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_priorities'
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
54 has_many :container_requests,
55 class_name: 'ContainerRequest',
56 foreign_key: 'container_uuid',
59 class_name: 'ApiClientAuthorization',
60 foreign_key: 'auth_uuid',
64 api_accessible :user, extend: :common do |t|
66 t.add :container_image
78 t.add :runtime_constraints
83 t.add :scheduling_parameters
84 t.add :runtime_user_uuid
85 t.add :runtime_auth_scopes
87 t.add :gateway_address
88 t.add :interactive_session_started
89 t.add :output_storage_classes
90 t.add :output_properties
92 t.add :subrequests_cost
95 # Supported states for a container
100 (Running = 'Running'),
101 (Complete = 'Complete'),
102 (Cancelled = 'Cancelled')
105 State_transitions = {
107 Queued => [Locked, Cancelled],
108 Locked => [Queued, Running, Cancelled],
109 Running => [Complete, Cancelled],
110 Complete => [Cancelled]
113 def self.limit_index_columns_read
117 def self.full_text_searchable_columns
118 super - ["secret_mounts", "secret_mounts_md5", "runtime_token", "gateway_address", "output_storage_classes"]
121 def self.searchable_columns *args
122 super - ["secret_mounts_md5", "runtime_token", "gateway_address", "output_storage_classes"]
125 def logged_attributes
126 super.except('secret_mounts', 'runtime_token')
129 def state_transitions
133 # Container priority is the highest "computed priority" of any
134 # matching request. The computed priority of a container-submitted
135 # request is the priority of the submitting container. The computed
136 # priority of a user-submitted request is a function of
137 # user-assigned priority and request creation time.
139 update_priorities uuid
143 # Create a new container (or find an existing one) to satisfy the
144 # given container request.
145 def self.resolve(req)
146 if req.runtime_token.nil?
147 runtime_user = if req.modified_by_user_uuid.nil?
150 User.find_by_uuid(req.modified_by_user_uuid)
152 runtime_auth_scopes = ["all"]
154 auth = ApiClientAuthorization.validate(token: req.runtime_token)
156 raise ArgumentError.new "Invalid runtime token"
158 runtime_user = User.find_by_id(auth.user_id)
159 runtime_auth_scopes = auth.scopes
161 c_attrs = act_as_user runtime_user do
163 command: req.command,
165 environment: req.environment,
166 output_path: req.output_path,
167 container_image: resolve_container_image(req.container_image),
168 mounts: resolve_mounts(req.mounts),
169 runtime_constraints: resolve_runtime_constraints(req.runtime_constraints),
170 scheduling_parameters: req.scheduling_parameters,
171 secret_mounts: req.secret_mounts,
172 runtime_token: req.runtime_token,
173 runtime_user_uuid: runtime_user.uuid,
174 runtime_auth_scopes: runtime_auth_scopes,
175 output_storage_classes: req.output_storage_classes,
178 act_as_system_user do
179 if req.use_existing && (reusable = find_reusable(c_attrs))
182 Container.create!(c_attrs)
187 # Return a runtime_constraints hash that complies with requested but
188 # is suitable for saving in a container record, i.e., has specific
189 # values instead of ranges.
191 # Doing this as a step separate from other resolutions, like "git
192 # revision range to commit hash", makes sense only when there is no
193 # opportunity to reuse an existing container (e.g., container reuse
194 # is not implemented yet, or we have already found that no existing
195 # containers are suitable).
196 def self.resolve_runtime_constraints(runtime_constraints)
198 runtime_constraints.each do |k, v|
205 if rc['keep_cache_ram'] == 0
206 rc['keep_cache_ram'] = Rails.configuration.Containers.DefaultKeepCacheRAM
208 if rc['keep_cache_disk'] == 0 and rc['keep_cache_ram'] == 0
209 rc['keep_cache_disk'] = bound_keep_cache_disk(rc['ram'])
214 # Return a mounts hash suitable for a Container, i.e., with every
215 # readonly collection UUID resolved to a PDH.
