neo4j-ruby-driver 6.0.3.alpha.0 → 6.2.1.beta.2

data/lib/neo4j/driver/session.rb

@@ -0,0 +1,577 @@
+# frozen_string_literal: true
+
+module Neo4j
+  module Driver
+    # Per-operation connection acquisition: each `run` / `begin_transaction` /
+    # `execute_read` / `execute_write` acquires a fresh connection from the
+    # driver with that operation's access mode and database, and releases it
+    # back to the pool when the operation completes (Result fully consumed,
+    # or transaction committed/rolled back). The session itself does not
+    # hold a connection — only its in-flight operation does.
+    class Session
+      include Internal::DurationNormalizer
+
+      def initialize(connection_provider, options = {}, clock: Internal::Clock.new)
+        @connection_provider = connection_provider
+        @options = options
+        @clock = clock
+        @transaction = nil
+        @open = true
+        @last_bookmarks = Set.new
+        @current_result = nil
+        # The home database this session pinned after its first resolution
+        # (Optimization:HomeDatabaseCache); nil until then / for explicit-db sessions.
+        @pinned_database = nil
+        @used_guess = false # whether an operation optimistically guessed the home db
+        @bookmark_manager = options[:bookmark_manager]
+        # The bookmark snapshot we sent on the most recent BEGIN — used
+        # as `previous_bookmarks` when forwarding to the manager so it
+        # set-difference-drops what it knew before. Computed at BEGIN
+        # time (not update time) because the manager may have grown
+        # between the two via another session.
+        @bookmarks_used_on_begin = nil
+
+        # Initialize with provided bookmarks if any
+        return unless options[:bookmarks]
+
+        initial_bookmarks = options[:bookmarks]
+        initial_bookmarks = [initial_bookmarks] unless initial_bookmarks.is_a?(Enumerable)
+        initial_bookmarks.each do |bookmark|
+          @last_bookmarks << (bookmark.is_a?(Bookmark) ? bookmark : Bookmark.from(bookmark))
+        end
+      end
+
+      def run(query, parameters = {}, config = {})
+        Internal::Validator.require_query_text!(query)
+        parameters ||= {}
+
+        unless parameters.is_a?(Hash)
+          raise ArgumentError,
+                "The parameters should be provided as Map type. Unsupported parameters type: #{parameters.class}"
+        end
+
+        raise Exceptions::ClientException, 'Session is closed' unless @open
+        if @transaction&.open?
+          raise Exceptions::ClientException,
+                'You cannot run a query directly on a session while a transaction is open'
+        end
+
+        drain_current_result
+
+        timeout = config.delete(:timeout)
+        tx_metadata = config.delete(:metadata)
+
+        # `mode` is sent for read sessions (matches routing-server
+        # expectations); writers are the default and omit the field.
+        # `bookmarks` is the current bookmark snapshot so server-side
+        # causal consistency works across auto-commit runs without an
+        # explicit transaction.
+        # `imp_user` (Bolt 4.4+) makes the server run the query as if
+        # the named user had issued it — auth as the session's user,
+        # authz as the impersonated one.
+        bookmarks = current_bookmarks_for_extra
+        # Acquire (with its own bookmark snapshot) so the home-db cache can pick
+        # the routing table: on a cache guess the RUN sends db=nil (server
+        # re-resolves), otherwise it pins the discovered name. run_db carries
+        # whichever applies; run_extra keeps this operation's bookmark snapshot.
+        connection, run_db = acquire_for_operation(session_access_mode)
+        run_extra = {
+          db: run_db,
+          mode: (session_access_mode == :read ? 'r' : nil),
+          tx_timeout: timeout_to_milliseconds(timeout),
+          tx_metadata:,
+          imp_user: @options[:impersonated_user],
+          bookmarks: bookmarks
+        }
+        run_extra.reject!(&Internal::Extras::BLANK)
+
+        fetch_size = effective_fetch_size
+        # Feature:IdempotentRetries — an auto-commit RUN whose failure the server
+        # flags `_idempotent` may be retried once: RESET clears the FAILED state
+        # (and drains the pipelined PULL's IGNORED), then RUN+PULL are re-sent
+        # (TELEMETRY is not). Only the RUN reply is eligible — a telemetry failure
+        # or a later stream (PULL) failure is raised as usual.
