1 # Copyright (C) The Arvados Authors. All rights reserved.
3 # SPDX-License-Identifier: AGPL-3.0
5 require 'log_reuse_info'
6 require 'whitelist_update'
8 require 'update_priority'
10 class Container < ArvadosModel
11 include ArvadosModelUpdates
14 include CommonApiTemplate
15 include WhitelistUpdate
16 extend CurrentApiClient
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: {}
28 serialize :environment, Hash
29 serialize :mounts, Hash
30 serialize :runtime_constraints, Hash
31 serialize :command, Array
32 serialize :scheduling_parameters, Hash
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 after_save :propagate_priority
54 after_commit { UpdatePriority.run_update_thread }
56 has_many :container_requests, :foreign_key => :container_uuid, :class_name => 'ContainerRequest', :primary_key => :uuid
57 belongs_to :auth, :class_name => 'ApiClientAuthorization', :foreign_key => :auth_uuid, :primary_key => :uuid
59 api_accessible :user, extend: :common do |t|
61 t.add :container_image
73 t.add :runtime_constraints
78 t.add :scheduling_parameters
79 t.add :runtime_user_uuid
80 t.add :runtime_auth_scopes
82 t.add :gateway_address
83 t.add :interactive_session_started
84 t.add :output_storage_classes
85 t.add :output_properties
87 t.add :subrequests_cost
90 # Supported states for a container
95 (Running = 'Running'),
96 (Complete = 'Complete'),
97 (Cancelled = 'Cancelled')
100 State_transitions = {
102 Queued => [Locked, Cancelled],
103 Locked => [Queued, Running, Cancelled],
104 Running => [Complete, Cancelled],
105 Complete => [Cancelled]
108 def self.limit_index_columns_read
112 def self.full_text_searchable_columns
113 super - ["secret_mounts", "secret_mounts_md5", "runtime_token", "gateway_address", "output_storage_classes"]
116 def self.searchable_columns *args
117 super - ["secret_mounts_md5", "runtime_token", "gateway_address", "output_storage_classes"]
120 def logged_attributes
121 super.except('secret_mounts', 'runtime_token')
124 def state_transitions
128 # Container priority is the highest "computed priority" of any
129 # matching request. The computed priority of a container-submitted
130 # request is the priority of the submitting container. The computed
131 # priority of a user-submitted request is a function of
132 # user-assigned priority and request creation time.
134 return if ![Queued, Locked, Running].include?(state)
135 p = ContainerRequest.
136 where('container_uuid=? and priority>0', uuid).
137 includes(:requesting_container).
140 if cr.requesting_container
141 cr.requesting_container.priority
143 (cr.priority << 50) - (cr.created_at.to_time.to_f * 1000).to_i
146 update_attributes!(priority: p)
149 def propagate_priority
150 return true unless saved_change_to_priority?
151 act_as_system_user do
152 # Update the priority of child container requests to match new
153 # priority of the parent container (ignoring requests with no
154 # container assigned, because their priority doesn't matter).
156 where(requesting_container_uuid: self.uuid,
157 state: ContainerRequest::Committed).
158 where('container_uuid is not null').
159 includes(:container).
161 map(&:update_priority!)
165 # Create a new container (or find an existing one) to satisfy the
166 # given container request.
167 def self.resolve(req)
168 if req.runtime_token.nil?
169 runtime_user = if req.modified_by_user_uuid.nil?
172 User.find_by_uuid(req.modified_by_user_uuid)
174 runtime_auth_scopes = ["all"]
176 auth = ApiClientAuthorization.validate(token: req.runtime_token)
178 raise ArgumentError.new "Invalid runtime token"
180 runtime_user = User.find_by_id(auth.user_id)
181 runtime_auth_scopes = auth.scopes
183 c_attrs = act_as_user runtime_user do
185 command: req.command,
187 environment: req.environment,
188 output_path: req.output_path,
189 container_image: resolve_container_image(req.container_image),
190 mounts: resolve_mounts(req.mounts),
191 runtime_constraints: resolve_runtime_constraints(req.runtime_constraints),
192 scheduling_parameters: req.scheduling_parameters,
193 secret_mounts: req.secret_mounts,
194 runtime_token: req.runtime_token,
195 runtime_user_uuid: runtime_user.uuid,
196 runtime_auth_scopes: runtime_auth_scopes,
197 output_storage_classes: req.output_storage_classes,
200 act_as_system_user do
201 if req.use_existing && (reusable = find_reusable(c_attrs))
204 Container.create!(c_attrs)
209 # Return a runtime_constraints hash that complies with requested but
210 # is suitable for saving in a container record, i.e., has specific
211 # values instead of ranges.
