services/api/app/models/container.rb

   1 # Copyright (C) The Arvados Authors. All rights reserved.
   2 #
   3 # SPDX-License-Identifier: AGPL-3.0
   4
   5 require 'log_reuse_info'
   6 require 'whitelist_update'
   7 require 'safe_json'
   8 require 'update_priorities'
   9
  10 class Container < ArvadosModel
  11   include ArvadosModelUpdates
  12   include HasUuid
  13   include KindAndEtag
  14   include CommonApiTemplate
  15   include WhitelistUpdate
  16   extend CurrentApiClient
  17   extend DbCurrentTime
  18   extend LogReuseInfo
  19
  20   # Posgresql JSONB columns should NOT be declared as serialized, Rails 5
  21   # already know how to properly treat them.
  22   attribute :secret_mounts, :jsonbHash, default: {}
  23   attribute :runtime_status, :jsonbHash, default: {}
  24   attribute :runtime_auth_scopes, :jsonbArray, default: []
  25   attribute :output_storage_classes, :jsonbArray, default: lambda { Rails.configuration.DefaultStorageClasses }
  26   attribute :output_properties, :jsonbHash, default: {}
  27
  28   serialize :environment, Hash
  29   serialize :mounts, Hash
  30   serialize :runtime_constraints, Hash
  31   serialize :command, Array
  32   serialize :scheduling_parameters, Hash
  33
  34   after_find :fill_container_defaults_after_find
  35   before_validation :fill_field_defaults, :if => :new_record?
  36   before_validation :set_timestamps
  37   before_validation :check_lock
  38   before_validation :check_unlock
  39   validates :command, :container_image, :output_path, :cwd, :priority, { presence: true }
  40   validates :priority, numericality: { only_integer: true, greater_than_or_equal_to: 0 }
  41   validate :validate_runtime_status
  42   validate :validate_state_change
  43   validate :validate_change
  44   validate :validate_lock
  45   validate :validate_output
  46   after_validation :assign_auth
  47   before_save :sort_serialized_attrs
  48   before_save :update_secret_mounts_md5
  49   before_save :scrub_secrets
  50   before_save :clear_runtime_status_when_queued
  51   after_save :update_cr_logs
  52   after_save :handle_completed
  53
  54   has_many :container_requests, {
  55              class_name: 'ContainerRequest',
  56              foreign_key: 'container_uuid',
  57              primary_key: 'uuid',
  58            }
  59   belongs_to :auth, {
  60                class_name: 'ApiClientAuthorization',
  61                foreign_key: 'auth_uuid',
  62                primary_key: 'uuid',
  63                optional: true,
  64              }
  65
  66   api_accessible :user, extend: :common do |t|
  67     t.add :command
  68     t.add :container_image
  69     t.add :cwd
  70     t.add :environment
  71     t.add :exit_code
  72     t.add :finished_at
  73     t.add :locked_by_uuid
  74     t.add :log
  75     t.add :mounts
  76     t.add :output
  77     t.add :output_path
  78     t.add :priority
  79     t.add :progress
  80     t.add :runtime_constraints
  81     t.add :runtime_status
  82     t.add :started_at
  83     t.add :state
  84     t.add :auth_uuid
  85     t.add :scheduling_parameters
  86     t.add :runtime_user_uuid
  87     t.add :runtime_auth_scopes
  88     t.add :lock_count
  89     t.add :gateway_address
  90     t.add :interactive_session_started
  91     t.add :output_storage_classes
  92     t.add :output_properties
  93     t.add :cost
  94     t.add :subrequests_cost
  95   end
  96
  97   # Supported states for a container
  98   States =
  99     [
 100      (Queued = 'Queued'),
 101      (Locked = 'Locked'),
 102      (Running = 'Running'),
 103      (Complete = 'Complete'),
 104      (Cancelled = 'Cancelled')
 105     ]
 106
 107   State_transitions = {
 108     nil => [Queued],
 109     Queued => [Locked, Cancelled],
 110     Locked => [Queued, Running, Cancelled],
 111     Running => [Complete, Cancelled],
 112     Complete => [Cancelled]
 113   }
 114
 115   def self.limit_index_columns_read
 116     ["mounts"]
 117   end
 118
 119   def self.full_text_searchable_columns
 120     super - ["secret_mounts", "secret_mounts_md5", "runtime_token", "gateway_address", "output_storage_classes"]
 121   end
 122
 123   def self.searchable_columns *args
 124     super - ["secret_mounts_md5", "runtime_token", "gateway_address", "output_storage_classes"]
 125   end
 126
 127   def logged_attributes
 128     super.except('secret_mounts', 'runtime_token')
 129   end
 130
 131   def state_transitions
 132     State_transitions
 133   end
 134
 135   # Container priority is the highest "computed priority" of any
 136   # matching request. The computed priority of a container-submitted
 137   # request is the priority of the submitting container. The computed
 138   # priority of a user-submitted request is a function of
 139   # user-assigned priority and request creation time.