+        retries_left = auto_commit_retries_enabled? ? 1 : 0
+
+        buffer = handler = run_response = nil
+        telemetry_sent = false # TELEMETRY 2 = auto-commit; sent once, never resent on a retry
+        telemetry_pending = false # its (opted-in) SUCCESS is still owed
+        loop do
+          # A fresh stream per attempt — a prior attempt's buffer saw the IGNORED.
+          buffer = Bolt::RecordBuffer.new(fetch_size: fetch_size)
+          handler = Bolt::StreamHandler.new(buffer)
+          retry_run = false
+          # `stage` tells the rescue which reply raised: only a :run failure is
+          # idempotent-retryable (not :telemetry, nor a send/flush transport error).
+          stage = nil
+          run_response =
+            begin
+              # TELEMETRY is pipelined ahead of the first RUN when the server opted
+              # in; kept inside this block so a send failure still releases the lease.
+              unless telemetry_sent
+                telemetry_sent = true
+                telemetry_pending = connection.telemetry(2, disabled: @options[:telemetry_disabled])
+              end
+              # Auto-commit RUN carries this session's NotificationsConfig (5.2+);
+              # the tx path puts it on BEGIN instead. nil / non-5.2 => absent.
+              connection.send_message(connection.protocol.build_run(query, parameters, run_extra,
+                                                                    notification_config: @options[:notification_config]))
+              connection.send_message(connection.protocol.build_pull(n: fetch_size), handler)
+              connection.flush
+              if telemetry_pending
+                telemetry_pending = false
+                stage = :telemetry
+                connection.fetch_response.assert_success!
+              end
+              stage = :run
+              connection.fetch_response.assert_success!
+            rescue Exceptions::Neo4jException => e
+              # Classify first: this notifies the auth-token manager (and, for a
+              # security failure, sets auth_failed so the guards below hold) and,
+              # for routed connections, fires on_write_failure / deactivate and
+              # swaps NotALeader for SessionExpired. assert_success! raises
+              # *outside* RoutedConnection's wrapper, so this is the only place the
+              # classifier sees FAILURE responses to RUN.
+              classified = connection.classify_failure(e)
+              if stage == :run && retries_left.positive? && idempotent_error?(e) && !connection.auth_failed
+                retries_left -= 1
+                # RESET clears FAILED and drains the abandoned RUN/PULL replies.
+                connection.reset!
+                retry_run = true
+                nil
+              else
+                # Server is in FAILED state; RESET so the connection is immediately
+                # reusable — but not if it's being discarded (auth failure: the
+                # server closes it, RESET would just error).
+                connection.reset! unless connection.auth_failed
+                @connection_provider.release(connection)
+                raise classified
+              end
+            rescue StandardError
+              # Transport-level failure (IO/socket) — the connection is likely
+              # dead. Return it to the pool either way so this lease doesn't leak;
+              # the next user will rediscover the breakage.
+              @connection_provider.release(connection)
+              raise
+            end
+          break unless retry_run
+        end
+
+        # A home-db RUN that sent db=nil comes back with the server's resolved
+        # home database — cache it so the next same-identity session can guess.
+        cache_home_db_from(run_response)
+
+        keys = (run_response.metadata[:fields] || run_response.metadata['fields'] || []).map(&:to_sym)
+        @current_result = Result.new(
+          connection, keys, buffer: buffer, handler: handler,
+                            query_text: query, parameters: parameters, run_metadata: run_response.metadata,
+                            fetch_size: fetch_size,
+                            on_summary: method(:harvest_auto_commit_bookmark),
+                            on_release: -> { @connection_provider.release(connection) }
+        )
+      end
+
+      def begin_transaction(timeout: nil, metadata: nil, &block)
+        raise Exceptions::ClientException, 'Session is closed' unless @open
+        if @transaction&.open?
+          raise Exceptions::ClientException,
+                'You cannot begin a transaction on a session with an open transaction; either run from within the transaction or use a different session.'