216 def self.resolve_mounts(mounts)
218 mounts.each do |k, mount|
221 if mount['kind'] != 'collection'
225 uuid = mount.delete 'uuid'
227 if mount['portable_data_hash'].nil? and !uuid.nil?
228 # PDH not supplied, try by UUID
230 readable_by(current_user).
232 select(:portable_data_hash).
235 raise ArvadosModel::UnresolvableContainerError.new "cannot mount collection #{uuid.inspect}: not found"
237 mount['portable_data_hash'] = c.portable_data_hash
243 # Return a container_image PDH suitable for a Container.
244 def self.resolve_container_image(container_image)
245 coll = Collection.for_latest_docker_image(container_image)
247 raise ArvadosModel::UnresolvableContainerError.new "docker image #{container_image.inspect} not found"
249 coll.portable_data_hash
252 def self.find_reusable(attrs)
253 log_reuse_info { "starting with #{Container.all.count} container records in database" }
254 candidates = Container.where_serialized(:command, attrs[:command], md5: true)
255 log_reuse_info(candidates) { "after filtering on command #{attrs[:command].inspect}" }
257 candidates = candidates.where('cwd = ?', attrs[:cwd])
258 log_reuse_info(candidates) { "after filtering on cwd #{attrs[:cwd].inspect}" }
260 candidates = candidates.where_serialized(:environment, attrs[:environment], md5: true)
261 log_reuse_info(candidates) { "after filtering on environment #{attrs[:environment].inspect}" }
263 candidates = candidates.where('output_path = ?', attrs[:output_path])
264 log_reuse_info(candidates) { "after filtering on output_path #{attrs[:output_path].inspect}" }
266 image = resolve_container_image(attrs[:container_image])
267 candidates = candidates.where('container_image = ?', image)
268 log_reuse_info(candidates) { "after filtering on container_image #{image.inspect} (resolved from #{attrs[:container_image].inspect})" }
270 candidates = candidates.where_serialized(:mounts, resolve_mounts(attrs[:mounts]), md5: true)
271 log_reuse_info(candidates) { "after filtering on mounts #{attrs[:mounts].inspect}" }
273 secret_mounts_md5 = Digest::MD5.hexdigest(SafeJSON.dump(self.deep_sort_hash(attrs[:secret_mounts])))
274 candidates = candidates.where('secret_mounts_md5 = ?', secret_mounts_md5)
275 log_reuse_info(candidates) { "after filtering on secret_mounts_md5 #{secret_mounts_md5.inspect}" }
277 resolved_runtime_constraints = resolve_runtime_constraints(attrs[:runtime_constraints])
278 # Ideally we would completely ignore Keep cache constraints when making
279 # reuse considerations, but our database structure makes that impractical.
280 # The best we can do is generate a search that matches on all likely values.
281 runtime_constraint_variations = {
283 # Check for constraints without keep_cache_disk
284 # (containers that predate the constraint)
286 # Containers that use keep_cache_ram instead
289 bound_keep_cache_disk(resolved_runtime_constraints['ram']),
290 # The minimum default bound
291 bound_keep_cache_disk(0),
292 # The maximum default bound (presumably)
293 bound_keep_cache_disk(1 << 60),
294 # The requested value
295 resolved_runtime_constraints.delete('keep_cache_disk'),
298 # Containers that use keep_cache_disk instead
301 Rails.configuration.Containers.DefaultKeepCacheRAM,
302 # The requested value
303 resolved_runtime_constraints.delete('keep_cache_ram'),
306 resolved_cuda = resolved_runtime_constraints['cuda']
307 if resolved_cuda.nil? or resolved_cuda['device_count'] == 0
308 runtime_constraint_variations[:cuda] = [
309 # Check for constraints without cuda
310 # (containers that predate the constraint)
312 # The default "don't need CUDA" value
315 'driver_version' => '',
316 'hardware_capability' => '',
318 # The requested value
319 resolved_runtime_constraints.delete('cuda')
322 reusable_runtime_constraints = hash_product(**runtime_constraint_variations)
323 .map { |v| resolved_runtime_constraints.merge(v) }
325 candidates = candidates.where_serialized(:runtime_constraints, reusable_runtime_constraints, md5: true, multivalue: true)
326 log_reuse_info(candidates) { "after filtering on runtime_constraints #{attrs[:runtime_constraints].inspect}" }
328 log_reuse_info { "checking for state=Complete with readable output and log..." }
330 select_readable_pdh = Collection.