 140   def update_priority!
 141     update_priorities uuid
 142     reload
 143   end
 144
 145   # Create a new container (or find an existing one) to satisfy the
 146   # given container request.
 147   def self.resolve(req)
 148     if req.runtime_token.nil?
 149       runtime_user = if req.modified_by_user_uuid.nil?
 150                        current_user
 151                      else
 152                        User.find_by_uuid(req.modified_by_user_uuid)
 153                      end
 154       runtime_auth_scopes = ["all"]
 155     else
 156       auth = ApiClientAuthorization.validate(token: req.runtime_token)
 157       if auth.nil?
 158         raise ArgumentError.new "Invalid runtime token"
 159       end
 160       runtime_user = User.find_by_id(auth.user_id)
 161       runtime_auth_scopes = auth.scopes
 162     end
 163     c_attrs = act_as_user runtime_user do
 164       {
 165         command: req.command,
 166         cwd: req.cwd,
 167         environment: req.environment,
 168         output_path: req.output_path,
 169         container_image: resolve_container_image(req.container_image),
 170         mounts: resolve_mounts(req.mounts),
 171         runtime_constraints: resolve_runtime_constraints(req.runtime_constraints),
 172         scheduling_parameters: req.scheduling_parameters,
 173         secret_mounts: req.secret_mounts,
 174         runtime_token: req.runtime_token,
 175         runtime_user_uuid: runtime_user.uuid,
 176         runtime_auth_scopes: runtime_auth_scopes,
 177         output_storage_classes: req.output_storage_classes,
 178       }
 179     end
 180     act_as_system_user do
 181       if req.use_existing && (reusable = find_reusable(c_attrs))
 182         reusable
 183       else
 184         Container.create!(c_attrs)
 185       end
 186     end
 187   end
 188
 189   # Return a runtime_constraints hash that complies with requested but
 190   # is suitable for saving in a container record, i.e., has specific
 191   # values instead of ranges.
 192   #
 193   # Doing this as a step separate from other resolutions, like "git
 194   # revision range to commit hash", makes sense only when there is no
 195   # opportunity to reuse an existing container (e.g., container reuse
 196   # is not implemented yet, or we have already found that no existing
 197   # containers are suitable).
 198   def self.resolve_runtime_constraints(runtime_constraints)
 199     rc = {}
 200     runtime_constraints.each do |k, v|
 201       if v.is_a? Array
 202         rc[k] = v[0]
 203       else
 204         rc[k] = v
 205       end
 206     end
 207     if rc['keep_cache_ram'] == 0
 208       rc['keep_cache_ram'] = Rails.configuration.Containers.DefaultKeepCacheRAM
 209     end
 210     if rc['keep_cache_disk'] == 0 and rc['keep_cache_ram'] == 0
 211       rc['keep_cache_disk'] = bound_keep_cache_disk(rc['ram'])
 212     end
 213     rc
 214   end
 215
 216   # Return a mounts hash suitable for a Container, i.e., with every
 217   # readonly collection UUID resolved to a PDH.
 218   def self.resolve_mounts(mounts)
 219     c_mounts = {}
 220     mounts.each do |k, mount|
 221       mount = mount.dup
 222       c_mounts[k] = mount
 223       if mount['kind'] != 'collection'
 224         next
 225       end
 226
 227       uuid = mount.delete 'uuid'
 228
 229       if mount['portable_data_hash'].nil? and !uuid.nil?