+        end
+
+        drain_current_result
+        @current_result = nil
+
+        tx_options = @options.merge(
+          # The resolved home database is added in open_transaction, where the
+          # BEGIN bookmark snapshot is computed (so resolution and BEGIN share
+          # one snapshot).
+          timeout: timeout_to_milliseconds(timeout),
+          metadata:,
+          # BEGIN carries `mode: "r"` for read sessions (write is the
+          # server default and omits it), matching the auto-commit path.
+          # The Transaction reads tx_options[:access_mode]; without this an
+          # explicit read transaction dropped the field (session stores the
+          # mode under :default_access_mode, not :access_mode).
+          access_mode: (session_access_mode == :read ? 'r' : nil)
+        ).compact
+
+        # TELEMETRY 1 = unmanaged (explicit) transaction.
+        @transaction = open_transaction(session_access_mode, tx_options, telemetry_api: 1)
+
+        if block_given?
+          begin
+            result = yield @transaction
+            # Explicit-block transactions default to rollback; user must call
+            # tx.commit to persist changes (matches Java driver semantics).
+            @transaction.rollback if @transaction.open?
+            result
+          rescue StandardError => e
+            @transaction.rollback if @transaction.open?
+            raise e
+          ensure
+            @transaction = nil
+          end
+        else
+          @transaction
+        end
+      end
+
+      def execute_read(timeout: nil, metadata: nil, &)
+        execute_transaction(AccessMode::READ, timeout:, metadata:, &)
+      end
+
+      def execute_write(timeout: nil, metadata: nil, &)
+        execute_transaction(AccessMode::WRITE, timeout:, metadata:, &)
+      end
+
+      # Close any pending result by asking the server to abandon
+      # remaining records (DISCARD) rather than pulling them just to
+      # throw them away client-side. Matches Java's behaviour and is
+      # required by testkit's session_run.test_discard_on_session_close_*
+      # scripts which lie about has_more=true to verify the driver
+      # eventually sends DISCARD. With pagination, draining via buffer
+      # would loop forever on those scripts.
+      def close
+        return unless @open
+
+        pending_error = nil
+        begin
+          @transaction&.close # tx releases its own connection
+          if @current_result
+            connection = @current_result.connection
+            begin
+              @current_result.consume
+            rescue Exceptions::Neo4jException => e
+              pending_error = e
+            end
+            # Any failure during the consume leaves the connection in
+            # the server's FAILED state; RESET so the next borrower starts
+            # from a clean READY state.
+            connection.reset! if @current_result.failed?
+            @current_result.discard! # idempotent; releases the connection
+          end
+        ensure
+          @open = false
+        end
+
+        raise pending_error if pending_error
+      end
+
+      def open?
+        @open
+      end
+
+      attr_reader :last_bookmarks
+
+      def update_bookmarks(bookmarks)
+        # Replace bookmarks (don't accumulate) — matches Java driver behavior.
+        # Each committed transaction generates a new bookmark that replaces the previous one.
+        new_set = Set.new(Array(bookmarks).map(&Bookmark.method(:from)))
+        # Forward to the BookmarkManager (if configured) so cross-session
+        # causal consistency works. Pass the bookmarks we *sent* in BEGIN
+        # as `previous` — Java's manager uses set-difference, so passing
+        # the session's own bookmarks alongside the manager's is safe.
+        @bookmark_manager&.update_bookmarks(@bookmarks_used_on_begin || @last_bookmarks, new_set)
+        @last_bookmarks = new_set
+      end
+
+      private
+
+      # The database name to put on the wire for this session's operations.
+      # An explicit `:database` is used as-is. For a home-db session (nil) on
+      # a routing driver, resolve the name from discovery once and reuse it
+      # (Bolt 4.4+/5.x return it in the ROUTE reply); the direct provider and
+      # the procedure-based protocols return nil, leaving the server to
+      # resolve the home db from an absent `db`.
+      # `bookmarks` is the snapshot the caller is already using for this
+      # operation's wire `bookmarks` — passed in (not recomputed) so a single
+      # operation never takes two different BookmarkManager snapshots.
+      # Acquire the operation's connection and decide the db it should send
+      # (Optimization:HomeDatabaseCache). Returns [connection, run_db]:
+      #   - explicit :database  -> acquire against it, send it.
