• Fix bug with reloading models with all-queries default scopes. The previous implementation allowed non-all-queries on the current scope to leak into the scope used for reloading.

  Andrew Novoselac and Matthew Draper

• insert! now accepts the :unique_by option, consistent with insert.

  Kenta Ishizaki

• Fix reading a store_accessor on a NULL structured column (json, jsonb, or hstore) marking the record as changed and overwriting the NULL with an empty hash on the next save.

  Kenta Ishizaki

• Fix update_all corrupting the optimistic locking column when it is set through an alias_attribute.

  Kenta Ishizaki

• Fix serialize with coder: ActiveRecord::Coders::JSON silently double-encoding a native json/jsonb column.

  Kenta Ishizaki

• Fix update_all / delete_all ignoring group and having, updating or deleting every row in the table instead of only the rows that satisfy the HAVING clause.

  Kenta Ishizaki

• Fix ActiveRecord::Relation#cache_key / cache_version for a loaded collection containing a record without a timestamp.

  Kenta Ishizaki

• Fix ::ActiveRecord::MessagePack serialization raising NoMethodError for any record with a populated time column, which made such records uncacheable through the MessagePack cache serializer.

  Kenta Ishizaki

• Fix replacing or clearing a polymorphic has_one leaving a stale type column on the removed record.

  Kenta Ishizaki

• Add sql_notifications connection configuration option to disable SQL notifications for specific database connections.

  Libraries like Solid Cache and Solid Queue that use separate database connections via connects_to can suppress internal SQL notification overhead by setting sql_notifications: false in database.yml:

  cache:
    adapter: postgresql
    database: myapp_cache
    sql_notifications: false
  

  Rosa Gutierrez

• Allow the query log tags format to be configured per connection pool.

  A query_log_tags_format key in a database.yml entry overrides the global config.active_record.query_log_tags_format for connections in that pool, so different databases can emit :legacy or :sqlcommenter formatted comments.

  production:
    primary:
      database: primary
    analytics:
      database: analytics
      query_log_tags_format: sqlcommenter
  

  Hartley McGuire

• Fix update_attribute/update_attribute! to raise for a readonly attribute referenced by an alias.

  Kenta Ishizaki

• Fix reset_column_sequences! for tables in quoted schemas.

  Kenta Ishizaki

• Fix accepts_nested_attributes_for :limit miscounting a single-record hash.

  Kenta Ishizaki

• Fix rename_index to preserve a partial index's WHERE for SQLite and older MySQL/MariaDB versions.

  Kenta Ishizaki

• Fix PostgreSQL foreign_keys returning a corrupted to_table for a foreign key that references a table in a quoted schema.

  Kenta Ishizaki

• Fix grouped calculations (e.g. count) grouped by a belongs_to association that points to a composite primary key model.

  Book.group(:order).count, where Book belongs_to :order and Order has a composite primary key, raised ArgumentError: Expected corresponding value for ["shop_id", "id"] to be an Array. The grouped result is now keyed by the associated records as it already is for single-column keys.

  Kenta Ishizaki

• Fix has_many/has_one :through associations with disable_joins: true silently returning an empty result when the source points to a composite primary key model.

  Kenta Ishizaki

• Fix collection association ids= writers (e.g. author.book_ids=) raising ::ActiveRecord::RecordNotFound for existing records when a composite primary key model is assigned string ids (the shape ids take from request params).

  Each id component is now cast against its own column type, instead of casting the whole tuple against an array of column names, which type_for_attribute can't resolve to a real type (so the cast was a no-op and the string ids never matched the loaded records).

  Kenta Ishizaki

• Fix find with multiple composite primary key ids passed as strings silently returning [].

  Model.find([["1", "10"], ["1", "20"]]) (the shape ids take when they come from request parameters) returned an empty array instead of the records and without raising RecordNotFound. The ids were cast against the array of key column names as a whole — which is a no-op — so the string tuples never compared equal to the records' integer ids when ordering the result. Each component is now cast against its own column type, matching the documented coercion already performed for single-column keys.