 230         # PDH not supplied, try by UUID
 231         c = Collection.
 232           readable_by(current_user).
 233           where(uuid: uuid).
 234           select(:portable_data_hash).
 235           first
 236         if !c
 237           raise ArvadosModel::UnresolvableContainerError.new "cannot mount collection #{uuid.inspect}: not found"
 238         end
 239         mount['portable_data_hash'] = c.portable_data_hash
 240       end
 241     end
 242     return c_mounts
 243   end
 244
 245   # Return a container_image PDH suitable for a Container.
 246   def self.resolve_container_image(container_image)
 247     coll = Collection.for_latest_docker_image(container_image)
 248     if !coll
 249       raise ArvadosModel::UnresolvableContainerError.new "docker image #{container_image.inspect} not found"
 250     end
 251     coll.portable_data_hash
 252   end
 253
 254   def self.find_reusable(attrs)
 255     log_reuse_info { "starting with #{Container.all.count} container records in database" }
 256     candidates = Container.where_serialized(:command, attrs[:command], md5: true)
 257     log_reuse_info(candidates) { "after filtering on command #{attrs[:command].inspect}" }
 258
 259     candidates = candidates.where('cwd = ?', attrs[:cwd])
 260     log_reuse_info(candidates) { "after filtering on cwd #{attrs[:cwd].inspect}" }
 261
 262     candidates = candidates.where_serialized(:environment, attrs[:environment], md5: true)
 263     log_reuse_info(candidates) { "after filtering on environment #{attrs[:environment].inspect}" }
 264
 265     candidates = candidates.where('output_path = ?', attrs[:output_path])
 266     log_reuse_info(candidates) { "after filtering on output_path #{attrs[:output_path].inspect}" }
 267
 268     image = resolve_container_image(attrs[:container_image])
 269     candidates = candidates.where('container_image = ?', image)
 270     log_reuse_info(candidates) { "after filtering on container_image #{image.inspect} (resolved from #{attrs[:container_image].inspect})" }
 271
 272     candidates = candidates.where_serialized(:mounts, resolve_mounts(attrs[:mounts]), md5: true)
 273     log_reuse_info(candidates) { "after filtering on mounts #{attrs[:mounts].inspect}" }
 274
 275     secret_mounts_md5 = Digest::MD5.hexdigest(SafeJSON.dump(self.deep_sort_hash(attrs[:secret_mounts])))
 276     candidates = candidates.where('secret_mounts_md5 = ?', secret_mounts_md5)
 277     log_reuse_info(candidates) { "after filtering on secret_mounts_md5 #{secret_mounts_md5.inspect}" }
 278
 279     resolved_runtime_constraints = resolve_runtime_constraints(attrs[:runtime_constraints])
 280     # Ideally we would completely ignore Keep cache constraints when making
 281     # reuse considerations, but our database structure makes that impractical.
 282     # The best we can do is generate a search that matches on all likely values.
 283     runtime_constraint_variations = {
 284       keep_cache_disk: [
 285         # Check for constraints without keep_cache_disk
 286         # (containers that predate the constraint)
 287         nil,
 288         # Containers that use keep_cache_ram instead
 289         0,
 290         # The default value
 291         bound_keep_cache_disk(resolved_runtime_constraints['ram']),
 292         # The minimum default bound
 293         bound_keep_cache_disk(0),
 294         # The maximum default bound (presumably)
 295         bound_keep_cache_disk(1 << 60),
 296         # The requested value
 297         resolved_runtime_constraints.delete('keep_cache_disk'),
 298       ].uniq,
 299       keep_cache_ram: [
 300         # Containers that use keep_cache_disk instead
 301         0,
 302         # The default value
 303         Rails.configuration.Containers.DefaultKeepCacheRAM,
 304         # The requested value
 305         resolved_runtime_constraints.delete('keep_cache_ram'),
 306       ].uniq,
 307     }
 308     resolved_cuda = resolved_runtime_constraints['cuda']
 309     if resolved_cuda.nil? or resolved_cuda['device_count'] == 0
 310       runtime_constraint_variations[:cuda] = [
 311         # Check for constraints without cuda
 312         # (containers that predate the constraint)
 313         nil,
 314         # The default "don't need CUDA" value
 315         {
 316           'device_count' => 0,
 317           'driver_version' => '',
 318           'hardware_capability' => '',
 319         },
 320         # The requested value
 321         resolved_runtime_constraints.delete('cuda')
 322       ].uniq
 323     end
 324     reusable_runtime_constraints = hash_product(runtime_constraint_variations)
 325                                      .map { |v| resolved_runtime_constraints.merge(v) }
 326
 327     candidates = candidates.where_serialized(:runtime_constraints, reusable_runtime_constraints, md5: true, multivalue: true)
 328     log_reuse_info(candidates) { "after filtering on runtime_constraints #{attrs[:runtime_constraints].inspect}" }
 329
 330     log_reuse_info { "checking for state=Complete with readable output and log..." }
 331
 332     select_readable_pdh = Collection.
