trellis-pg 0.1.0.pre.alpha

data/lib/trellis/error.rb

@@ -0,0 +1,103 @@
+# frozen_string_literal: true
+
+module Trellis
+  # An error from Trellis: the engine's message and a stable #code, and
+  # nothing else (ADR-0010 decision 4). Every error the binding raises is one
+  # of the subclasses below, so `rescue Trellis::Error` catches them all, and
+  # `rescue Trellis::NotFoundError` just one kind.
+  #
+  # The engine's error codes are an open set, so a code this version of the
+  # binding doesn't know arrives as an UnknownError, with the engine's code
+  # named in the message, rather than crashing.
+  class Error < StandardError
+    # The code this class stands for, as a symbol; see #code.
+    CODE = nil
+
+    # One of :parse, :validation, :connectivity, :conflict, :not_found,
+    # :internal, :timeout, or :unknown for a code newer than this binding.
+    def code
+      self.class::CODE
+    end
+
+    # The exception for the native extension's `(code, message)` pair.
+    # Not part of the public API.
+    def self.from_native(code, message)
+      klass = BY_CODE[code]
+      return klass.new(message) if klass
+
+      UnknownError.new("(error code #{code.inspect}) #{message}")
+    end
+  end
+
+  # The statement text isn't valid grammar.
+  class ParseError < Error
+    CODE = :parse
+  end
+
+  # Well-formed but rejected: a bad option, an unknown column, a type
+  # mismatch, a call on a handle that has been shut down.
+  class ValidationError < Error
+    CODE = :validation
+  end
+
+  # The database couldn't be reached, or the connection dropped.
+  class ConnectivityError < Error
+    CODE = :connectivity
+  end
+
+  # Clashes with something that already exists.
+  class ConflictError < Error
+    CODE = :conflict
+  end
+
+  # Names something that doesn't exist.
+  class NotFoundError < Error
+    CODE = :not_found
+  end
+
+  # A Trellis bug or an unexpected database failure.
+  class InternalError < Error
+    CODE = :internal
+  end
+
+  # A bounded wait ran out of time, such as Trellis.await_converged's
+  # timeout_ms, or Trellis.release_key's wait for the pages in flight on its
+  # table, or a call ran out of its 30-second deadline (the server stopped
+  # what the call started and rolled it back). Expected, not a bug: retry it.
+  class TimeoutError < Error
+    CODE = :timeout
+  end
+
+  # An error code newer than this binding. The message names the code.
+  class UnknownError < Error
+    CODE = :unknown
+  end
+
+  # A call on a handle this process inherited through `fork` rather than
+  # connected itself. Rust threads don't cross `fork`, so the handle has
+  # nothing left to answer it: every call raises this instead of hanging.
+  # Call Trellis.connect in the forked process (Puma's before_worker_boot,
+  # Passenger's starting_worker_process: "Forking servers" in the README).
+  # A ValidationError, like any other call on a handle that can't serve it.
+  #
+  # Also Trellis.connect's error in a process forked while its parent had a
+  # handle running, which may have inherited a lock that handle's threads
+  # held (issue #600): shut down before forking (Puma's before_fork).
+  class ForkedHandleError < ValidationError
+  end
+
+  class Error
+    # The engine's codes, written out rather than derived, so adding one is a
+    # deliberate change here. test/error_test.rb fails when the engine
+    # reports a code this map lacks.
+    BY_CODE = {
+      "parse" => ParseError,
+      "validation" => ValidationError,
+      "connectivity" => ConnectivityError,
+      "conflict" => ConflictError,
+      "not_found" => NotFoundError,
+      "internal" => InternalError,
+      "timeout" => TimeoutError
+    }.freeze
+  end
+end

data/lib/trellis/instance.rb

@@ -0,0 +1,540 @@
+# frozen_string_literal: true
+
+require_relative "error"
+require_relative "values"
+
+module Trellis
+  # The largest `limit:` and `timeout_ms:` the engine takes (an i64 and a
+  # u64 millisecond count); anything larger is refused as a ValidationError
+  # rather than failing to convert in the extension.
+  MAX_LIMIT = (2**63) - 1
+  MAX_TIMEOUT_MS = (2**64) - 1
+  SELF_CHECK_MODES = %i[standard strict].freeze
+
+  # The argument checks every call makes before it reaches the extension.