  Kenta Ishizaki

• Fix PostgreSQL daterange / tsrange / tstzrange schema dump producing invalid Ruby.

  The schema dumper used to render range defaults via Range#inspect, which falls back to Date#inspect / Time#inspect for the bounds. The resulting schema.rb literal (for example default: Mon, 01 Jan 2024..Wed, 01 Jan 2025) raised a SyntaxError on db:schema:load. The bounds are now rendered via the subtype's type_cast_for_schema, so date and timestamp range defaults round-trip through schema.rb like any other column.

  Kenta Ishizaki

• Respect schema_search_path on rails dbconsole for PostgreSQL.

  Gabriel Sobrinho

• Preserve IPv6 prefix when dumping PostgreSQL cidr / inet defaults to schema.rb.

  The schema dumper used to omit the prefix whenever it equaled /32, which is the full mask for IPv4 but not for IPv6. As a result, an IPv6 default such as "::/32" was written as "::" in schema.rb and reloaded as ::/128, silently dropping the subnet information. The prefix is now only omitted when it covers the full address (/32 for IPv4, /128 for IPv6).

  Kenta Ishizaki

• Fix deadlock when pool-less connection materializes while fetching database server version.

  Hartley McGuire

• Add #default_order query method and association option which can be used to order records when no other order is specified.

  This extends the functionality offered by #implicit_order_column to scopes and associations.

  Usage:

  class Post < ApplicationRecord
    has_many :comments, default_order: :likes
    # .. or ..
    has_many :comments, -> { default_order(:likes) }
  end
  
  post = Post.first
  
  post.comments # comments ordered by `likes`
  post.comments.order(:created_at) # comments ordered by `created_at`

  Dan Ungureanu

• Raise ::ActiveRecord::MultiparameterAssignmentErrors instead of NoMethodError when assigning a malformed multiparameter attribute name.

  A key routed to the multiparameter code path but missing a closing parenthesis (e.g. "written_on(") used to crash with NoMethodError: undefined method 'first' for nil inside find_parameter_position. It now raises the same MultiparameterAssignmentErrors already used for other invalid multiparameter input, so callers can rescue a single documented error class.

  Kenta Ishizaki

• Include the list of valid values in the ArgumentError raised when assigning an invalid value to an enum attribute.

  Book.new(status: "bogus")
  # ArgumentError: 'bogus' is not a valid status. Valid values are: "proposed", "written", "published"

  Hammad Khan

• Include the mismatched keys in the error raised by insert_all / upsert_all when the inserted objects have inconsistent attributes.

  Book.insert_all [{ name: "Rework", author_id: 1 }, { name: "Remote", author_id: 1, isbn: "x" }]
  # ArgumentError: All objects being inserted must have the same keys (extra: [:isbn])

  Hammad Khan

• Allow create_join_table to accept a primary key.

  This is useful for databases like PostgreSQL where logical replication requires primary keys on all tables.

  create_join_table :assemblies, :parts, primary_key: [:assembly_id, :part_id]

  This generates a join table with a composite primary key on both foreign key columns.

  Genadi Samokovarov

• Fix Active Record Pool Reaper thread leak after Parallelization#shutdown.

  After parallelized test runs, the parent process leaked the Active Record Pool Reaper thread, holding open connection pools and file descriptors for up to reaping_frequency seconds (or indefinitely if never reaped).

  Three fixes: Parallelization#shutdown now calls run_cleanup_hooks; Active Record registers a cleanup hook that discards all connection pools; and ConnectionPool#discard! now calls Reaper.discard_pool, which immediately kills and joins the reaper thread when no pools remain at that frequency.

  Ruy Rocha

• Reset lock_version after a nested savepoint rollback.