331 readable_by(current_user).
332 select(:portable_data_hash).
335 usable = candidates.where(state: Complete, exit_code: 0)
336 log_reuse_info(usable) { "with state=Complete, exit_code=0" }
338 usable = usable.where("log IN (#{select_readable_pdh})")
339 log_reuse_info(usable) { "with readable log" }
341 usable = usable.where("output IN (#{select_readable_pdh})")
342 log_reuse_info(usable) { "with readable output" }
344 usable = usable.order('finished_at ASC').limit(1).first
346 log_reuse_info { "done, reusing container #{usable.uuid} with state=Complete" }
350 # Check for non-failing Running candidates and return the most likely to finish sooner.
351 log_reuse_info { "checking for state=Running..." }
352 running = candidates.where(state: Running).
353 where("(runtime_status->'error') is null and priority > 0").
354 order('progress desc, started_at asc').
357 log_reuse_info { "done, reusing container #{running.uuid} with state=Running" }
360 log_reuse_info { "have no containers in Running state" }
363 # Check for Locked or Queued ones and return the most likely to start first.
364 locked_or_queued = candidates.
365 where("state IN (?)", [Locked, Queued]).
366 order('state asc, priority desc, created_at asc').
368 if !attrs[:scheduling_parameters]['preemptible']
369 locked_or_queued = locked_or_queued.
370 where("not ((scheduling_parameters::jsonb)->>'preemptible')::boolean")
372 chosen = locked_or_queued.first
374 log_reuse_info { "done, reusing container #{chosen.uuid} with state=#{chosen.state}" }
377 log_reuse_info { "have no containers in Locked or Queued state" }
380 log_reuse_info { "done, no reusable container found" }
386 if self.state != Queued
387 raise LockFailedError.new("cannot lock when #{self.state}")
389 self.update!(state: Locked)
394 if state_was == Queued and state == Locked
395 if self.priority <= 0
396 raise LockFailedError.new("cannot lock when priority<=0")
398 self.lock_count = self.lock_count+1
404 if self.state != Locked
405 raise InvalidStateTransitionError.new("cannot unlock when #{self.state}")
407 self.update!(state: Queued)
412 if state_was == Locked and state == Queued
413 if self.locked_by_uuid != current_api_client_authorization.uuid
414 raise ArvadosModel::PermissionDeniedError.new("locked by a different token")
416 if self.lock_count >= Rails.configuration.Containers.MaxDispatchAttempts
417 self.state = Cancelled
418 self.runtime_status = {error: "Failed to start container. Cancelled after exceeding 'Containers.MaxDispatchAttempts' (lock_count=#{self.lock_count})"}
423 def self.readable_by(*users_list)
424 # Load optional keyword arguments, if they exist.
425 if users_list.last.is_a? Hash
426 kwargs = users_list.pop
430 if users_list.select { |u| u.is_admin }.any?
433 Container.where(ContainerRequest.readable_by(*users_list).where("containers.uuid = container_requests.container_uuid").arel.exists)
437 [Complete, Cancelled].include?(self.state)
440 def self.for_current_token
441 return if !current_api_client_authorization
442 _, _, _, container_uuid = Thread.current[:token].split('/')
443 if container_uuid.nil?