+      #   - home-db + cache guess (SSR pool) -> acquire against the guessed home
+      #     db (reuses its routing table, no discovery) but send db=nil so the
+      #     server re-resolves; unless the acquired connection turns out non-SSR
+      #     (a mixed cluster can't re-route), in which case fall back to explicit
+      #     discovery and pin the resolved name.
+      #   - home-db, no guess -> discover, pin, send the resolved name.
+      def acquire_for_operation(access_mode)
+        # Its own bookmark snapshot for discovery/acquire — distinct from the
+        # RUN/BEGIN extra's snapshot, so a BookmarkManager supplier is consulted
+        # once per acquire (matching the reference drivers / the bookmark-manager
+        # tests), and home-db discovery routes with a fresh set.
+        bookmarks = current_bookmarks_for_extra
+        explicit = @options[:database]
+        return [do_acquire(access_mode, explicit, bookmarks), explicit] if explicit
+        # Once this session has resolved its home db, pin it for the rest of the
+        # session — later runs send it explicitly so a re-resolving server can't
+        # move the session to a different database mid-flight.
+        return [do_acquire(access_mode, @pinned_database, bookmarks), @pinned_database] if @pinned_database
+
+        imp_user = @options[:impersonated_user]
+        auth = @options[:auth_token]
+        # Guess only pays off when we already hold a fresh routing table for it:
+        # acquire against that table and send db=nil so the server resolves the
+        # real home db. If its table is stale (or the connection turns out
+        # non-SSR), fall through to explicit discovery — a ROUTE with db=nil that
+        # authoritatively resolves and pins the home db. The optimistic acquire
+        # and the fallback share one acquisition-timeout deadline.
+        guess = @connection_provider.home_db_guess(imp_user, auth)
+        if guess && @connection_provider.routing_table_fresh?(guess, access_mode)
+          deadline = acquisition_deadline
+          connection = do_acquire(access_mode, guess, bookmarks, deadline)
+          if connection.ssr_enabled?
+            # A real guess: the reply's db will pin the session (cache_home_db_from).
+            @used_guess = true
+            return [connection, nil]
+          end
+
+          @connection_provider.release(connection)
+          resolved = @connection_provider.home_database(bookmarks, imp_user, auth)
+          @pinned_database = resolved
+          return [do_acquire(access_mode, resolved, bookmarks, deadline), resolved]
+        end
+        resolved = @connection_provider.home_database(bookmarks, imp_user, auth)
+        @pinned_database = resolved
+        [do_acquire(access_mode, resolved, bookmarks), resolved]
+      end
+
+      def do_acquire(access_mode, database, bookmarks, deadline = nil)
+        @connection_provider.acquire(access_mode: access_mode, database: database, bookmarks: bookmarks,
+                                     imp_user: @options[:impersonated_user], auth: @options[:auth_token],
+                                     deadline: deadline)
+      end
+
+      # A single acquisition-timeout deadline (monotonic seconds) so a guessed
+      # acquire and its fallback share one budget, matching the reference drivers.
+      def acquisition_deadline
+        @clock.monotonic + (@options[:connection_acquisition_timeout]&.to_f || 60)
+      end
+
+      # Record the server's resolved home database from a RUN/BEGIN reply, so a
+      # later same-identity session guesses it. No-op for explicit-db sessions
+      # (their reply omits db) and the direct provider (cache_home_db is a no-op).
+      def cache_home_db_from(response)
+        db = response.metadata[:db]
+        return if db.nil? || @options[:database]
+
+        # First guess-based run learns the real home db from the reply — pin it
+        # for the session (only when we actually guessed; direct drivers never
+        # pin) and remember it for the driver-wide cache.
+        @pinned_database ||= db if @used_guess
+        @connection_provider.cache_home_db(@options[:impersonated_user], @options[:auth_token], db)
+      end
+
+      # Current bookmark snapshot as a plain array of strings, ready
+      # for the wire. Returns nil when empty so the BEGIN/RUN extras
+      # hash's `reject!` strips the key entirely (testkit's stub
+      # scripts strictly compare against omission).