  When a record was saved inside a transaction(requires_new: true) block that later rolled back, the in-memory lock_version was left at the incremented value while the database row was reverted by the savepoint. Saving the same instance again raised ::ActiveRecord::StaleObjectError because the WHERE clause used the bumped value that no longer existed in the database.

  Restore the locking column from the snapshot on savepoint rollback in addition to the outermost transaction rollback handled in #57363.

  Kenta Ishizaki

• Fix duplicate WHERE conditions in create_or_find_by.

  When create_or_find_by catches a RecordNotUnique error and retries the query, it now uses rewhere and take! to prevent duplicating existing scope conditions.

  Yavor Dashev

• Allow to pass array values to in_order_of

  Passing arrays allows to group records and order those groups with another query:

  Posts
    .in_order_of(:state, [[:published, :canceled], :archived])
    .order(created_at: :desc)
  # =>
  # SELECT "posts".* FROM "posts" WHERE "posts"."state" IN (1, 2, 3)
  # ORDER BY CASE
  #   WHEN "posts"."state" IN (1, 2) THEN 1
  #   WHEN "posts"."state" = 3 THEN 2
  #  END ASC, "posts"."created_at" DESC

  Markus Doits

• On PostgreSQL 18.4+, disable_referential_integrity uses NOT ENFORCED/ENFORCED instead of DISABLE TRIGGER ALL/ENABLE TRIGGER ALL, requiring only table ownership rather than superuser privileges. Only currently ENFORCED foreign keys are toggled; intentionally NOT ENFORCED foreign keys are left unchanged.

  Unlike ENABLE TRIGGER ALL, restoring ENFORCED checks existing rows against the constraint, so FK violations raise ::ActiveRecord::InvalidForeignKey rather than RuntimeError. The toggle and restore run in a single transaction so a restoration failure rolls back the initial NOT ENFORCED toggle too — preventing originally- ENFORCED constraints from being left in a NOT ENFORCED state indistinguishable from intentional ones.

  Fixtures sharing FK relationships must be passed together to FixtureSet.create_fixtures to ensure all referenced rows are present when enforcement is restored.

  check_all_foreign_keys_valid! skips NOT ENFORCED constraints on PostgreSQL 18.4+, as VALIDATE CONSTRAINT cannot be applied to them.

  Unlike SET CONSTRAINTS ALL DEFERRED (the approach attempted in rails/rails#27636 and reverted), NOT ENFORCED also suppresses referential actions such as ON DELETE CASCADE, SET NULL, and SET DEFAULT.

  Yasuo Honda

• Add exclusion_constraint_exists? and unique_constraint_exists? helpers

  fatkodima

• Add enforced: option to add_foreign_key and change_foreign_key for PostgreSQL 18.4+.

  NOT ENFORCED foreign keys are available since PostgreSQL 18.0, but DEFERRABLE was lost on them until 18.4 ("Fix loss of deferrability of foreign-key triggers", https://www.postgresql.org/docs/release/18.4/). Rails therefore requires PostgreSQL 18.4 or later for this feature.

  When enforced: false is passed to add_foreign_key, the constraint is created as NOT ENFORCED, meaning PostgreSQL skips referential integrity checks during DML. PostgreSQL marks NOT ENFORCED constraints as NOT VALID internally (and VALIDATE CONSTRAINT does not apply to them), so the schema dumper outputs both enforced: false and validate: false.

  change_foreign_key toggles the enforced status of an existing foreign key. Without this method, users would need to issue raw ALTER TABLE ... ALTER CONSTRAINT SQL to disable enforcement temporarily (e.g., for bulk DML that loads the referenced and referencing tables in arbitrary order) and then re-enable it.

  add_foreign_key :articles, :authors, enforced: false
  # => ALTER TABLE "articles" ADD CONSTRAINT ... FOREIGN KEY ("author_id") REFERENCES "authors" ("id") NOT ENFORCED
  
  change_foreign_key :articles, :authors, enforced: true
  # => ALTER TABLE "articles" ALTER CONSTRAINT "fk_rails_..." ENFORCED
  
  change_foreign_key :articles, :authors, enforced: false
  # => ALTER TABLE "articles" ALTER CONSTRAINT "fk_rails_..." NOT ENFORCED