 333       readable_by(current_user).
 334       select(:portable_data_hash).
 335       to_sql
 336
 337     usable = candidates.where(state: Complete, exit_code: 0)
 338     log_reuse_info(usable) { "with state=Complete, exit_code=0" }
 339
 340     usable = usable.where("log IN (#{select_readable_pdh})")
 341     log_reuse_info(usable) { "with readable log" }
 342
 343     usable = usable.where("output IN (#{select_readable_pdh})")
 344     log_reuse_info(usable) { "with readable output" }
 345
 346     usable = usable.order('finished_at ASC').limit(1).first
 347     if usable
 348       log_reuse_info { "done, reusing container #{usable.uuid} with state=Complete" }
 349       return usable
 350     end
 351
 352     # Check for non-failing Running candidates and return the most likely to finish sooner.
 353     log_reuse_info { "checking for state=Running..." }
 354     running = candidates.where(state: Running).
 355               where("(runtime_status->'error') is null and priority > 0").
 356               order('progress desc, started_at asc').
 357               limit(1).first
 358     if running
 359       log_reuse_info { "done, reusing container #{running.uuid} with state=Running" }
 360       return running
 361     else
 362       log_reuse_info { "have no containers in Running state" }
 363     end
 364
 365     # Check for Locked or Queued ones and return the most likely to start first.
 366     locked_or_queued = candidates.
 367                        where("state IN (?)", [Locked, Queued]).
 368                        order('state asc, priority desc, created_at asc').
 369                        limit(1)
 370     if !attrs[:scheduling_parameters]['preemptible']
 371       locked_or_queued = locked_or_queued.
 372                            where("not ((scheduling_parameters::jsonb)->>'preemptible')::boolean")
 373     end
 374     chosen = locked_or_queued.first
 375     if chosen
 376       log_reuse_info { "done, reusing container #{chosen.uuid} with state=#{chosen.state}" }
 377       return chosen
 378     else
 379       log_reuse_info { "have no containers in Locked or Queued state" }
 380     end
 381
 382     log_reuse_info { "done, no reusable container found" }
 383     nil
 384   end
 385
 386   def lock
 387     self.with_lock do
 388       if self.state != Queued
 389         raise LockFailedError.new("cannot lock when #{self.state}")
 390       end
 391       self.update!(state: Locked)
 392     end
 393   end
 394
 395   def check_lock
 396     if state_was == Queued and state == Locked
 397       if self.priority <= 0
 398         raise LockFailedError.new("cannot lock when priority<=0")
 399       end
 400       self.lock_count = self.lock_count+1
 401     end
 402   end
 403
 404   def unlock
 405     self.with_lock do
 406       if self.state != Locked
 407         raise InvalidStateTransitionError.new("cannot unlock when #{self.state}")
 408       end
 409       self.update!(state: Queued)
 410     end
 411   end
 412
 413   def check_unlock
 414     if state_was == Locked and state == Queued
 415       if self.locked_by_uuid != current_api_client_authorization.uuid
 416         raise ArvadosModel::PermissionDeniedError.new("locked by a different token")
 417       end
 418       if self.lock_count >= Rails.configuration.Containers.MaxDispatchAttempts
 419         self.state = Cancelled
 420         self.runtime_status = {error: "Failed to start container.  Cancelled after exceeding 'Containers.MaxDispatchAttempts' (lock_count=#{self.lock_count})"}
 421       end
 422     end
 423   end
 424
 425   def self.readable_by(*users_list)
 426     # Load optional keyword arguments, if they exist.