+  module Checks # :nodoc:
+    private
+
+    def string_option!(name, value)
+      return if value.is_a?(String)
+
+      raise ValidationError, "#{name} must be a String, got: #{value.inspect}"
+    end
+
+    def string!(what, value)
+      raise ValidationError, "#{what} must be a String, got: #{value.inspect}" unless value.is_a?(String)
+    end
+
+    def limit!(limit)
+      return if limit.is_a?(Integer) && limit.between?(1, MAX_LIMIT)
+
+      raise ValidationError, "limit must be a positive Integer, got: #{limit.inspect}"
+    end
+
+    def timeout_ms!(timeout_ms)
+      return if timeout_ms.is_a?(Integer) && timeout_ms.between?(0, MAX_TIMEOUT_MS)
+
+      raise ValidationError, "timeout_ms must be a non-negative Integer, got: #{timeout_ms.inspect}"
+    end
+
+    def cursor!(cursor, what)
+      return if cursor.nil? || cursor.is_a?(String)
+
+      raise ValidationError, "after must be #{what}, got: #{cursor.inspect}"
+    end
+  end
+
+  # One Trellis instance: a catalog schema in a database, and the native
+  # handle that runs it from this process. A process holds as many as it has
+  # catalog schemas (one per database, or several in one database), each with
+  # its own connections and threads and no shared budget
+  # (docs/decisions/0010-embeddable-clients.md, decision 3).
+  #
+  #   analytics = Trellis::Instance.connect(url: "host=localhost dbname=app", schema: "analytics")
+  #   billing = Trellis::Instance.connect(url: "host=localhost dbname=app", schema: "billing")
+  #   analytics.define("TRANSFORM widget_prices FROM widgets SELECT price AS price")
+  #   billing.shutdown # analytics keeps running
+  #
+  # Every call the `Trellis` module makes is a method here: the module
+  # delegates to its default instance (see Trellis.connect). The surface
+  # mirrors the Rust crate's BlockingTrellis:
+  #
+  # - Lifecycle: connect, connected?, migrate, config, shutdown.
+  # - Every statement form: apply runs one statement of Trellis's grammar
+  #   (TRANSFORM, PAUSE TRANSFORM, RESUME TRANSFORM, DROP TRANSFORM, ALTER
+  #   TRANSFORM) and reports what it did
+  #   as an Applied. define is the one guarded convenience: an apply that
+  #   refuses anything but a TRANSFORM statement before applying it.
+  # - Reads: status, definitions, dump_definitions_sql.
+  # - Quarantine: quarantined, quarantine_status, sample_quarantined,
+  #   poisoned_since, release_key. Pausing and resuming are statements:
+  #   instance.apply("RESUME TRANSFORM order_totals.total").
+  # - Operations: request_backfill, detect_capture_disruptions, has_live_drain_workers?,
+  #   has_live_staging_worker?, the read-your-writes pair watermark_token and
+  #   await_converged, and the audit self_check, a background job that
+  #   self_check_job reads back.
+  #
+  # Every call blocks on the database with the GVL released, so other threads
+  # run meanwhile, and Thread#kill, Thread#raise, Timeout.timeout and Ctrl-C
+  # interrupt it, leaving no thread behind. An interrupt abandons the wait, not
+  # the work: the call itself ends within its 30-second deadline, where the
+  # server stops it. Every failure raises a Trellis::Error subclass, a call that
+  # runs out of time a Trellis::TimeoutError.
+  #
+  # Conventions: times are Times in UTC, at microsecond precision; a
+  # transform is named by its target table, bare ("order_totals") or qualified as
+  # #define returns it ("public.order_totals"), wherever a call takes one (#status,
+  # #quarantine_status, #sample_quarantined, #release_key, #self_check and the
+  # PAUSE, RESUME, DROP and ALTER statements); a quarantine target is an
+  # address string, a transform's name or "transform.column"
+  # ("order_totals.total", always the bare transform); the
+  # cursor sample_quarantined returns, and watermark_token's token, are opaque
+  # strings to pass back unchanged; and every symbol in a result comes from a
+  # closed set, never from a string the database returned.
+  #
+  # An instance doesn't survive `fork`. Shut every instance down before
+  # forking (Instance.shutdown_all; Puma's before_fork) and connect after
+  # (Puma's before_worker_boot, Passenger's starting_worker_process). A
+  # forked child's calls on an instance it inherited raise
+  # Trellis::ForkedHandleError, and its shutdown leaves the parent's handle
+  # alone and does nothing. So does Instance.connect in a child forked while
+  # any instance of its parent was running (issue #600).