213 # Doing this as a step separate from other resolutions, like "git
214 # revision range to commit hash", makes sense only when there is no
215 # opportunity to reuse an existing container (e.g., container reuse
216 # is not implemented yet, or we have already found that no existing
217 # containers are suitable).
218 def self.resolve_runtime_constraints(runtime_constraints)
220 runtime_constraints.each do |k, v|
227 if rc['keep_cache_ram'] == 0
228 rc['keep_cache_ram'] = Rails.configuration.Containers.DefaultKeepCacheRAM
233 # Return a mounts hash suitable for a Container, i.e., with every
234 # readonly collection UUID resolved to a PDH.
235 def self.resolve_mounts(mounts)
237 mounts.each do |k, mount|
240 if mount['kind'] != 'collection'
244 uuid = mount.delete 'uuid'
246 if mount['portable_data_hash'].nil? and !uuid.nil?
247 # PDH not supplied, try by UUID
249 readable_by(current_user).
251 select(:portable_data_hash).
254 raise ArvadosModel::UnresolvableContainerError.new "cannot mount collection #{uuid.inspect}: not found"
256 mount['portable_data_hash'] = c.portable_data_hash
262 # Return a container_image PDH suitable for a Container.
263 def self.resolve_container_image(container_image)
264 coll = Collection.for_latest_docker_image(container_image)
266 raise ArvadosModel::UnresolvableContainerError.new "docker image #{container_image.inspect} not found"
268 coll.portable_data_hash
271 def self.find_reusable(attrs)
272 log_reuse_info { "starting with #{Container.all.count} container records in database" }
273 candidates = Container.where_serialized(:command, attrs[:command], md5: true)
274 log_reuse_info(candidates) { "after filtering on command #{attrs[:command].inspect}" }
276 candidates = candidates.where('cwd = ?', attrs[:cwd])
277 log_reuse_info(candidates) { "after filtering on cwd #{attrs[:cwd].inspect}" }
279 candidates = candidates.where_serialized(:environment, attrs[:environment], md5: true)
280 log_reuse_info(candidates) { "after filtering on environment #{attrs[:environment].inspect}" }
282 candidates = candidates.where('output_path = ?', attrs[:output_path])
283 log_reuse_info(candidates) { "after filtering on output_path #{attrs[:output_path].inspect}" }
285 image = resolve_container_image(attrs[:container_image])
286 candidates = candidates.where('container_image = ?', image)
287 log_reuse_info(candidates) { "after filtering on container_image #{image.inspect} (resolved from #{attrs[:container_image].inspect})" }
289 candidates = candidates.where_serialized(:mounts, resolve_mounts(attrs[:mounts]), md5: true)
290 log_reuse_info(candidates) { "after filtering on mounts #{attrs[:mounts].inspect}" }
292 secret_mounts_md5 = Digest::MD5.hexdigest(SafeJSON.dump(self.deep_sort_hash(attrs[:secret_mounts])))
293 candidates = candidates.where('secret_mounts_md5 = ?', secret_mounts_md5)
294 log_reuse_info(candidates) { "after filtering on secret_mounts_md5 #{secret_mounts_md5.inspect}" }
296 if attrs[:runtime_constraints]['cuda'].nil?
297 attrs[:runtime_constraints]['cuda'] = {
299 'driver_version' => '',
300 'hardware_capability' => '',
303 resolved_runtime_constraints = [resolve_runtime_constraints(attrs[:runtime_constraints])]
304 if resolved_runtime_constraints[0]['cuda']['device_count'] == 0
305 # If no CUDA requested, extend search to include older container
306 # records that don't have a 'cuda' section in runtime_constraints
307 resolved_runtime_constraints << resolved_runtime_constraints[0].except('cuda')
310 candidates = candidates.where_serialized(:runtime_constraints, resolved_runtime_constraints, md5: true, multivalue: true)
311 log_reuse_info(candidates) { "after filtering on runtime_constraints #{attrs[:runtime_constraints].inspect}" }
313 log_reuse_info { "checking for state=Complete with readable output and log..." }
315 select_readable_pdh = Collection.