  Yasuo Honda

• Expose cursor, order and use_ranges attributes for BatchEnumerator

  fatkodima

• Fix ActiveRecord::QueryMethods#in_order_of when passing an out-of-range Integer

  To match the behavior of the Enumerable version, in_order_of now ignores an out-of-range Integer.

  tejanium

• Revert alphabetical sorting of table columns inside schema.rb.

  Alphabetical sorting of table columns inside the schema creates improper production tables when using db:prepare.

  Bert McCutchen

• Deprecated ActiveRecord::ConnectionAdapters::Column#auto_populated? in favor of auto_populated_on_insert?

  Rafael Mendonça França

• Reload virtual columns on update in PostgreSQL

  Automatically reload virtual columns on update when using PostgreSQL. This is done by issuing a single UPDATE query that includes a RETURNING clause.

  Given a Post model represented by the following schema:

  create_table :posts do |t|
    t.integer :upvotes_count
    t.integer :downvotes_count
    t.virtual :total_votes_count, type: :integer, as: "upvotes_count + downvotes_count", stored: true
  end

  total_votes_count will reflect the sum of upvotes and downvotes after update is successfully called. Prior to this change calling reload would have been necessary to obtain the new value calculated by the database.

  post = Post.find(1)
  post.update(upvotes_count: 2, downvotes_count: 2)
  # Calling `post.reload` no longer necessary
  post.total_votes => 4

  Alex Baldwin

• Reset the optimistic locking column when a transaction is rolled back.

  Previously, when a record with optimistic locking was successfully saved inside a transaction that later rolled back, the in-memory lock_version was left at the incremented value while the database row was reverted to the previous one. Saving the same instance again then raised ::ActiveRecord::StaleObjectError because the WHERE clause used the incremented value that no longer existed in the database.

  The locking column is now restored from the snapshot taken at the start of the transaction, so retrying a save on the same record after a rollback works without an explicit reload.

  Kenta Ishizaki

• Fix handling of expressions in array syntax for add_index.

  This change allows passing expressions in array syntax for add_index method. For example, add_index :users, [ "lower(email)" ] now works the same as add_index :users, "lower(email)".

  # This now works properly:
  add_index :users, [ "lower(email)" ]
  
  # As does this:
  add_index :users, [ "lower(email)", :status ]

  Alexandre Camillo

• Fix strict_loading violations ignored when using pluck

  Johnson Chan

• Move the defaulting of prevent_writes to true when using the reading role into the parameters of the role switching methods, and raise an ArgumentError if prevent_writes: false is provided with the reading role.

  Joshua Young

• Fix incorrect callback execution order when config.active_record.run_after_transaction_callbacks_in_order_defined = true and using after_commit and after_rollback callbacks with prepend: true.

  Joshua Young

• Accept encryption credentials as ENV

  Taking advantage of Rails.apps.creds (#56455), the primary_key, deterministic_key and key_derivation_salt required by ActiveRecord::Encryption can now also be provided through environment variables named ACTIVE_RECORD_ENCRYPTION__PRIMARY_KEY, ACTIVE_RECORD_ENCRYPTION__DETERMINISTIC_KEY, ACTIVE_RECORD_ENCRYPTION__KEY_DERIVATION_SALT.

  Claudio Baccigalupo

• Treat nil values as "" during multi-parameter attribute assignment

  topic = Topic.where.not(last_read: nil).first
  topic.attributes = { "last_read(1i)" => nil, "last_read(2i)" => nil, "last_read(3i)" => nil }
  topic.last_read # => nil

  Sean Doyle

• Include record ID in error when uniqueness validation fails

  When a uniqueness validation fails, the errors.details hash for the attribute now includes an :existing_id key, holding the ID of the record that caused the conflict.

