= New Features

* Dataset#for_no_key_update and #for_key_share have been added on
  PostgreSQL.  If you are not deleting a row or modifying a key
  column in a row, it is recommended to switch existing #for_update
  calls to #for_no_key_update, so they don't take stronger locks than
  necessary (#for_update will block concurrent INSERTs that reference
  a locked row).

* Check and foreign key constraints now support a :not_enforced option
  on PostgreSQL 18+, for using NOT ENFORCED. Constraints that are not
  enforced serve mostly documentation purposes, but they can be
  enforced later by altering the constraint.

* On PostgreSQL, alter_table blocks now support an alter_constraint
  method, which accepts the following options for altering the
  constraint:

  :deferrable :: Modify deferrable setting for constraint
                 (PostgreSQL 9.4+):
                 true :: DEFERRABLE INITIALLY DEFERRED
                 false :: NOT DEFERRABLE
                 :immediate :: DEFERRABLE INITIALLY IMMEDIATE
  :enforced :: Set true to use ENFORCED, or false to use NOT ENFORCED
               (PostgreSQL 18+)
  :inherit :: Set true to use INHERIT, or false to use NO INHERIT
              (PostgreSQL 18+)

* Entries in Database#foreign_key_list, #reverse_foreign_key_list,
  and #check_constraint arrays on PostgreSQL now include :validated
  and :enforced entries to indicate whether the constraint is
  validated and enforced, respectively.

* :primary_key and :unique column options now support hash values in
  the schema generators.  Using a hash value allows you to provide
  column constraint-specific options in the hash.  This currently
  supports the :name and :deferrable options for both constraint
  types. Additional options will be supported in the future.

= Other Improvements

* Dataset#for_share on PostgreSQL and MySQL now caches the returned
  dataset, similar to the caching done by #for_update.

* Dataset#nolock on Microsoft SQL Server now caches the returned
  dataset, similar to the caching done by #for_update.

* Many minor performance improvements have been applied, mostly
  fixed and/or flagged by rubocop-performance:

  * sort_by! instead of sort!
  * delete/tr instead of gsub
  * yield instead of block.call
  * start_with? accepts multiple arguments
  * Hash#[]= instead of merge!
  * Array.new instead of Integer#times.map
  * Hoist literal arrays inside loops to before the loop
  * Use block instead of Method#to_proc
  * Use start_with instead of regexp with \A
  * Use String#include? instead of =~

* In the jdbc adapter, Database#foreign_key_list has been optimized,
  reducing the number of allocations.

* On DB2, multiple UNIQUE constraints on the same table are now
  handled when automatically marking columns as NOT NULL.