 427     if users_list.last.is_a? Hash
 428       kwargs = users_list.pop
 429     else
 430       kwargs = {}
 431     end
 432     if users_list.select { |u| u.is_admin }.any?
 433       return super
 434     end
 435     Container.where(ContainerRequest.readable_by(*users_list).where("containers.uuid = container_requests.container_uuid").arel.exists)
 436   end
 437
 438   def final?
 439     [Complete, Cancelled].include?(self.state)
 440   end
 441
 442   def self.for_current_token
 443     return if !current_api_client_authorization
 444     _, _, _, container_uuid = Thread.current[:token].split('/')
 445     if container_uuid.nil?
 446       Container.where(auth_uuid: current_api_client_authorization.uuid).first
 447     else
 448       Container.where('auth_uuid=? or (uuid=? and runtime_token=?)',
 449                       current_api_client_authorization.uuid,
 450                       container_uuid,
 451                       current_api_client_authorization.token).first
 452     end
 453   end
 454
 455   protected
 456
 457   def self.bound_keep_cache_disk(value)
 458     value ||= 0
 459     min_value = 2 << 30
 460     max_value = 32 << 30
 461     if value < min_value
 462       min_value
 463     elsif value > max_value
 464       max_value
 465     else
 466       value
 467     end
 468   end
 469
 470   def self.hash_product(**kwargs)
 471     # kwargs is a hash that maps parameters to an array of values.
 472     # This function enumerates every possible hash where each key has one of
 473     # the values from its array.
 474     # The output keys are strings since that's what container hash attributes
 475     # want.
 476     # A nil value yields a hash without that key.
 477     [[:_, nil]].product(
 478       *kwargs.map { |(key, values)| [key.to_s].product(values) },
 479     ).map { |param_pairs| Hash[param_pairs].compact }
 480   end
 481
 482   def fill_field_defaults
 483     self.state ||= Queued
 484     self.environment ||= {}
 485     self.runtime_constraints ||= {}
 486     self.mounts ||= {}
 487     self.cwd ||= "."
 488     self.priority ||= 0
 489     self.scheduling_parameters ||= {}
 490   end
 491
 492   def permission_to_create
 493     current_user.andand.is_admin
 494   end
 495
 496   def permission_to_destroy
 497     current_user.andand.is_admin
 498   end
 499
 500   def ensure_owner_uuid_is_permitted
 501     # validate_change ensures owner_uuid can't be changed at all --
 502     # except during create, which requires admin privileges. Checking
 503     # permission here would be superfluous.
 504     true
 505   end
 506
 507   def set_timestamps
 508     if self.state_changed? and self.state == Running
 509       self.started_at ||= db_current_time
 510     end
 511
 512     if self.state_changed? and [Complete, Cancelled].include? self.state
 513       self.finished_at ||= db_current_time
 514     end
 515   end
 516
 517   # Check that well-known runtime status keys have desired data types
 518   def validate_runtime_status
 519     [
 520       'error', 'errorDetail', 'warning', 'warningDetail', 'activity'
 521     ].each do |k|
 522       if self.runtime_status.andand.include?(k) && !self.runtime_status[k].is_a?(String)
 523         errors.add(:runtime_status, "'#{k}' value must be a string")
 524       end
 525     end
 526   end
 527
 528   def validate_change
 529     permitted = [:state]
 530     final_attrs = [:finished_at]
 531     progress_attrs = [:progress, :runtime_status, :subrequests_cost, :cost,
 532                       :log, :output, :output_properties, :exit_code]
 533
 534     if self.new_record?