  # Before
  errors.details[:name]
  # => [{error: :taken, value: "John Doe"}]
  
  # After
  errors.details[:name]
  # => [{error: :taken, value: "John Doe", existing_id: 123}]

  Bruno Vicenzo

• Bump the minimum PostgreSQL version to 10.0.

  As part of this change, supports_pgcrypto_uuid? is deprecated because pgcrypto provides gen_random_uuid() since PostgreSQL 9.4, which is below the new 10.0 minimum.

  Yasuo Honda

• Let add_column raise ArgumentError if :null is set to a true value when defining a primary key.

  Primary keys get a NOT NULL constraint unconditionally. In particular, null: true was being ignored, thus not doing what the user specified. We should rather raise to let the user know the call is invalid.

  Xavier Noria

• Deprecate the schema_order option in PostgreSQL database configurations.

  Use schema_search_path instead. The schema_order alias will be removed in Rails 8.3.

  Eileen M. Uchitelle

• Deprecate the strict option in MySQL database configurations.

  The strict option for MySQL will be removed in Rails 8.3 because it is the default behavior.

  To change the default behavior of strict, use variables: { sql_mode: "..." } to configure sql_mode directly.

  strict: false can be replaced with variables: { sql_mode: "" }, and strict: :default can be replaced with variables: { sql_mode: :default }.

  Eileen M. Uchitelle

• Allow configuring SET queriers for the PostgreSQL and MySQL adapters.

  Individual settings can be skipped by setting them to false in database.yml, which is useful when connecting through a load balancer or proxy that handles configuration:

  PostgreSQL example:

  production:
    adapter: postgresql
    standard_conforming_strings: false
    intervalstyle: false
    min_messages: false
    schema_search_path: false
  

  MySQL example:

  production:
    adapter: mysql2
    wait_timeout: false
    variables:
      sql_mode: false
  

  Also deprecates set_standard_conforming_strings — it is now handled automatically through the consolidated settings hash.

  Eileen M. Uchitelle, Matthew Draper

• MySQL error 1046 (ER_NO_DB_ERROR: No database selected) is now retryable as a ConnectionFailed exception

  Clay Harmon

• Batch SQL statements when creating tables to improve performance.

  Andrew Novoselac

• Support PostgreSQL RESET on readonly queries.

  ActiveRecord::Base.connected_to(role: :reading, prevent_writes: true) do
    ActiveRecord::Base.with_connection do |c|
      c.execute("SET statement_timeout = '7s'")
      # some queries
      c.execute("RESET statement_timeout")
      # => no longer raises ActiveRecord::ReadOnlyError
    end
  end

  Francesco Rodriguez

• Add MySQL lock: option for add_index, remove_index, and ALTER TABLE column operations (add_column, remove_column, change_column, rename_column).

  Also extend algorithm: option support to ALTER TABLE column operations on MySQL.

  MySQL supports ALGORITHM = {DEFAULT|COPY|INPLACE|INSTANT} and LOCK = {DEFAULT|NONE|SHARED|EXCLUSIVE} to control how DDL operations are performed, enabling online schema changes without blocking reads or writes.

  add_index :users, :email, algorithm: :inplace, lock: :none
  remove_index :users, :email, algorithm: :inplace, lock: :none
  add_column :users, :name, :string, algorithm: :instant, lock: :none
  change_column :users, :name, :string, null: false, algorithm: :inplace, lock: :none
  remove_column :users, :name, algorithm: :inplace, lock: :none
  rename_column :users, :name, :full_name, algorithm: :inplace, lock: :none

  Dominik Darnel

• Avoid issuing a ROLLBACK statement following TransactionRollbackError during COMMIT.

  This prevents the unnecessary "WARNING: there is no transaction in progress" log spilled to stderr directly from libpq.