444 Container.where(auth_uuid: current_api_client_authorization.uuid).first
446 Container.where('auth_uuid=? or (uuid=? and runtime_token=?)',
447 current_api_client_authorization.uuid,
449 current_api_client_authorization.token).first
455 def self.bound_keep_cache_disk(value)
461 elsif value > max_value
468 def self.hash_product(**kwargs)
469 # kwargs is a hash that maps parameters to an array of values.
470 # This function enumerates every possible hash where each key has one of
471 # the values from its array.
472 # The output keys are strings since that's what container hash attributes
474 # A nil value yields a hash without that key.
476 *kwargs.map { |(key, values)| [key.to_s].product(values) },
477 ).map { |param_pairs| Hash[param_pairs].compact }
480 def fill_field_defaults
481 self.state ||= Queued
482 self.environment ||= {}
483 self.runtime_constraints ||= {}
487 self.scheduling_parameters ||= {}
490 def permission_to_create
491 current_user.andand.is_admin
494 def permission_to_destroy
495 current_user.andand.is_admin
498 def ensure_owner_uuid_is_permitted
499 # validate_change ensures owner_uuid can't be changed at all --
500 # except during create, which requires admin privileges. Checking
501 # permission here would be superfluous.
506 if self.state_changed? and self.state == Running
507 self.started_at ||= db_current_time
510 if self.state_changed? and [Complete, Cancelled].include? self.state
511 self.finished_at ||= db_current_time
515 # Check that well-known runtime status keys have desired data types
516 def validate_runtime_status
518 'error', 'errorDetail', 'warning', 'warningDetail', 'activity'
520 if self.runtime_status.andand.include?(k) && !self.runtime_status[k].is_a?(String)
521 errors.add(:runtime_status, "'#{k}' value must be a string")
528 final_attrs = [:finished_at]
529 progress_attrs = [:progress, :runtime_status, :subrequests_cost, :cost,
530 :log, :output, :output_properties, :exit_code]
533 permitted.push(:owner_uuid, :command, :container_image, :cwd,
534 :environment, :mounts, :output_path, :priority,
535 :runtime_constraints, :scheduling_parameters,
536 :secret_mounts, :runtime_token,
537 :runtime_user_uuid, :runtime_auth_scopes,
538 :output_storage_classes)
543 permitted.push :priority, :runtime_status, :log, :lock_count
546 permitted.push :priority
549 permitted.push :priority, :output_properties, :gateway_address, *progress_attrs
550 if self.state_changed?
551 permitted.push :started_at
553 if !self.interactive_session_started_was
554 permitted.push :interactive_session_started
558 if self.state_was == Running
559 permitted.push *final_attrs, *progress_attrs
565 permitted.push :finished_at, *progress_attrs
567 permitted.push :finished_at, :log, :runtime_status, :cost
571 # The state_transitions check will add an error message for this
575 if self.state_was == Running &&
576 !current_api_client_authorization.nil? &&
577 (current_api_client_authorization.uuid == self.auth_uuid ||
578 current_api_client_authorization.token == self.runtime_token)
579 # The contained process itself can write final attrs but can't
580 # change priority or log.
581 permitted.push *final_attrs
582 permitted = permitted - [:log, :priority]
583 elsif !current_user.andand.is_admin
584 raise PermissionDeniedError
585 elsif self.locked_by_uuid && self.locked_by_uuid != current_api_client_authorization.andand.uuid
586 # When locked, progress fields cannot be updated by the wrong
587 # dispatcher, even though it has admin privileges.
588 permitted = permitted - progress_attrs
590 check_update_whitelist permitted
594 if [Locked, Running].include? self.state
595 # If the Container was already locked, locked_by_uuid must not
596 # changes. Otherwise, the current auth gets the lock.
597 need_lock = locked_by_uuid_was || current_api_client_authorization.andand.uuid
602 # The caller can provide a new value for locked_by_uuid, but only
603 # if it's exactly what we expect. This allows a caller to perform
604 # an update like {"state":"Unlocked","locked_by_uuid":null}.