 535       permitted.push(:owner_uuid, :command, :container_image, :cwd,
 536                      :environment, :mounts, :output_path, :priority,
 537                      :runtime_constraints, :scheduling_parameters,
 538                      :secret_mounts, :runtime_token,
 539                      :runtime_user_uuid, :runtime_auth_scopes,
 540                      :output_storage_classes)
 541     end
 542
 543     case self.state
 544     when Locked
 545       permitted.push :priority, :runtime_status, :log, :lock_count
 546
 547     when Queued
 548       permitted.push :priority
 549
 550     when Running
 551       permitted.push :priority, :output_properties, :gateway_address, *progress_attrs
 552       if self.state_changed?
 553         permitted.push :started_at
 554       end
 555       if !self.interactive_session_started_was
 556         permitted.push :interactive_session_started
 557       end
 558
 559     when Complete
 560       if self.state_was == Running
 561         permitted.push *final_attrs, *progress_attrs
 562       end
 563
 564     when Cancelled
 565       case self.state_was
 566       when Running
 567         permitted.push :finished_at, *progress_attrs
 568       when Queued, Locked
 569         permitted.push :finished_at, :log, :runtime_status, :cost
 570       end
 571
 572     else
 573       # The state_transitions check will add an error message for this
 574       return false
 575     end
 576
 577     if self.state_was == Running &&
 578        !current_api_client_authorization.nil? &&
 579        (current_api_client_authorization.uuid == self.auth_uuid ||
 580         current_api_client_authorization.token == self.runtime_token)
 581       # The contained process itself can write final attrs but can't
 582       # change priority or log.
 583       permitted.push *final_attrs
 584       permitted = permitted - [:log, :priority]
 585     elsif !current_user.andand.is_admin
 586       raise PermissionDeniedError
 587     elsif self.locked_by_uuid && self.locked_by_uuid != current_api_client_authorization.andand.uuid
 588       # When locked, progress fields cannot be updated by the wrong
 589       # dispatcher, even though it has admin privileges.
 590       permitted = permitted - progress_attrs
 591     end
 592     check_update_whitelist permitted
 593   end
 594
 595   def validate_lock
 596     if [Locked, Running].include? self.state
 597       # If the Container was already locked, locked_by_uuid must not
 598       # changes. Otherwise, the current auth gets the lock.
 599       need_lock = locked_by_uuid_was || current_api_client_authorization.andand.uuid
 600     else
 601       need_lock = nil
 602     end
 603
 604     # The caller can provide a new value for locked_by_uuid, but only
 605     # if it's exactly what we expect. This allows a caller to perform
 606     # an update like {"state":"Unlocked","locked_by_uuid":null}.
 607     if self.locked_by_uuid_changed?
 608       if self.locked_by_uuid != need_lock
 609         return errors.add :locked_by_uuid, "can only change to #{need_lock}"
 610       end
 611     end
 612     self.locked_by_uuid = need_lock
 613   end
 614
 615   def validate_output
 616     # Output must exist and be readable by the current user.  This is so
 617     # that a container cannot "claim" a collection that it doesn't otherwise
 618     # have access to just by setting the output field to the collection PDH.
 619     if output_changed?
 620       c = Collection.
 621             readable_by(current_user, {include_trash: true}).
 622             where(portable_data_hash: self.output).
 623             first
 624       if !c
 625         errors.add :output, "collection must exist and be readable by current user."
 626       end
 627     end
 628   end
 629
 630   def update_cr_logs
 631     # If self.final?, this update is superfluous: the final log/output
 632     # update will be done when handle_completed calls finalize! on
 633     # each requesting CR.
 634     return if self.final? || !saved_change_to_log?
 635     leave_modified_by_user_alone do
 636       ContainerRequest.where(container_uuid: self.uuid, state: ContainerRequest::Committed).each do |cr|
 637         cr.update_collections(container: self, collections: ['log'])
 638         cr.save!
 639       end
 640     end
 641   end
 642
 643   def assign_auth
 644     if self.auth_uuid_changed?
 645          return errors.add :auth_uuid, 'is readonly'
 646     end
 647     if not [Locked, Running].include? self.state
 648       # Don't need one. If auth already exists, expire it.
