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
230 if rc['keep_cache_disk'] == 0 and rc['keep_cache_ram'] == 0
231 # Only set if keep_cache_ram isn't set.
232 rc['keep_cache_disk'] = Rails.configuration.Containers.DefaultKeepCacheDisk
237 # Return a mounts hash suitable for a Container, i.e., with every
238 # readonly collection UUID resolved to a PDH.
239 def self.resolve_mounts(mounts)
241 mounts.each do |k, mount|
244 if mount['kind'] != 'collection'
248 uuid = mount.delete 'uuid'
250 if mount['portable_data_hash'].nil? and !uuid.nil?
251 # PDH not supplied, try by UUID
253 readable_by(current_user).
255 select(:portable_data_hash).
258 raise ArvadosModel::UnresolvableContainerError.new "cannot mount collection #{uuid.inspect}: not found"
260 mount['portable_data_hash'] = c.portable_data_hash
266 # Return a container_image PDH suitable for a Container.
267 def self.resolve_container_image(container_image)
268 coll = Collection.for_latest_docker_image(container_image)
270 raise ArvadosModel::UnresolvableContainerError.new "docker image #{container_image.inspect} not found"
272 coll.portable_data_hash
275 def self.find_reusable(attrs)
276 log_reuse_info { "starting with #{Container.all.count} container records in database" }
277 candidates = Container.where_serialized(:command, attrs[:command], md5: true)
278 log_reuse_info(candidates) { "after filtering on command #{attrs[:command].inspect}" }
280 candidates = candidates.where('cwd = ?', attrs[:cwd])
281 log_reuse_info(candidates) { "after filtering on cwd #{attrs[:cwd].inspect}" }
283 candidates = candidates.where_serialized(:environment, attrs[:environment], md5: true)
284 log_reuse_info(candidates) { "after filtering on environment #{attrs[:environment].inspect}" }
286 candidates = candidates.where('output_path = ?', attrs[:output_path])
287 log_reuse_info(candidates) { "after filtering on output_path #{attrs[:output_path].inspect}" }
289 image = resolve_container_image(attrs[:container_image])
290 candidates = candidates.where('container_image = ?', image)
291 log_reuse_info(candidates) { "after filtering on container_image #{image.inspect} (resolved from #{attrs[:container_image].inspect})" }
293 candidates = candidates.where_serialized(:mounts, resolve_mounts(attrs[:mounts]), md5: true)
294 log_reuse_info(candidates) { "after filtering on mounts #{attrs[:mounts].inspect}" }
296 secret_mounts_md5 = Digest::MD5.hexdigest(SafeJSON.dump(self.deep_sort_hash(attrs[:secret_mounts])))
297 candidates = candidates.where('secret_mounts_md5 = ?', secret_mounts_md5)
298 log_reuse_info(candidates) { "after filtering on secret_mounts_md5 #{secret_mounts_md5.inspect}" }
300 if attrs[:runtime_constraints]['cuda'].nil?
301 attrs[:runtime_constraints]['cuda'] = {
303 'driver_version' => '',
304 'hardware_capability' => '',
307 resolved_runtime_constraints = [resolve_runtime_constraints(attrs[:runtime_constraints])]
308 if resolved_runtime_constraints[0]['cuda']['device_count'] == 0
309 # If no CUDA requested, extend search to include older container
310 # records that don't have a 'cuda' section in runtime_constraints
311 resolved_runtime_constraints << resolved_runtime_constraints[0].except('cuda')
313 if resolved_runtime_constraints[0]['keep_cache_disk'] == 0
314 # If no disk cache requested, extend search to include older container
315 # records that don't have a 'keep_cache_disk' field in runtime_constraints
316 if resolved_runtime_constraints.length == 2
317 # exclude the one that also excludes CUDA
318 resolved_runtime_constraints << resolved_runtime_constraints[1].except('keep_cache_disk')
320 resolved_runtime_constraints << resolved_runtime_constraints[0].except('keep_cache_disk')
323 candidates = candidates.where_serialized(:runtime_constraints, resolved_runtime_constraints, md5: true, multivalue: true)
324 log_reuse_info(candidates) { "after filtering on runtime_constraints #{attrs[:runtime_constraints].inspect}" }
326 log_reuse_info { "checking for state=Complete with readable output and log..." }
328 select_readable_pdh = Collection.