+  class Instance
+    extend Checks
+    include Checks
+
+    @instances = []
+    @registry = Mutex.new
+    @at_exit_installed = false
+
+    class << self
+      # Connects a new instance and returns it. Raises ForkedHandleError in
+      # a process forked while its parent had an instance running
+      # (connected, connecting, or not yet fully shut down): it may have
+      # inherited a lock one of that instance's threads held (issue #600).
+      #
+      # - url: where the database is, as a libpq connection string
+      #   ("host=... dbname=...") or URL. Nothing is read from the environment.
+      # - schema: the schema Trellis keeps this instance's own tables in. Not
+      #   "public", not target_schema, and not another instance's schema or
+      #   target_schema; migrate refuses those.
+      # - target_schema: the schema a bare target table name is created in.
+      # - staging: whether this process runs the instance's staging worker,
+      #   which installs change capture on the source tables and starts each
+      #   new transform's backfill. Exactly one process in a fleet should,
+      #   per instance.
+      # - drain_threads: how many threads apply staged changes to the targets.
+      #   Some process must run at least one per instance, or no definition
+      #   reaches :live.
+      # - worker_threads: the worker threads of each Rust runtime this
+      #   instance owns (its calls', and its background client's), invisible
+      #   to Ruby's own thread sizing, so kept small and explicit. Each
+      #   instance has its own; nothing is shared between instances.
+      #
+      # The defaults (staging: false, drain_threads: 0) run nothing in the
+      # background, which is what a migration step or a console wants. The
+      # staging worker reads Trellis's tables as it starts, so a staging: true
+      # connect to an unmigrated schema fails: migrate first.
+      def connect(url:, schema: "trellis", target_schema: "public", staging: false,
+                  drain_threads: 0, worker_threads: 2)
+        string_option!(:url, url)
+        string_option!(:schema, schema)
+        string_option!(:target_schema, target_schema)
+        unless [true, false].include?(staging)
+          raise ValidationError, "staging must be true or false, got: #{staging.inspect}"
+        end
+        unless drain_threads.is_a?(Integer) && drain_threads >= 0
+          raise ValidationError,
+                "drain_threads must be a non-negative Integer, got: #{drain_threads.inspect}"
+        end
+        unless worker_threads.is_a?(Integer) && worker_threads >= 1
+          raise ValidationError,
+                "worker_threads must be a positive Integer, got: #{worker_threads.inspect}"
+        end
+
+        handle = Native.connect(url, schema, target_schema, staging, drain_threads, worker_threads)
+        instance = new(handle, schema)
+        register(instance)
+        instance
+      end
+
+      # The instances this process holds, the default one included: every
+      # one it connected and hasn't shut down, minus any a forked child only
+      # inherited.
+      def connected
+        snapshot.select(&:connected?)
+      end
+
+      # Shuts down every instance this process holds, the `Trellis` module's
+      # default instance included, and waits for each. Does nothing for an
+      # instance a forked child only inherited. Every instance is attempted;
+      # if some fail, the first error is raised once the rest are done and
+      # the others are warned about. What a forking server calls before it
+      # forks (Puma's before_fork), and the at_exit hook's backstop.
+      def shutdown_all
+        errors = snapshot.filter_map do |instance|
+          instance.shutdown
+          nil
+        rescue Error => e
+          e
+        end
+        return nil if errors.empty?
+
+        errors.drop(1).each { |e| warn "trellis: shutting down an instance failed: #{e.class}: #{e.message}" }
+        raise errors.first
+      end
+
+      private
+
+      # Removes a shut-down instance from the registry.
+      def deregister(instance)
+        @registry.synchronize { @instances.delete(instance) }
+      end
+
+      def register(instance)
+        @registry.synchronize do
+          @instances << instance
+          install_at_exit
+        end
+      end
+
+      def snapshot
+        @registry.synchronize { @instances.dup }
+      end
+
+      def install_at_exit
+        return if @at_exit_installed
+
+        @at_exit_installed = true
+        at_exit { shutdown_all_at_exit }
+      end
+
+      # The backstop for a process that exits without shutting its instances
+      # down. A forked child that inherited them leaves them alone (see
+      # #shutdown).