 649       #
 650       # We use db_transaction_time here (not db_current_time) to
 651       # ensure the token doesn't validate later in the same
 652       # transaction (e.g., in a test case) by satisfying expires_at >
 653       # transaction timestamp.
 654       self.auth.andand.update(expires_at: db_transaction_time)
 655       self.auth = nil
 656       return
 657     elsif self.auth
 658       # already have one
 659       return
 660     end
 661     if self.runtime_token.nil?
 662       if self.runtime_user_uuid.nil?
 663         # legacy behavior, we don't have a runtime_user_uuid so get
 664         # the user from the highest priority container request, needed
 665         # when performing an upgrade and there are queued containers,
 666         # and some tests.
 667         cr = ContainerRequest.
 668                where('container_uuid=? and priority>0', self.uuid).
 669                order('priority desc').
 670                first
 671         if !cr
 672           return errors.add :auth_uuid, "cannot be assigned because priority <= 0"
 673         end
 674         self.runtime_user_uuid = cr.modified_by_user_uuid
 675         self.runtime_auth_scopes = ["all"]
 676       end
 677
 678       # Generate a new token. This runs with admin credentials as it's done by a
 679       # dispatcher user, so expires_at isn't enforced by API.MaxTokenLifetime.
 680       self.auth = ApiClientAuthorization.
 681                     create!(user_id: User.find_by_uuid(self.runtime_user_uuid).id,
 682                             api_client_id: 0,
 683                             scopes: self.runtime_auth_scopes)
 684     end
 685   end
 686
 687   def sort_serialized_attrs
 688     if self.environment_changed?
 689       self.environment = self.class.deep_sort_hash(self.environment)
 690     end
 691     if self.mounts_changed?
 692       self.mounts = self.class.deep_sort_hash(self.mounts)
 693     end
 694     if self.runtime_constraints_changed?
 695       self.runtime_constraints = self.class.deep_sort_hash(self.runtime_constraints)
 696     end
 697     if self.scheduling_parameters_changed?
 698       self.scheduling_parameters = self.class.deep_sort_hash(self.scheduling_parameters)
 699     end
 700     if self.runtime_auth_scopes_changed?
 701       self.runtime_auth_scopes = self.runtime_auth_scopes.sort
 702     end
 703   end
 704
 705   def update_secret_mounts_md5
 706     if self.secret_mounts_changed?
 707       self.secret_mounts_md5 = Digest::MD5.hexdigest(
 708         SafeJSON.dump(self.class.deep_sort_hash(self.secret_mounts)))
 709     end
 710   end
 711
 712   def scrub_secrets
 713     # this runs after update_secret_mounts_md5, so the
 714     # secret_mounts_md5 will still reflect the secrets that are being
 715     # scrubbed here.
 716     if self.state_changed? && self.final?
 717       self.secret_mounts = {}
 718       self.runtime_token = nil
 719     end
 720   end
 721
 722   def clear_runtime_status_when_queued
 723     # Avoid leaking status messages between different dispatch attempts
 724     if self.state_was == Locked && self.state == Queued
 725       self.runtime_status = {}
 726     end
 727   end
 728
 729   def handle_completed
 730     # This container is finished so finalize any associated container requests
 731     # that are associated with this container.
 732     if saved_change_to_state? and self.final?
 733       # These get wiped out by with_lock (which reloads the record),
 734       # so record them now in case we need to schedule a retry.
 735       prev_secret_mounts = secret_mounts_before_last_save
 736       prev_runtime_token = runtime_token_before_last_save
 737
 738       # Need to take a lock on the container to ensure that any
 739       # concurrent container requests that might try to reuse this
 740       # container will block until the container completion
 741       # transaction finishes.  This ensure that concurrent container
 742       # requests that try to reuse this container are finalized (on
 743       # Complete) or don't reuse it (on Cancelled).
 744       self.with_lock do
 745         act_as_system_user do
 746           if self.state == Cancelled
 747             # Cancelled means the container didn't run to completion.
 748             # This happens either because it was cancelled by the user
 749             # or because there was an infrastructure failure.  We want
 750             # to retry infrastructure failures automatically.
 751             #
 752             # Seach for live container requests to determine if we
 753             # should retry the container.