329 readable_by(current_user).
330 select(:portable_data_hash).
333 usable = candidates.where(state: Complete, exit_code: 0)
334 log_reuse_info(usable) { "with state=Complete, exit_code=0" }
336 usable = usable.where("log IN (#{select_readable_pdh})")
337 log_reuse_info(usable) { "with readable log" }
339 usable = usable.where("output IN (#{select_readable_pdh})")
340 log_reuse_info(usable) { "with readable output" }
342 usable = usable.order('finished_at ASC').limit(1).first
344 log_reuse_info { "done, reusing container #{usable.uuid} with state=Complete" }
348 # Check for non-failing Running candidates and return the most likely to finish sooner.
349 log_reuse_info { "checking for state=Running..." }
350 running = candidates.where(state: Running).
351 where("(runtime_status->'error') is null").
352 order('progress desc, started_at asc').
355 log_reuse_info { "done, reusing container #{running.uuid} with state=Running" }
358 log_reuse_info { "have no containers in Running state" }
361 # Check for Locked or Queued ones and return the most likely to start first.
362 locked_or_queued = candidates.
363 where("state IN (?)", [Locked, Queued]).
364 order('state asc, priority desc, created_at asc').
367 log_reuse_info { "done, reusing container #{locked_or_queued.uuid} with state=#{locked_or_queued.state}" }
368 return locked_or_queued
370 log_reuse_info { "have no containers in Locked or Queued state" }
373 log_reuse_info { "done, no reusable container found" }
379 if self.state != Queued
380 raise LockFailedError.new("cannot lock when #{self.state}")
382 self.update_attributes!(state: Locked)
387 if state_was == Queued and state == Locked
388 if self.priority <= 0
389 raise LockFailedError.new("cannot lock when priority<=0")
391 self.lock_count = self.lock_count+1
397 if self.state != Locked
398 raise InvalidStateTransitionError.new("cannot unlock when #{self.state}")
400 self.update_attributes!(state: Queued)
405 if state_was == Locked and state == Queued
406 if self.locked_by_uuid != current_api_client_authorization.uuid
407 raise ArvadosModel::PermissionDeniedError.new("locked by a different token")
409 if self.lock_count >= Rails.configuration.Containers.MaxDispatchAttempts
410 self.state = Cancelled
411 self.runtime_status = {error: "Failed to start container. Cancelled after exceeding 'Containers.MaxDispatchAttempts' (lock_count=#{self.lock_count})"}
416 def self.readable_by(*users_list)
417 # Load optional keyword arguments, if they exist.
418 if users_list.last.is_a? Hash
419 kwargs = users_list.pop
423 if users_list.select { |u| u.is_admin }.any?
426 Container.where(ContainerRequest.readable_by(*users_list).where("containers.uuid = container_requests.container_uuid").arel.exists)
430 [Complete, Cancelled].include?(self.state)
433 def self.for_current_token
434 return if !current_api_client_authorization
435 _, _, _, container_uuid = Thread.current[:token].split('/')
436 if container_uuid.nil?
437 Container.where(auth_uuid: current_api_client_authorization.uuid).first
439 Container.where('auth_uuid=? or (uuid=? and runtime_token=?)',
440 current_api_client_authorization.uuid,
442 current_api_client_authorization.token).first
448 def fill_field_defaults
449 self.state ||= Queued
450 self.environment ||= {}
451 self.runtime_constraints ||= {}
455 self.scheduling_parameters ||= {}
458 def permission_to_create
459 current_user.andand.is_admin
462 def permission_to_destroy
463 current_user.andand.is_admin
466 def ensure_owner_uuid_is_permitted
467 # validate_change ensures owner_uuid can't be changed at all --
468 # except during create, which requires admin privileges. Checking
469 # permission here would be superfluous.