  Sorah Fukumori

• Add implicit_persistence_transaction hook for customizing transaction behavior.

  A new protected method implicit_persistence_transaction has been added that wraps persistence operations (save, destroy, touch) in a transaction. This method can be overridden in models to customize transaction behavior, such as setting a specific isolation level or skipping transaction creation when one is already open.

  Example skipping transaction creation if one is already open:

  class Account < ApplicationRecord
    private
      def implicit_persistence_transaction(connection, &block)
        if connection.transaction_open?
          yield
        else
          super
        end
      end
  end

  Israel P Valverde

• Pass sql query to query log tags.

  config.active_record.query_log_tags = [
    sql_length: ->(context) { context[:sql].length }
  ]

  fatkodima

• Speedup ActiveRecord::Migration.maintain_test_schema! when using multiple databases.

  Previously, Active Record would inefficiently connect twice to each database, now it only connects once per database to reverify the schema.

  Iliana Hadzhiatanasova

• Add unique_by option to insert_all!.

  Chedli Bourguiba

• Fix PostgreSQL schema dumping to handle schema-qualified table names in foreign_key references that span different schemas.

  # before
  add_foreign_key "hst.event_log_attributes", "hst.event_logs" # emits correctly because they're in the same schema (hst)
  add_foreign_key "hst.event_log_attributes", "hst.usr.user_profiles", column: "created_by_id" # emits hst.user.* when user.* is expected
  
  # after
  add_foreign_key "hst.event_log_attributes", "hst.event_logs"
  add_foreign_key "hst.event_log_attributes", "usr.user_profiles", column: "created_by_id"

  Chiperific

• Add PostgreSQLAdapter.register_type_mapping for custom SQL type registration.

  Third-party gems can now register custom type mappings without prepending internal methods:

  ActiveRecord::ConnectionAdapters::PostgreSQLAdapter.register_type_mapping do |type_map|
    type_map.register_type("geometry") do |oid, fmod, sql_type|
      MyGeometryType.new(sql_type)
    end
  end

  Callbacks execute in registration order.

  Abdelkader Boudih

• Yield the transaction object to the block when using with_lock.

  Ngan Pham

• Fix bug when current_transaction.isolation would not have been reset in test env.

  Additionally, extending the change in #55549 to handle requires_new: true.

  Kir Shatrov

• Allow schema_dump configuration to be an absolute path.

  Previously, the schema_dump configuration was always joined with the db_dir path. Now, if an absolute path is provided, it will be used as-is.

  Mike Dalessio

• Decode PostgreSQL bytea and money columns when they appear in direct query results.

  bytea columns are now decoded to binary-encoded Strings, and money columns are decoded to BigDecimal instead of String.

  ActiveRecord::Base.connection
       .select_value("select '\\x48656c6c6f'::bytea").encoding #=> Encoding::BINARY
  
  ActiveRecord::Base.connection
       .select_value("select '12.34'::money").class #=> BigDecimal

  Matthew Draper

• Add support for configuring migration strategy on a per-adapter basis.

  migration_strategy can now be set on individual adapter classes, overriding the global ActiveRecord.migration_strategy. This allows individual databases to customize migration execution logic:

  class CustomPostgresStrategy < ActiveRecord::Migration::DefaultStrategy
    def drop_table(*)
      # Custom logic specific to PostgreSQL
    end
  end
  
  ActiveRecord::ConnectionAdapters::PostgreSQLAdapter.migration_strategy = CustomPostgresStrategy

  Adrianna Chang

• Allow either explain format syntax for EXPLAIN queries.

  MySQL uses FORMAT=JSON whereas Postgres uses FORMAT JSON. We should be able to accept both formats as options.

  Gannon McGibbon

• On MySQL parallel test database table reset to use DELETE instead of TRUNCATE.

  Truncating on MySQL is very slow even on empty or nearly empty tables.