+      #
+      # Folds in the BookmarkManager's bookmarks too (when configured)
+      # so cross-session causal consistency works. Records the merged
+      # set on @bookmarks_used_on_begin so update_bookmarks can pass
+      # it as `previous` to the manager. A custom manager can return
+      # plain strings rather than Bookmark instances, so coerce
+      # everything via Bookmark.from before reading .value.
+      def current_bookmarks_for_extra
+        from_manager = Array(@bookmark_manager&.bookmarks).map(&Bookmark.method(:from))
+        merged = @last_bookmarks | from_manager
+        @bookmarks_used_on_begin = merged
+        bookmarks = merged.to_a.map(&:value)
+        bookmarks unless bookmarks.empty?
+      end
+
+      # Records per PULL batch. Matches Java/Python defaults (1000). The
+      # user can override per-driver via the `fetch_size` option and the
+      # special value `-1` means "pull all records in one batch".
+      def effective_fetch_size
+        size = @options[:fetch_size]
+        size.nil? ? 1000 : size
+      end
+
+      # Feature:IdempotentRetries. A session-level AutoCommitRetriesMode wins;
+      # otherwise fall back to the driver-level `auto_commit_retries_disabled`
+      # default (merged into the session options), which defaults to enabled.
+      def auto_commit_retries_enabled?
+        case @options[:auto_commit_retries_mode]
+        when AutoCommitRetriesMode::ENABLED then true
+        when AutoCommitRetriesMode::DISABLED then false
+        else !@options[:auto_commit_retries_disabled]
+        end
+      end
+
+      # A server FAILURE the server marks retryable via `_idempotent` in its
+      # diagnostic record.
+      def idempotent_error?(error)
+        error.diagnostic_record&.[](:_idempotent) == true
+      end
+
+      # Default access mode for the session — drives connection routing for
+      # auto-commit `run` calls. execute_read / execute_write override per
+      # operation when going through `execute_transaction`.
+      def session_access_mode
+        case @options[:default_access_mode]
+        when AccessMode::READ then :read
+        else :write
+        end
+      end
+
+      # Auto-commit hook called from Result when its stream ends in SUCCESS.
+      # Explicit-tx Results don't trigger this — Transaction#commit harvests
+      # the bookmark from the COMMIT response itself.
+      def harvest_auto_commit_bookmark(summary)
+        # `metadata` is private on ResultSummary (deliberately not part of
+        # the public Java-shaped API). Bookmark only ever lives on the wire
+        # SUCCESS metadata; no public Summary accessor for it because Java's
+        # ResultSummary doesn't expose one either. send() bypasses the
+        # private check for this internal-only consumer.
+        bookmark = summary.send(:metadata)[:bookmark]
+        update_bookmarks(bookmark) if bookmark
+      end
+
+      # Pull any pending auto-commit result's records into memory so they
+      # remain accessible from the user's reference after the connection is
+      # released. The Result self-releases its connection on the SUCCESS
+      # this drives. If draining surfaces a FAILURE we RESET the connection
+      # before releasing so it goes back to the pool clean.
+      def drain_current_result
+        return unless @current_result
+
+        # Detach first so any failure path doesn't leave @current_result
+        # pointing at a connection we've already returned to the pool —
+        # a later session.close would otherwise reset!/discard! a slot
+        # that may already belong to another session.
+        result = @current_result
+        @current_result = nil
+        connection = result.connection
+
+        begin
+          result.buffer
+        rescue Exceptions::Neo4jException
+          connection.reset!
+          result.discard! # idempotent; routes through Result#release_connection
+          raise
+        end
+
+        return unless result.failed?
+
+        connection.reset!
+        result.discard!
+
+        # Successful drain: Result already released its connection from on_success.
+      end
+
+      def execute_transaction(access_mode, timeout: nil, metadata: nil, &)
+        raise Exceptions::ClientException, 'Session is closed' unless @open
+
+        # Drain any pending auto-commit result so its connection is released
+        # and its bookmark is harvested before the managed tx acquires its
+        # own connection (likely a different server in routing mode).
+        drain_current_result
+
+        # Default matches Java's Config.DEFAULT_MAX_TRANSACTION_RETRY_TIME
+        # (30s). The window matters for cluster failover: a leader election
+        # can take several seconds, during which the router keeps returning
+        # the old/no leader, so the managed-tx retry must keep rediscovering
+        # (with exponential backoff) well past a second or two.