 754             retryable_requests = ContainerRequest.
 755                                    joins('left outer join containers as requesting_container on container_requests.requesting_container_uuid = requesting_container.uuid').
 756                                    where("container_requests.container_uuid = ? and "+
 757                                          "container_requests.priority > 0 and "+
 758                                          "container_requests.owner_uuid not in (select group_uuid from trashed_groups) and "+
 759                                          "(requesting_container.priority is null or (requesting_container.state = 'Running' and requesting_container.priority > 0)) and "+
 760                                          "container_requests.state = 'Committed' and "+
 761                                          "container_requests.container_count < container_requests.container_count_max", uuid).
 762                                    order('container_requests.uuid asc')
 763           else
 764             retryable_requests = []
 765           end
 766
 767           if retryable_requests.any?
 768             scheduling_parameters = {
 769               # partitions: empty if any are empty, else the union of all parameters
 770               "partitions": retryable_requests
 771                               .map { |req| req.scheduling_parameters["partitions"] || [] }
 772                               .reduce { |cur, new| (cur.empty? or new.empty?) ? [] : (cur | new) },
 773
 774               # preemptible: true if all are true, else false
 775               "preemptible": retryable_requests
 776                                .map { |req| req.scheduling_parameters["preemptible"] }
 777                                .all?,
 778
 779               # supervisor: true if all any true, else false
 780               "supervisor": retryable_requests
 781                                .map { |req| req.scheduling_parameters["supervisor"] }
 782                                .any?,
 783
 784               # max_run_time: 0 if any are 0 (unlimited), else the maximum
 785               "max_run_time": retryable_requests
 786                                 .map { |req| req.scheduling_parameters["max_run_time"] || 0 }
 787                                 .reduce do |cur, new|
 788                 if cur == 0 or new == 0
 789                   0
 790                 elsif new > cur
 791                   new
 792                 else
 793                   cur
 794                 end
 795               end,
 796             }
 797
 798             c_attrs = {
 799               command: self.command,
 800               cwd: self.cwd,
 801               environment: self.environment,
 802               output_path: self.output_path,
 803               container_image: self.container_image,
 804               mounts: self.mounts,
 805               runtime_constraints: self.runtime_constraints,
 806               scheduling_parameters: scheduling_parameters,
 807               secret_mounts: prev_secret_mounts,
 808               runtime_token: prev_runtime_token,
 809               runtime_user_uuid: self.runtime_user_uuid,
 810               runtime_auth_scopes: self.runtime_auth_scopes
 811             }
 812             c = Container.create! c_attrs
 813             retryable_requests.each do |cr|
 814               cr.with_lock do
 815                 leave_modified_by_user_alone do
 816                   # Use row locking because this increments container_count
 817                   cr.cumulative_cost += self.cost + self.subrequests_cost
 818                   cr.container_uuid = c.uuid
 819                   cr.save!
 820                 end
 821               end
 822             end
 823           end
 824
 825           # Notify container requests associated with this container
 826           ContainerRequest.where(container_uuid: uuid,
 827                                  state: ContainerRequest::Committed).each do |cr|
 828             leave_modified_by_user_alone do
 829               cr.finalize!
 830             end
 831           end
 832
 833           # Cancel outstanding container requests made by this container.
 834           ContainerRequest.
 835             where(requesting_container_uuid: uuid,
 836                   state: ContainerRequest::Committed).
 837             in_batches(of: 15).each_record do |cr|
 838             leave_modified_by_user_alone do
 839               cr.set_priority_zero
 840               container_state = Container.where(uuid: cr.container_uuid).pluck(:state).first
 841               if container_state == Container::Queued || container_state == Container::Locked
 842                 # If the child container hasn't started yet, finalize the
 843                 # child CR now instead of leaving it "on hold", i.e.,
 844                 # Queued with priority 0.  (OTOH, if the child is already
 845                 # running, leave it alone so it can get cancelled the
 846                 # usual way, get a copy of the log collection, etc.)
 847                 cr.update!(state: ContainerRequest::Final)
 848               end
 849             end
 850           end
 851         end
 852       end
 853     end
 854   end
 855 end