474 if self.state_changed? and self.state == Running
475 self.started_at ||= db_current_time
478 if self.state_changed? and [Complete, Cancelled].include? self.state
479 self.finished_at ||= db_current_time
483 # Check that well-known runtime status keys have desired data types
484 def validate_runtime_status
486 'error', 'errorDetail', 'warning', 'warningDetail', 'activity'
488 if self.runtime_status.andand.include?(k) && !self.runtime_status[k].is_a?(String)
489 errors.add(:runtime_status, "'#{k}' value must be a string")
496 final_attrs = [:finished_at]
497 progress_attrs = [:progress, :runtime_status, :subrequests_cost, :cost,
498 :log, :output, :output_properties, :exit_code]
501 permitted.push(:owner_uuid, :command, :container_image, :cwd,
502 :environment, :mounts, :output_path, :priority,
503 :runtime_constraints, :scheduling_parameters,
504 :secret_mounts, :runtime_token,
505 :runtime_user_uuid, :runtime_auth_scopes,
506 :output_storage_classes)
511 permitted.push :priority, :runtime_status, :log, :lock_count
514 permitted.push :priority
517 permitted.push :priority, :output_properties, :gateway_address, *progress_attrs
518 if self.state_changed?
519 permitted.push :started_at
521 if !self.interactive_session_started_was
522 permitted.push :interactive_session_started
526 if self.state_was == Running
527 permitted.push *final_attrs, *progress_attrs
533 permitted.push :finished_at, *progress_attrs
535 permitted.push :finished_at, :log, :runtime_status, :cost
539 # The state_transitions check will add an error message for this
543 if self.state_was == Running &&
544 !current_api_client_authorization.nil? &&
545 (current_api_client_authorization.uuid == self.auth_uuid ||
546 current_api_client_authorization.token == self.runtime_token)
547 # The contained process itself can write final attrs but can't
548 # change priority or log.
549 permitted.push *final_attrs
550 permitted = permitted - [:log, :priority]
551 elsif !current_user.andand.is_admin
552 raise PermissionDeniedError
553 elsif self.locked_by_uuid && self.locked_by_uuid != current_api_client_authorization.andand.uuid
554 # When locked, progress fields cannot be updated by the wrong
555 # dispatcher, even though it has admin privileges.
556 permitted = permitted - progress_attrs
558 check_update_whitelist permitted
562 if [Locked, Running].include? self.state
563 # If the Container was already locked, locked_by_uuid must not
564 # changes. Otherwise, the current auth gets the lock.
565 need_lock = locked_by_uuid_was || current_api_client_authorization.andand.uuid
570 # The caller can provide a new value for locked_by_uuid, but only
571 # if it's exactly what we expect. This allows a caller to perform
572 # an update like {"state":"Unlocked","locked_by_uuid":null}.
573 if self.locked_by_uuid_changed?
574 if self.locked_by_uuid != need_lock
575 return errors.add :locked_by_uuid, "can only change to #{need_lock}"
578 self.locked_by_uuid = need_lock
582 # Output must exist and be readable by the current user. This is so
583 # that a container cannot "claim" a collection that it doesn't otherwise
584 # have access to just by setting the output field to the collection PDH.
587 readable_by(current_user, {include_trash: true}).
588 where(portable_data_hash: self.output).
591 errors.add :output, "collection must exist and be readable by current user."
597 # If self.final?, this update is superfluous: the final log/output
598 # update will be done when handle_completed calls finalize! on
599 # each requesting CR.
600 return if self.final? || !saved_change_to_log?
601 leave_modified_by_user_alone do
602 ContainerRequest.where(container_uuid: self.uuid).each do |cr|
603 cr.update_collections(container: self, collections: ['log'])
610 if self.auth_uuid_changed?
611 return errors.add :auth_uuid, 'is readonly'
613 if not [Locked, Running].include? self.state
614 # Don't need one. If auth already exists, expire it.
616 # We use db_transaction_time here (not db_current_time) to
617 # ensure the token doesn't validate later in the same
618 # transaction (e.g., in a test case) by satisfying expires_at >
619 # transaction timestamp.
620 self.auth.andand.update_attributes(expires_at: db_transaction_time)
627 if self.runtime_token.nil?
628 if self.runtime_user_uuid.nil?
629 # legacy behavior, we don't have a runtime_user_uuid so get
630 # the user from the highest priority container request, needed
631 # when performing an upgrade and there are queued containers,
633 cr = ContainerRequest.
634 where('container_uuid=? and priority>0', self.uuid).
635 order('priority desc').