+      def shutdown_all_at_exit
+        shutdown_all
+      rescue Error => e
+        warn "trellis: shutting down at exit failed: #{e.class}: #{e.message}"
+      end
+    end
+
+    def initialize(handle, schema)
+      @handle = handle
+      @schema = schema
+      @lock = Mutex.new
+    end
+    private_class_method :new
+
+    # Whether this instance holds a live handle. False after shutdown, and
+    # in a forked child that hasn't connected its own.
+    def connected?
+      handle = @handle
+      !handle.nil? && handle.owner_pid == Process.pid
+    end
+
+    # Creates or upgrades Trellis's own tables in the configured schema. Safe
+    # to run on every deploy.
+    def migrate
+      handle.migrate
+      nil
+    end
+
+    # The Config this instance connected with.
+    def config
+      Config.new(**handle.config)
+    end
+
+    # Registers a transform from a `TRANSFORM` statement, creates its target
+    # table, and returns the target table (fully qualified, "schema.table"),
+    # at :waiting_to_backfill. Its backfill runs in the background: poll
+    # #status with the returned name for :live.
+    #
+    # Only a `TRANSFORM` statement: any other form (DROP, PAUSE, ...) raises
+    # ValidationError without being applied, and a statement that doesn't
+    # parse raises ParseError. #apply takes every form.
+    #
+    # Trellis runs this on its own connections, not in the caller's
+    # transaction: if an enclosing migration rolls back, the definition stays.
+    def define(statement)
+      string!("the statement", statement)
+      handle.define(statement)
+    end
+
+    # Runs one statement of Trellis's grammar, whatever its form, and returns
+    # an Applied saying what it did.
+    #
+    #   trellis.apply("TRANSFORM order_totals FROM orders SELECT price + tax AS total")
+    #   # => #<data Trellis::Applied kind=:transform_defined, target_table="public.order_totals", ...>
+    #   trellis.apply("RESUME TRANSFORM order_totals.total").columns
+    #   # => ["order_totals.total"]
+    #
+    # A statement that doesn't parse raises ParseError, and nothing is
+    # applied. Like #define, this runs on Trellis's own connections, not in
+    # the caller's transaction.
+    #
+    # An Applied of kind :unknown is a success like any other: the statement
+    # took effect, and only its outcome is newer than this binding can
+    # describe. Don't retry it; read the result back with #status or
+    # #definitions if you need it.
+    def apply(statement)
+      string!("the statement", statement)
+      Applied.from_native(handle.apply(statement))
+    end
+
+    # The Status of the definition writing target_table, or nil when none
+    # does. target_table is either the fully qualified name #define and #apply
+    # return ("public.order_totals") or the bare table name ("order_totals");
+    # a qualified name with another schema matches nothing.
+    def status(target_table)
+      string!("the target table", target_table)
+      hash = handle.status(target_table)
+      hash && Status.from_native(hash)
+    end
+
+    # Every registered transform definition, oldest first, as
+    # DefinitionSummary values.
+    def definitions
+      handle.definitions.map { |hash| DefinitionSummary.from_native(hash) }
+    end
+
+    # The SQL that carries this instance's definitions in a schema-only dump
+    # of its database: Trellis's migration history, instance marker,
+    # transform definitions, and the seed rows capture
+    # needs. Append it to the output of `pg_dump --schema-only` (with the
+    # instance's schema and target schema in scope) and a database loaded
+    # from the file has every definition, each at :waiting_to_backfill, as
+    # `define` leaves it; the worker builds them. Runtime state (the ledger,
+    # ring, chunks, quarantine, claims and statuses) is never in it.
+    #
+    # Read from one snapshot, but a different one from the pg_dump's: a
+    # define or drop that commits between the two leaves them disagreeing.
+    # The Rails hook (Trellis::SchemaDump) reads it on both sides of the
+    # pg_dump and retries when the two differ.
+    def dump_definitions_sql
+      handle.dump_definitions_sql
+    end
+
+    # Rebuilds every definition that reads source_table (a bare table name,
+    # resolved like one in a statement), and returns once that rebuild is
+    # queued.
+    #
+    # Each :live reader reports :backfilling (rebuilt by the call itself)
+    # until the re-read has re-derived every row the table still has and
+    # deleted any target row it no longer backs. A newly defined transform
+    # doesn't need this: its backfill is queued for it. Only a table Trellis
+    # already captures can be re-read; any other raises NotFoundError.
+    def request_backfill(source_table)
+      string!("the source table", source_table)
+      handle.request_backfill(source_table)
+      nil
+    end
+
+    # Releases one key `transform` holds in quarantine, once its cause is
+    # fixed. source_table ("public.orders" or "orders") and key are as
+    # sample_quarantined and poisoned_since report them.