+        max_retry_time = @options[:max_transaction_retry_time] || 30
+        start_time = @clock.realtime
+        errors = []
+        # TELEMETRY 0 (managed tx) is reported until the server acknowledges it:
+        # re-sent on each attempt whose telemetry never landed (e.g. the
+        # connection died before its SUCCESS), but not once acked — matching the
+        # reference drivers and the retry stub script. The callback flips this
+        # the moment a Transaction reads the telemetry SUCCESS.
+        telemetry_acked = false
+        on_telemetry_ack = -> { telemetry_acked = true }
+
+        op_mode = (access_mode == AccessMode::READ ? :read : :write)
+        tx_options = @options.merge(
+          # BEGIN carries `mode: "r"` for read; write is the server default
+          # and omits the field (reference drivers and the stub scripts
+          # expect `BEGIN {"db": ...}`, not `…"mode": "w"`). Matches the
+          # auto-commit and explicit-begin paths; `compact` drops the nil.
+          access_mode: (access_mode == AccessMode::READ ? 'r' : nil),
+          timeout: timeout_to_milliseconds(timeout),
+          metadata:
+        ).compact
+
+        loop do
+          # api 0 = managed tx; execute_query's session overrides it to 3
+          # (DRIVER_EXECUTE_QUERY). nil once the server has acked the report.
+          telemetry_api = telemetry_acked ? nil : (@options[:telemetry_api] || 0)
+          return run_managed_transaction(op_mode, tx_options, telemetry_api, on_telemetry_ack, &)
+        rescue Exceptions::ServiceUnavailableException, Exceptions::SessionExpiredException,
+               Exceptions::TransientException,
+               # AuthorizationExpired (server authz-cache expiry) is always
+               # retryable; SecurityRetryableException is an auth failure the
+               # token manager has flagged retryable (token refreshed) — both
+               # warrant another attempt on a fresh connection/identity.
+               Exceptions::AuthorizationExpiredException, Exceptions::SecurityRetryableException => e
+          errors << e
+
+          if @clock.realtime - start_time >= max_retry_time
+            # Retries exhausted — re-raise the LAST error with its real
+            # type (SessionExpired / Transient / ServiceUnavailable)
+            # rather than flattening everything to ServiceUnavailable;
+            # Java's transaction executor likewise rethrows the last
+            # error. testkit asserts on the type (e.g. a routed reader
+            # interruption must surface as SessionExpired), so the
+            # generic wrapper masked it. Earlier attempts ride along as
+            # suppressed.
+            e.add_suppressed(*errors[0...-1])
+            raise e
+          end
+
+          # Exponential backoff: 1s, 2s, 4s, ... matching Java driver defaults
+          sleep(2**(errors.size - 1))
+        end
+      end
+
+      def run_managed_transaction(op_mode, tx_options, telemetry_api, on_telemetry_ack, &)
+        if @transaction&.open?
+          raise Exceptions::ClientException,
+                'You cannot begin a transaction on a session with an open transaction'
+        end
+
+        # TELEMETRY 0 = managed transaction function; api is nil once acked so a
+        # retry doesn't repeat it.
+        @transaction = open_transaction(op_mode, tx_options, telemetry_api: telemetry_api,
+                                                             telemetry_ack: on_telemetry_ack)
+
+        begin
+          result = yield @transaction
+          @transaction.commit if @transaction.open?
+          result
+        rescue StandardError => e
+          @transaction.rollback if @transaction.open?
+          raise e
+        ensure
+          @transaction = nil
+        end
+      end
+
+      def open_transaction(op_mode, tx_options, telemetry_api:, telemetry_ack: nil)
+        # Acquire and decide the BEGIN db together (Optimization:HomeDatabaseCache):
+        # a cache guess picks the routing table and sends db=nil; otherwise the
+        # discovered/pinned name is sent. It takes its own bookmark snapshot for
+        # discovery. begin_db is dropped when nil via compact.
+        connection, begin_db = acquire_for_operation(op_mode)
+        # The BEGIN's own snapshot, taken last so current_bookmarks_for_extra
+        # records @bookmarks_used_on_begin as the set actually sent on BEGIN (the
+        # commit-time update_bookmarks reports it as `previous`).