605 if self.locked_by_uuid_changed?
606 if self.locked_by_uuid != need_lock
607 return errors.add :locked_by_uuid, "can only change to #{need_lock}"
610 self.locked_by_uuid = need_lock
614 # Output must exist and be readable by the current user. This is so
615 # that a container cannot "claim" a collection that it doesn't otherwise
616 # have access to just by setting the output field to the collection PDH.
619 readable_by(current_user, {include_trash: true}).
620 where(portable_data_hash: self.output).
623 errors.add :output, "collection must exist and be readable by current user."
629 # If self.final?, this update is superfluous: the final log/output
630 # update will be done when handle_completed calls finalize! on
631 # each requesting CR.
632 return if self.final? || !saved_change_to_log?
633 leave_modified_by_user_alone do
634 ContainerRequest.where(container_uuid: self.uuid, state: ContainerRequest::Committed).each do |cr|
635 cr.update_collections(container: self, collections: ['log'])
642 if self.auth_uuid_changed?
643 return errors.add :auth_uuid, 'is readonly'
645 if not [Locked, Running].include? self.state
646 # Don't need one. If auth already exists, expire it.
648 # We use db_transaction_time here (not db_current_time) to
649 # ensure the token doesn't validate later in the same
650 # transaction (e.g., in a test case) by satisfying expires_at >
651 # transaction timestamp.
652 self.auth.andand.update(expires_at: db_transaction_time)
659 if self.runtime_token.nil?
660 if self.runtime_user_uuid.nil?
661 # legacy behavior, we don't have a runtime_user_uuid so get
662 # the user from the highest priority container request, needed
663 # when performing an upgrade and there are queued containers,
665 cr = ContainerRequest.
666 where('container_uuid=? and priority>0', self.uuid).
667 order('priority desc').
670 return errors.add :auth_uuid, "cannot be assigned because priority <= 0"
672 self.runtime_user_uuid = cr.modified_by_user_uuid
673 self.runtime_auth_scopes = ["all"]
676 # Generate a new token. This runs with admin credentials as it's done by a
677 # dispatcher user, so expires_at isn't enforced by API.MaxTokenLifetime.
678 self.auth = ApiClientAuthorization.
679 create!(user_id: User.find_by_uuid(self.runtime_user_uuid).id,
681 scopes: self.runtime_auth_scopes)
685 def sort_serialized_attrs
686 if self.environment_changed?
687 self.environment = self.class.deep_sort_hash(self.environment)
689 if self.mounts_changed?
690 self.mounts = self.class.deep_sort_hash(self.mounts)
692 if self.runtime_constraints_changed?
693 self.runtime_constraints = self.class.deep_sort_hash(self.runtime_constraints)
695 if self.scheduling_parameters_changed?
696 self.scheduling_parameters = self.class.deep_sort_hash(self.scheduling_parameters)
698 if self.runtime_auth_scopes_changed?
699 self.runtime_auth_scopes = self.runtime_auth_scopes.sort
703 def update_secret_mounts_md5
704 if self.secret_mounts_changed?
705 self.secret_mounts_md5 = Digest::MD5.hexdigest(
706 SafeJSON.dump(self.class.deep_sort_hash(self.secret_mounts)))
711 # this runs after update_secret_mounts_md5, so the
712 # secret_mounts_md5 will still reflect the secrets that are being
714 if self.state_changed? && self.final?
715 self.secret_mounts = {}
716 self.runtime_token = nil
720 def clear_runtime_status_when_queued
721 # Avoid leaking status messages between different dispatch attempts
722 if self.state_was == Locked && self.state == Queued
723 self.runtime_status = {}
728 # This container is finished so finalize any associated container requests
729 # that are associated with this container.
730 if saved_change_to_state? and self.final?
731 # These get wiped out by with_lock (which reloads the record),
732 # so record them now in case we need to schedule a retry.