638 return errors.add :auth_uuid, "cannot be assigned because priority <= 0"
640 self.runtime_user_uuid = cr.modified_by_user_uuid
641 self.runtime_auth_scopes = ["all"]
644 # Generate a new token. This runs with admin credentials as it's done by a
645 # dispatcher user, so expires_at isn't enforced by API.MaxTokenLifetime.
646 self.auth = ApiClientAuthorization.
647 create!(user_id: User.find_by_uuid(self.runtime_user_uuid).id,
649 scopes: self.runtime_auth_scopes)
653 def sort_serialized_attrs
654 if self.environment_changed?
655 self.environment = self.class.deep_sort_hash(self.environment)
657 if self.mounts_changed?
658 self.mounts = self.class.deep_sort_hash(self.mounts)
660 if self.runtime_constraints_changed?
661 self.runtime_constraints = self.class.deep_sort_hash(self.runtime_constraints)
663 if self.scheduling_parameters_changed?
664 self.scheduling_parameters = self.class.deep_sort_hash(self.scheduling_parameters)
666 if self.runtime_auth_scopes_changed?
667 self.runtime_auth_scopes = self.runtime_auth_scopes.sort
671 def update_secret_mounts_md5
672 if self.secret_mounts_changed?
673 self.secret_mounts_md5 = Digest::MD5.hexdigest(
674 SafeJSON.dump(self.class.deep_sort_hash(self.secret_mounts)))
679 # this runs after update_secret_mounts_md5, so the
680 # secret_mounts_md5 will still reflect the secrets that are being
682 if self.state_changed? && self.final?
683 self.secret_mounts = {}
684 self.runtime_token = nil
688 def clear_runtime_status_when_queued
689 # Avoid leaking status messages between different dispatch attempts
690 if self.state_was == Locked && self.state == Queued
691 self.runtime_status = {}
696 # This container is finished so finalize any associated container requests
697 # that are associated with this container.
698 if saved_change_to_state? and self.final?
699 # These get wiped out by with_lock (which reloads the record),
700 # so record them now in case we need to schedule a retry.
701 prev_secret_mounts = secret_mounts_before_last_save
702 prev_runtime_token = runtime_token_before_last_save
704 # Need to take a lock on the container to ensure that any
705 # concurrent container requests that might try to reuse this
706 # container will block until the container completion
707 # transaction finishes. This ensure that concurrent container
708 # requests that try to reuse this container are finalized (on
709 # Complete) or don't reuse it (on Cancelled).
711 act_as_system_user do
712 if self.state == Cancelled
713 retryable_requests = ContainerRequest.where("container_uuid = ? and priority > 0 and state = 'Committed' and container_count < container_count_max", uuid)
715 retryable_requests = []
718 if retryable_requests.any?
720 command: self.command,
722 environment: self.environment,
723 output_path: self.output_path,
724 container_image: self.container_image,
726 runtime_constraints: self.runtime_constraints,
727 scheduling_parameters: self.scheduling_parameters,
728 secret_mounts: prev_secret_mounts,
729 runtime_token: prev_runtime_token,
730 runtime_user_uuid: self.runtime_user_uuid,
731 runtime_auth_scopes: self.runtime_auth_scopes
733 c = Container.create! c_attrs
734 retryable_requests.each do |cr|
736 leave_modified_by_user_alone do
737 # Use row locking because this increments container_count
738 cr.cumulative_cost += self.cost + self.subrequests_cost
739 cr.container_uuid = c.uuid
746 # Notify container requests associated with this container
747 ContainerRequest.where(container_uuid: uuid,
748 state: ContainerRequest::Committed).each do |cr|
749 leave_modified_by_user_alone do
754 # Cancel outstanding container requests made by this container.
756 includes(:container).
757 where(requesting_container_uuid: uuid,
758 state: ContainerRequest::Committed).each do |cr|
759 leave_modified_by_user_alone do
760 cr.update_attributes!(priority: 0)
762 if cr.container.state == Container::Queued || cr.container.state == Container::Locked
763 # If the child container hasn't started yet, finalize the
764 # child CR now instead of leaving it "on hold", i.e.,
765 # Queued with priority 0. (OTOH, if the child is already
766 # running, leave it alone so it can get cancelled the
767 # usual way, get a copy of the log collection, etc.)
768 cr.update_attributes!(state: ContainerRequest::Final)