+        bookmarks = current_bookmarks_for_extra
+        tx_options = tx_options.merge(database: begin_db).compact
+        Transaction.new(connection, self, bookmarks, tx_options, telemetry_api: telemetry_api,
+                                                                 telemetry_ack: telemetry_ack,
+                                                                 # executeQuery's session reports telemetry api 3 (DRIVER_EXECUTE_QUERY);
+                                                                 # only that path pipelines BEGIN + RUN + PULL (Optimization:ExecuteQueryPipelining).
+                                                                 pipelined: @options[:telemetry_api] == 3,
+                                                                 on_begin: method(:cache_home_db_from),
+                                                                 on_release: lambda {
+                                                                   @connection_provider.release(connection)
+                                                                 })
+      end
+    end
+  end
+end

data/lib/neo4j/driver/summary/database_info.rb

@@ -0,0 +1,25 @@
+# frozen_string_literal: true
+
+module Neo4j
+  module Driver
+    module Summary
+      # Which database the result came from. Java's
+      # org.neo4j.driver.summary.DatabaseInfo.
+      class DatabaseInfo
+        attr_reader :name
+
+        def initialize(db_data)
+          @name =
+            case db_data
+            when String then db_data
+            when Hash then db_data[:name]
+            end
+        end
+
+        def to_s
+          @name || 'unknown'
+        end
+      end
+    end
+  end
+end

data/lib/neo4j/driver/summary/gql_notification.rb

@@ -0,0 +1,38 @@
+# frozen_string_literal: true
+
+module Neo4j
+  module Driver
+    module Summary
+      # A GqlStatusObject whose status also maps to a Neo4j notification
+      # (the wire status carries a `neo4j_code`). It exposes the extra
+      # notification facets — position, classification and severity —
+      # which live in the diagnostic record under `_position` /
+      # `_classification` / `_severity`.
+      #
+      # The parsed accessors return nil for a value outside the known enum
+      # (the backend renders that as "UNKNOWN"); the raw_* accessors hand
+      # back whatever the server sent, untouched.
+      class GqlNotification < GqlStatusObject
+        def position
+          @data.dig(:diagnostic_record, :_position)&.then { Notification::Position.new(it) }
+        end
+
+        def raw_classification
+          @data.dig(:diagnostic_record, :_classification)
+        end
+
+        def classification
+          raw_classification if Notification::KNOWN_CATEGORIES.include?(raw_classification)
+        end
+
+        def raw_severity
+          @data.dig(:diagnostic_record, :_severity)
+        end
+
+        def severity
+          raw_severity if Notification::KNOWN_SEVERITIES.include?(raw_severity)
+        end
+      end
+    end
+  end
+end

data/lib/neo4j/driver/summary/gql_status_object.rb

@@ -0,0 +1,35 @@
+# frozen_string_literal: true
+
+module Neo4j
+  module Driver
+    module Summary
+      # A GQL status object, sourced from a Bolt 5.6+ summary `statuses`
+      # entry. Mirrors org.neo4j.driver.summary.GqlStatusObject. A plain
+      # status (e.g. "00000 — successful completion") carries only a status
+      # code, a description and a diagnostic record; the GqlNotification
+      # subtype adds the notification facets (position / classification /
+      # severity).
+      class GqlStatusObject
+        # The server may omit the standard diagnostic-record keys; the
+        # driver fills these defaults (matching Java) while leaving every
+        # other key — including ones the server set to an explicit null —
+        # exactly as received.
+        DIAGNOSTIC_RECORD_DEFAULTS = {
+          OPERATION: '', OPERATION_CODE: '0', CURRENT_SCHEMA: '/'
+        }.freeze
+
+        attr_reader :gql_status, :status_description
+
+        def initialize(data)
+          @data = data
+          @gql_status = data[:gql_status]
+          @status_description = data[:status_description]
+        end
+
+        def diagnostic_record
+          DIAGNOSTIC_RECORD_DEFAULTS.merge(@data[:diagnostic_record] || {})
+        end
+      end
+    end
+  end
+end