  As a result of this change auto increment counters are now no longer reset between test runs on MySQL and the SKIP_TEST_DATABASE_TRUNCATE environment variable no longer has any effect.

  Donal McBreen

• Fix inconsistency in PostgreSQL handling of unbounded time range types

  Use -infinity rather than NULL for the lower value of PostgreSQL time ranges when saving records with a Ruby range that begins with nil.

  create_table :products do |t|
    t.tsrange :period
  end
  class Product < ActiveRecord::Base; end
  
  t = Time.utc(2000)
  
  Product.create(period: t...nil)
  Product.create(period: nil...t)

  Previously this would create two records using different values to represent lower-unbounded and upper-unbounded ranges.

  ["2000-01-01 00:00:00",infinity)
  (NULL,"2000-01-01 00:00:00")

  Now both will use -infinity/infinity which are handled differently than NULL by some PostgreSQL range operators (e.g., lower_inf) and support both exclusive and inclusive bounds.

  ["2000-01-01 00:00:00",infinity)
  [-infinity,"2000-01-01 00:00:00")

  Martin-Alexander

• Database-specific shard swap prohibition

  In #43485 (v7.0.0), shard swapping prohibition was introduced as a global switch that applied to all databases.

  For the use case of a multi-database application, the global prohibition is overly broad, and so with this change the method prohibit_shard_swapping will scope the prohibition to the same connection class (i.e., connection_specification_name). This allows an application to prohibit shard swapping on a specific database while allowing it on all others.

  Mike Dalessio

• Fix upsert_all when using repeated timestamp attributes.

  Gannon McGibbon

• PostgreSQL enable drop database FORCE option.

  One of the benefits of developing with MySQL is that it allows dropping the current database without first disconnecting clients. As a result developers can use bin/rails db:reset and similar, without first shutting down instances of the app, Rails consoles, background workers, etc. By default PostgreSQL fails to drop a database when clients are connected and displays the following error:

  > PG::ObjectInUse: ERROR:  database "xyz" is being accessed by other users (PG::ObjectInUse)

  This is frustrating when working in development where the database may be dropped frequently.

  PostgreSQL 13 added the FORCE option to the DROP DATABASE statement (PostgreSQL docs) which automatically disconnects clients before dropping the database. This option is automatically enabled for supported PostgreSQL versions.

  Steven Webb

• Raise specific exception when a prohibited shard change is attempted.

  The new ShardSwapProhibitedError exception allows applications and connection-related libraries to more easily recover from this specific scenario. Previously an ArgumentError was raised, so the new exception subclasses ArgumentError for backwards compatibility.

  Mike Dalessio

• Fix SQLite3 data loss during table alterations with CASCADE foreign keys.

  When altering a table in SQLite3 that is referenced by child tables with ON DELETE CASCADE foreign keys, ActiveRecord would silently delete all data from the child tables. This occurred because SQLite requires table recreation for schema changes, and during this process the original table is temporarily dropped, triggering CASCADE deletes on child tables.

  The root cause was incorrect ordering of operations. The original code wrapped disable_referential_integrity inside a transaction, but PRAGMA foreign_keys cannot be modified inside a transaction in SQLite - attempting to do so simply has no effect. This meant foreign keys remained enabled during table recreation, causing CASCADE deletes to fire.

  The fix reverses the order to follow the official SQLite 12-step ALTER TABLE procedure: disable_referential_integrity now wraps the transaction instead of being wrapped by it. This ensures foreign keys are properly disabled before the transaction starts and re-enabled after it commits, preventing CASCADE deletes while maintaining data integrity through atomic transactions.

  Ruy Rocha

• Fix negative scopes for enums to include records with nil values.

  fatkodima

• Improve support for SQLite database URIs.

  The db:create and db:drop tasks now correctly handle SQLite database URIs, and the SQLite3Adapter will create the parent directory if it does not exist.

  Mike Dalessio

Please check [8-1-stable]) for previous changes.