+    #
+    # It stages a recompute of the key, which every transform reading the
+    # table applies from the key's current row, and discards the changes held
+    # for `transform` meanwhile. Another transform holding the same key keeps
+    # holding it. If the cause is still there, the key is poisoned again.
+    # Resuming the transform releases every key it holds.
+    #
+    # An unknown transform raises NotFoundError, and so does a key it doesn't
+    # hold, which changes nothing. The release first waits for the drain
+    # pages in flight on the table to commit; a wait past the lock timeout
+    # (30 seconds) raises TimeoutError and changes nothing: call it again.
+    #
+    #   trellis.release_key("order_totals", "public.orders", "42")
+    def release_key(transform, source_table, key)
+      string!("the transform", transform)
+      string!("the source table", source_table)
+      string!("the key", key)
+      handle.release_key(transform, source_table, key)
+      nil
+    end
+
+    # Every source row the apply path poisoned (gave up on) after `since` (a
+    # Time), oldest first, as PoisonEntry values. Poll it with the last
+    # entry's poisoned_at so a whole-table failure doesn't sit unnoticed.
+    def poisoned_since(since)
+      raise ValidationError, "since must be a Time, got: #{since.inspect}" unless since.is_a?(::Time)
+
+      micros = EpochMicros.from_time(since) or
+        raise ValidationError, "since is out of range: #{since.inspect}"
+      handle.poisoned_since(micros).map { |hash| PoisonEntry.from_native(hash) }
+    end
+
+    # Every quarantined transform and paused column, across every transform,
+    # as QuarantineEntry values. Cheap enough for a dashboard or health
+    # check to poll.
+    def quarantined
+      handle.quarantined.map { |hash| QuarantineEntry.from_native(hash) }
+    end
+
+    # The QuarantineEntry of one target: a transform ("order_totals") or one
+    # of its columns ("order_totals.total"). A column that isn't paused is
+    # :live.
+    def quarantine_status(target)
+      string!("the target", target)
+      QuarantineEntry.from_native(handle.quarantine_status(target))
+    end
+
+    # One SamplePage of the rows quarantined under target, to diagnose a
+    # quarantine's cause: for a column ("order_totals.total"), the rows that
+    # failed evaluating it; for a whole transform ("order_totals"), the keys
+    # poisoned from its source table.
+    #
+    # - limit: the most rows to return.
+    # - after: the previous page's next_cursor; nil for the first page.
+    #
+    #   page = trellis.sample_quarantined("order_totals.total", limit: 50)
+    #   page = trellis.sample_quarantined("order_totals.total", limit: 50, after: page.next_cursor)
+    def sample_quarantined(target, limit: 100, after: nil)
+      string!("the target", target)
+      limit!(limit)
+      cursor!(after, "a cursor from a previous page")
+      SamplePage.from_native(handle.sample_quarantined(target, after, limit))
+    end
+
+    # Whether at least one drain worker is alive anywhere in the fleet. With
+    # none, nothing reaches a target table: poll this from a health check.
+    def has_live_drain_workers?
+      handle.has_live_drain_workers
+    end
+
+    # Whether the staging worker (which installs change capture and starts
+    # backfills) is alive anywhere in the fleet. The other half of the health check.
+    def has_live_staging_worker?
+      handle.has_live_staging_worker
+    end
+
+    # Pauses, now, the definitions that read a table whose Trellis capture
+    # was changed outside Trellis (a trigger disabled, dropped or recreated, a
+    # capture function replaced, the table gone), and returns the qualified
+    # names of those tables. The staging worker's reconcile pass does the same
+    # on its own cadence, with a CaptureFailure of kind :disruption on each
+    # paused definition's status; this is for a caller that has just done
+    # something that disrupts capture, like a fixture load, and needs the
+    # readers paused before it goes on. Installs, repairs and rebuilds
+    # nothing, and resuming is still the caller's (Trellis::TestSupport does
+    # it for a test suite). Safe beside a running staging worker.
+    def detect_capture_disruptions
+      handle.detect_capture_disruptions
+    end
+
+    # A read-your-writes token (an opaque String) covering every write
+    # committed before this call. Take it after a source-table write commits,
+    # then pass it to #await_converged to wait for that write to reach its
+    # targets.