733 prev_secret_mounts = secret_mounts_before_last_save
734 prev_runtime_token = runtime_token_before_last_save
736 # Need to take a lock on the container to ensure that any
737 # concurrent container requests that might try to reuse this
738 # container will block until the container completion
739 # transaction finishes. This ensure that concurrent container
740 # requests that try to reuse this container are finalized (on
741 # Complete) or don't reuse it (on Cancelled).
743 act_as_system_user do
744 if self.state == Cancelled
745 # Cancelled means the container didn't run to completion.
746 # This happens either because it was cancelled by the user
747 # or because there was an infrastructure failure. We want
748 # to retry infrastructure failures automatically.
750 # Seach for live container requests to determine if we
751 # should retry the container.
752 retryable_requests = ContainerRequest.
753 joins('left outer join containers as requesting_container on container_requests.requesting_container_uuid = requesting_container.uuid').
754 where("container_requests.container_uuid = ? and "+
755 "container_requests.priority > 0 and "+
756 "container_requests.owner_uuid not in (select group_uuid from trashed_groups) and "+
757 "(requesting_container.priority is null or (requesting_container.state = 'Running' and requesting_container.priority > 0)) and "+
758 "container_requests.state = 'Committed' and "+
759 "container_requests.container_count < container_requests.container_count_max", uuid).
760 order('container_requests.uuid asc')
762 retryable_requests = []
765 if retryable_requests.any?
766 scheduling_parameters = {
767 # partitions: empty if any are empty, else the union of all parameters
768 "partitions": retryable_requests
769 .map { |req| req.scheduling_parameters["partitions"] || [] }
770 .reduce { |cur, new| (cur.empty? or new.empty?) ? [] : (cur | new) },
772 # preemptible: true if all are true, else false
773 "preemptible": retryable_requests
774 .map { |req| req.scheduling_parameters["preemptible"] }
777 # supervisor: true if all any true, else false
778 "supervisor": retryable_requests
779 .map { |req| req.scheduling_parameters["supervisor"] }
782 # max_run_time: 0 if any are 0 (unlimited), else the maximum
783 "max_run_time": retryable_requests
784 .map { |req| req.scheduling_parameters["max_run_time"] || 0 }
785 .reduce do |cur, new|
786 if cur == 0 or new == 0
797 command: self.command,
799 environment: self.environment,
800 output_path: self.output_path,
801 container_image: self.container_image,
803 runtime_constraints: self.runtime_constraints,
804 scheduling_parameters: scheduling_parameters,
805 secret_mounts: prev_secret_mounts,
806 runtime_token: prev_runtime_token,
807 runtime_user_uuid: self.runtime_user_uuid,
808 runtime_auth_scopes: self.runtime_auth_scopes
810 c = Container.create! c_attrs
811 retryable_requests.each do |cr|
813 leave_modified_by_user_alone do
814 # Use row locking because this increments container_count
815 cr.cumulative_cost += self.cost + self.subrequests_cost
816 cr.container_uuid = c.uuid
823 # Notify container requests associated with this container
824 ContainerRequest.where(container_uuid: uuid,
825 state: ContainerRequest::Committed).each do |cr|
826 leave_modified_by_user_alone do
831 # Cancel outstanding container requests made by this container.
833 where(requesting_container_uuid: uuid,
834 state: ContainerRequest::Committed).
835 in_batches(of: 15).each_record do |cr|
836 leave_modified_by_user_alone do
838 container_state = Container.where(uuid: cr.container_uuid).pluck(:state).first
839 if container_state == Container::Queued || container_state == Container::Locked
840 # If the child container hasn't started yet, finalize the
841 # child CR now instead of leaving it "on hold", i.e.,
842 # Queued with priority 0. (OTOH, if the child is already
843 # running, leave it alone so it can get cancelled the
844 # usual way, get a copy of the log collection, etc.)
845 cr.update!(state: ContainerRequest::Final)