316 readable_by(current_user).
317 select(:portable_data_hash).
320 usable = candidates.where(state: Complete, exit_code: 0)
321 log_reuse_info(usable) { "with state=Complete, exit_code=0" }
323 usable = usable.where("log IN (#{select_readable_pdh})")
324 log_reuse_info(usable) { "with readable log" }
326 usable = usable.where("output IN (#{select_readable_pdh})")
327 log_reuse_info(usable) { "with readable output" }
329 usable = usable.order('finished_at ASC').limit(1).first
331 log_reuse_info { "done, reusing container #{usable.uuid} with state=Complete" }
335 # Check for non-failing Running candidates and return the most likely to finish sooner.
336 log_reuse_info { "checking for state=Running..." }
337 running = candidates.where(state: Running).
338 where("(runtime_status->'error') is null").
339 order('progress desc, started_at asc').
342 log_reuse_info { "done, reusing container #{running.uuid} with state=Running" }
345 log_reuse_info { "have no containers in Running state" }
348 # Check for Locked or Queued ones and return the most likely to start first.
349 locked_or_queued = candidates.
350 where("state IN (?)", [Locked, Queued]).
351 order('state asc, priority desc, created_at asc').
354 log_reuse_info { "done, reusing container #{locked_or_queued.uuid} with state=#{locked_or_queued.state}" }
355 return locked_or_queued
357 log_reuse_info { "have no containers in Locked or Queued state" }
360 log_reuse_info { "done, no reusable container found" }
366 if self.state != Queued
367 raise LockFailedError.new("cannot lock when #{self.state}")
369 self.update_attributes!(state: Locked)
374 if state_was == Queued and state == Locked
375 if self.priority <= 0
376 raise LockFailedError.new("cannot lock when priority<=0")
378 self.lock_count = self.lock_count+1
384 if self.state != Locked
385 raise InvalidStateTransitionError.new("cannot unlock when #{self.state}")
387 self.update_attributes!(state: Queued)
392 if state_was == Locked and state == Queued
393 if self.locked_by_uuid != current_api_client_authorization.uuid
394 raise ArvadosModel::PermissionDeniedError.new("locked by a different token")
396 if self.lock_count >= Rails.configuration.Containers.MaxDispatchAttempts
397 self.state = Cancelled
398 self.runtime_status = {error: "Failed to start container. Cancelled after exceeding 'Containers.MaxDispatchAttempts' (lock_count=#{self.lock_count})"}
403 def self.readable_by(*users_list)
404 # Load optional keyword arguments, if they exist.
405 if users_list.last.is_a? Hash
406 kwargs = users_list.pop
410 if users_list.select { |u| u.is_admin }.any?
413 Container.where(ContainerRequest.readable_by(*users_list).where("containers.uuid = container_requests.container_uuid").arel.exists)
417 [Complete, Cancelled].include?(self.state)
420 def self.for_current_token
421 return if !current_api_client_authorization
422 _, _, _, container_uuid = Thread.current[:token].split('/')
423 if container_uuid.nil?
424 Container.where(auth_uuid: current_api_client_authorization.uuid).first
426 Container.where('auth_uuid=? or (uuid=? and runtime_token=?)',
427 current_api_client_authorization.uuid,
429 current_api_client_authorization.token).first
435 def fill_field_defaults
436 self.state ||= Queued
437 self.environment ||= {}
438 self.runtime_constraints ||= {}
442 self.scheduling_parameters ||= {}
445 def permission_to_create
446 current_user.andand.is_admin
449 def permission_to_destroy
450 current_user.andand.is_admin
453 def ensure_owner_uuid_is_permitted
454 # validate_change ensures owner_uuid can't be changed at all --
455 # except during create, which requires admin privileges. Checking
456 # permission here would be superfluous.