+    def watermark_token
+      handle.watermark_token
+    end
+
+    # Waits until every change committed at or before `token` has reached its
+    # target tables, or timeout_ms passes (a TimeoutError; retry it). Returns
+    # nil.
+    #
+    # It waits for captured changes only: a transform that isn't :live yet
+    # may still be missing rows when this returns (see #status).
+    #
+    # Like every call, it returns within 30 seconds: a longer timeout_ms is
+    # cut to that, and a caller that wants to wait longer calls again. Calls
+    # from other threads don't wait behind it.
+    def await_converged(token, timeout_ms:)
+      string!("the token", token)
+      timeout_ms!(timeout_ms)
+      handle.await_converged(token, timeout_ms)
+      nil
+    end
+
+    # Starts a background check of target_table against a fresh recompute of
+    # its definition from the source tables, and returns a SelfCheckJob at
+    # once. The comparison is a drain worker's, a page of keys at a time
+    # (with no drain worker anywhere in the fleet the job stays :queued, as
+    # a define does); poll self_check_job with the job's id until it is
+    # finished, and its report is the verdict.
+    #
+    # - timeout_ms: how long a page of the job waits for the target to catch
+    #   up, per wait. A :standard check waits up to twice per page, a
+    #   :strict one once.
+    # - mode: :standard re-checks anything that differs after a fresh wait,
+    #   so a change still in flight isn't reported; it is safe while the
+    #   source is being written. :strict skips the re-check, and is only
+    #   sound once writes to the audited tables have stopped.
+    #
+    # A target whose job is still :queued or :running gets that job back,
+    # whatever mode and timeout_ms this call passed; a finished job stays
+    # until the next self_check of its target replaces it. Only a
+    # one-row-per-source-key transform can be audited: an aggregate target
+    # raises ValidationError, and an unknown one NotFoundError. A column that is
+    # paused is left out of the comparison.
+    #
+    #   job = trellis.self_check("order_totals", timeout_ms: 30_000)
+    #   sleep 1 until (job = trellis.self_check_job(job.id)).finished?
+    #   job.report.outcome # => :converged
+    #
+    # The call reads nothing of the target, so it returns well inside the
+    # 30-second call limit, holding no connection afterwards.
+    def self_check(target_table, timeout_ms:, mode: :standard)
+      string!("the target table", target_table)
+      timeout_ms!(timeout_ms)
+      unless SELF_CHECK_MODES.include?(mode)
+        raise ValidationError, "mode must be :standard or :strict, got: #{mode.inspect}"
+      end
+
+      SelfCheckJob.from_native(handle.self_check(target_table, mode.to_s, timeout_ms))
+    end
+
+    # The SelfCheckJob that self_check returned with this id, as it stands
+    # now, or nil when there is none: a newer self_check of its target
+    # replaced it, or its transform was dropped.
+    def self_check_job(id)
+      unless id.is_a?(Integer) && id.between?(-(2**63), MAX_LIMIT)
+        raise ValidationError, "the job id must be an Integer, got: #{id.inspect}"
+      end
+
+      hash = handle.self_check_job(id)
+      hash && SelfCheckJob.from_native(hash)
+    end
+
+    # Stops this instance: its background work, its connections and its
+    # runtime thread, and waits for them. The other instances of the process
+    # are untouched. Does nothing if not connected, which includes a forked
+    # child holding only the instance it inherited: that one is the parent's,
+    # so it's left in place, untouched, and the child's other calls on it
+    # still raise ForkedHandleError. A child's cleanup code (Puma's
+    # on_worker_shutdown, say) can call this whether or not the child
+    # connected. A shut-down instance is spent: connect a new one.
+    def shutdown
+      @lock.synchronize do
+        return nil unless connected?
+
+        handle = @handle
+        begin
+          handle.shutdown
+        ensure
+          # Also when the shutdown raised or was interrupted: its call still
+          # takes the engine out of the handle, so the handle is spent.
+          @handle = nil
+          self.class.send(:deregister, self)
+        end
+      end
+      nil
+    end
+
+    # Shows the schema, never the connection string.
+    def inspect
+      "#<#{self.class.name} schema=#{@schema.inspect} #{@handle.nil? ? 'shut down' : 'connected'}>"
+    end
+    alias to_s inspect
+
+    private
+
+    def handle
+      @handle or raise ValidationError,
+                       "this Trellis instance (schema #{@schema.inspect}) is not connected: " \
+                       "it has been shut down"
+    end
+  end
+end