
  This can speed up model loaded in certain cases when the
  auto_validations plugin is used.

* A datetime_parse_to_time extension has been added, which parses
  strings without timezone offsets using DateTime.parse intead of
  Time.parse.  This can fix problems when the string being parsed
  represents a time not valid in the local timezone due to daylight
  savings time shifts.  Time.parse silently shifts such times by 1
  hour instead of raising an exception, resulting in incorrect
  behavior in that case.

  It only makes sense to use this extension when the times in the
  database are stored in UTC but not returned with timezone
  information, the timezone for the Database instance
  (or Sequel.database_timezone) is set to :utc (not the default),
  and Time is used as the datetime_class (the default).

* A pg_timestamptz extension has been added for switching the default
  generic timestamp type from timestamp to timestamptz.

* Sequel.date_{add,sub} in the date_arithmetic extension now supports
  a :cast option for setting the cast type.  This value defaults to
  Time for backwards compatibility, which uses the default generic
  timestamp type for the database.

* The class_table_inheritance plugin now supports an
  :ignore_subclass_columns option which takes an array of column
  symbols to ignore in subclasses.  This allows you to use
  the plugin when your table inheritance hierarchy includes
  non-primary key columns with the same name in different tables.

= Improvements

* Dataset#insert_select now returns false instead of nil if it runs
  an INSERT statement but does not return a value on Microsoft SQL
  Server or PostgreSQL.  This can happen on both databases if triggers
  are used.

  Model#save now checks for a false value returned by
  Dataset#insert_select, and does not issue another INSERT statement
  in that case.

* Database#indexes now correctly handles SQL::Identifier arguments on
  SQLite, Microsoft SQL Server, SQLAnywhere, and DB2.

* Dataset#to_json in the json_serializer plugin and Dataset#to_xml
  in the xml_serializer plugin now both handle datasets that use
  eager_graph.

* Dataset#nullify now caches the dataset it returns, for better
  performance if it is called more than once on the same dataset.

* Database#synchronize is now optimized on ruby 2.5+ and is about
  10% faster by relying on the new lazy proc allocation feature.

= Backwards Compatibility

* Fractional second timestamps are now enabled on DB2.  If you are
  connecting to a DB2 database that does not support fractional
  seconds, you should add the following code (where DB is your
  Sequel::Database instance):

    DB.extend_datasets do
      def supports_timestamp_usecs?
        false
      end
    end

* Calling a filtering method with no argument and a virtual row
  block that returns nil on a dataset with no existing filter now
  adds a WHERE NULL filter, to match the behavior if given a nil
  argument.  Previously, a deprecation warning was issued and a
  dataset with no filter was returned.