461 if self.state_changed? and self.state == Running
462 self.started_at ||= db_current_time
465 if self.state_changed? and [Complete, Cancelled].include? self.state
466 self.finished_at ||= db_current_time
470 # Check that well-known runtime status keys have desired data types
471 def validate_runtime_status
473 'error', 'errorDetail', 'warning', 'warningDetail', 'activity'
475 if self.runtime_status.andand.include?(k) && !self.runtime_status[k].is_a?(String)
476 errors.add(:runtime_status, "'#{k}' value must be a string")
483 final_attrs = [:finished_at]
484 progress_attrs = [:progress, :runtime_status, :subrequests_cost, :cost,
485 :log, :output, :output_properties, :exit_code]
488 permitted.push(:owner_uuid, :command, :container_image, :cwd,
489 :environment, :mounts, :output_path, :priority,
490 :runtime_constraints, :scheduling_parameters,
491 :secret_mounts, :runtime_token,
492 :runtime_user_uuid, :runtime_auth_scopes,
493 :output_storage_classes)
498 permitted.push :priority, :runtime_status, :log, :lock_count
501 permitted.push :priority
504 permitted.push :priority, :output_properties, :gateway_address, *progress_attrs
505 if self.state_changed?
506 permitted.push :started_at
508 if !self.interactive_session_started_was
509 permitted.push :interactive_session_started
513 if self.state_was == Running
514 permitted.push *final_attrs, *progress_attrs
520 permitted.push :finished_at, *progress_attrs
522 permitted.push :finished_at, :log, :runtime_status, :cost
526 # The state_transitions check will add an error message for this
530 if self.state_was == Running &&
531 !current_api_client_authorization.nil? &&
532 (current_api_client_authorization.uuid == self.auth_uuid ||
533 current_api_client_authorization.token == self.runtime_token)
534 # The contained process itself can write final attrs but can't
535 # change priority or log.
536 permitted.push *final_attrs
537 permitted = permitted - [:log, :priority]
538 elsif !current_user.andand.is_admin
539 raise PermissionDeniedError
540 elsif self.locked_by_uuid && self.locked_by_uuid != current_api_client_authorization.andand.uuid
541 # When locked, progress fields cannot be updated by the wrong
542 # dispatcher, even though it has admin privileges.
543 permitted = permitted - progress_attrs
545 check_update_whitelist permitted
549 if [Locked, Running].include? self.state
550 # If the Container was already locked, locked_by_uuid must not
551 # changes. Otherwise, the current auth gets the lock.
552 need_lock = locked_by_uuid_was || current_api_client_authorization.andand.uuid
557 # The caller can provide a new value for locked_by_uuid, but only
558 # if it's exactly what we expect. This allows a caller to perform
559 # an update like {"state":"Unlocked","locked_by_uuid":null}.
560 if self.locked_by_uuid_changed?
561 if self.locked_by_uuid != need_lock
562 return errors.add :locked_by_uuid, "can only change to #{need_lock}"
565 self.locked_by_uuid = need_lock
569 # Output must exist and be readable by the current user. This is so
570 # that a container cannot "claim" a collection that it doesn't otherwise
571 # have access to just by setting the output field to the collection PDH.
574 readable_by(current_user, {include_trash: true}).
575 where(portable_data_hash: self.output).
578 errors.add :output, "collection must exist and be readable by current user."
584 # If self.final?, this update is superfluous: the final log/output
585 # update will be done when handle_completed calls finalize! on
586 # each requesting CR.
587 return if self.final? || !saved_change_to_log?
588 leave_modified_by_user_alone do
589 ContainerRequest.where(container_uuid: self.uuid).each do |cr|
590 cr.update_collections(container: self, collections: ['log'])
597 if self.auth_uuid_changed?
598 return errors.add :auth_uuid, 'is readonly'
600 if not [Locked, Running].include? self.state
601 # Don't need one. If auth already exists, expire it.
603 # We use db_transaction_time here (not db_current_time) to
604 # ensure the token doesn't validate later in the same
605 # transaction (e.g., in a test case) by satisfying expires_at >
606 # transaction timestamp.
607 self.auth.andand.update_attributes(expires_at: db_transaction_time)
614 if self.runtime_token.nil?
615 if self.runtime_user_uuid.nil?
616 # legacy behavior, we don't have a runtime_user_uuid so get
617 # the user from the highest priority container request, needed
618 # when performing an upgrade and there are queued containers,
620 cr = ContainerRequest.
621 where('container_uuid=? and priority>0', self.uuid).
622 order('priority desc').
625 return errors.add :auth_uuid, "cannot be assigned because priority <= 0"
627 self.runtime_user_uuid = cr.modified_by_user_uuid
628 self.runtime_auth_scopes = ["all"]
631 # Generate a new token. This runs with admin credentials as it's done by a
632 # dispatcher user, so expires_at isn't enforced by API.MaxTokenLifetime.
633 self.auth = ApiClientAuthorization.
634 create!(user_id: User.find_by_uuid(self.runtime_user_uuid).id,
636 scopes: self.runtime_auth_scopes)
640 def sort_serialized_attrs
641 if self.environment_changed?
642 self.environment = self.class.deep_sort_hash(self.environment)
644 if self.mounts_changed?
645 self.mounts = self.class.deep_sort_hash(self.mounts)
647 if self.runtime_constraints_changed?
648 self.runtime_constraints = self.class.deep_sort_hash(self.runtime_constraints)
650 if self.scheduling_parameters_changed?
651 self.scheduling_parameters = self.class.deep_sort_hash(self.scheduling_parameters)
653 if self.runtime_auth_scopes_changed?
654 self.runtime_auth_scopes = self.runtime_auth_scopes.sort
658 def update_secret_mounts_md5
659 if self.secret_mounts_changed?
660 self.secret_mounts_md5 = Digest::MD5.hexdigest(
661 SafeJSON.dump(self.class.deep_sort_hash(self.secret_mounts)))
666 # this runs after update_secret_mounts_md5, so the
667 # secret_mounts_md5 will still reflect the secrets that are being
669 if self.state_changed? && self.final?
670 self.secret_mounts = {}
671 self.runtime_token = nil
675 def clear_runtime_status_when_queued
676 # Avoid leaking status messages between different dispatch attempts
677 if self.state_was == Locked && self.state == Queued
678 self.runtime_status = {}
683 # This container is finished so finalize any associated container requests
684 # that are associated with this container.
685 if saved_change_to_state? and self.final?
686 # These get wiped out by with_lock (which reloads the record),
687 # so record them now in case we need to schedule a retry.
688 prev_secret_mounts = secret_mounts_before_last_save
689 prev_runtime_token = runtime_token_before_last_save
691 # Need to take a lock on the container to ensure that any
692 # concurrent container requests that might try to reuse this
693 # container will block until the container completion
694 # transaction finishes. This ensure that concurrent container
695 # requests that try to reuse this container are finalized (on
696 # Complete) or don't reuse it (on Cancelled).
698 act_as_system_user do
699 if self.state == Cancelled
700 retryable_requests = ContainerRequest.where("container_uuid = ? and priority > 0 and state = 'Committed' and container_count < container_count_max", uuid)
702 retryable_requests = []
705 if retryable_requests.any?
707 command: self.command,
709 environment: self.environment,
710 output_path: self.output_path,
711 container_image: self.container_image,
713 runtime_constraints: self.runtime_constraints,
714 scheduling_parameters: self.scheduling_parameters,
715 secret_mounts: prev_secret_mounts,
716 runtime_token: prev_runtime_token,
717 runtime_user_uuid: self.runtime_user_uuid,
718 runtime_auth_scopes: self.runtime_auth_scopes
720 c = Container.create! c_attrs
721 retryable_requests.each do |cr|
723 leave_modified_by_user_alone do
724 # Use row locking because this increments container_count
725 cr.cumulative_cost += self.cost + self.subrequests_cost
726 cr.container_uuid = c.uuid
733 # Notify container requests associated with this container
734 ContainerRequest.where(container_uuid: uuid,
735 state: ContainerRequest::Committed).each do |cr|
736 leave_modified_by_user_alone do
741 # Cancel outstanding container requests made by this container.
743 includes(:container).
744 where(requesting_container_uuid: uuid,
745 state: ContainerRequest::Committed).each do |cr|
746 leave_modified_by_user_alone do
747 cr.update_attributes!(priority: 0)
749 if cr.container.state == Container::Queued || cr.container.state == Container::Locked
750 # If the child container hasn't started yet, finalize the
751 # child CR now instead of leaving it "on hold", i.e.,
752 # Queued with priority 0. (OTOH, if the child is already
753 # running, leave it alone so it can get cancelled the
754 # usual way, get a copy of the log collection, etc.)
755 cr.update_attributes!(state: ContainerRequest::Final)