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, :foreign_key => :container_uuid, :class_name => 'ContainerRequest', :primary_key => :uuid
55 belongs_to :auth, :class_name => 'ApiClientAuthorization', :foreign_key => :auth_uuid, :primary_key => :uuid
57 api_accessible :user, extend: :common do |t|
59 t.add :container_image
71 t.add :runtime_constraints
76 t.add :scheduling_parameters
77 t.add :runtime_user_uuid
78 t.add :runtime_auth_scopes
80 t.add :gateway_address
81 t.add :interactive_session_started
82 t.add :output_storage_classes
83 t.add :output_properties
85 t.add :subrequests_cost
88 # Supported states for a container
93 (Running = 'Running'),
94 (Complete = 'Complete'),
95 (Cancelled = 'Cancelled')
100 Queued => [Locked, Cancelled],
101 Locked => [Queued, Running, Cancelled],
102 Running => [Complete, Cancelled],
103 Complete => [Cancelled]
106 def self.limit_index_columns_read
110 def self.full_text_searchable_columns
111 super - ["secret_mounts", "secret_mounts_md5", "runtime_token", "gateway_address", "output_storage_classes"]
114 def self.searchable_columns *args
115 super - ["secret_mounts_md5", "runtime_token", "gateway_address", "output_storage_classes"]
118 def logged_attributes
119 super.except('secret_mounts', 'runtime_token')
122 def state_transitions
126 # Container priority is the highest "computed priority" of any
127 # matching request. The computed priority of a container-submitted
128 # request is the priority of the submitting container. The computed
129 # priority of a user-submitted request is a function of
130 # user-assigned priority and request creation time.
132 update_priorities uuid
136 # Create a new container (or find an existing one) to satisfy the
137 # given container request.
138 def self.resolve(req)
139 if req.runtime_token.nil?
140 runtime_user = if req.modified_by_user_uuid.nil?
143 User.find_by_uuid(req.modified_by_user_uuid)
145 runtime_auth_scopes = ["all"]
147 auth = ApiClientAuthorization.validate(token: req.runtime_token)
149 raise ArgumentError.new "Invalid runtime token"
151 runtime_user = User.find_by_id(auth.user_id)
152 runtime_auth_scopes = auth.scopes
154 c_attrs = act_as_user runtime_user do
156 command: req.command,
158 environment: req.environment,
159 output_path: req.output_path,
160 container_image: resolve_container_image(req.container_image),
161 mounts: resolve_mounts(req.mounts),
162 runtime_constraints: resolve_runtime_constraints(req.runtime_constraints),
163 scheduling_parameters: req.scheduling_parameters,
164 secret_mounts: req.secret_mounts,
165 runtime_token: req.runtime_token,
166 runtime_user_uuid: runtime_user.uuid,
167 runtime_auth_scopes: runtime_auth_scopes,
168 output_storage_classes: req.output_storage_classes,
171 act_as_system_user do
172 if req.use_existing && (reusable = find_reusable(c_attrs))
175 Container.create!(c_attrs)
180 # Return a runtime_constraints hash that complies with requested but
181 # is suitable for saving in a container record, i.e., has specific
182 # values instead of ranges.
184 # Doing this as a step separate from other resolutions, like "git
185 # revision range to commit hash", makes sense only when there is no
186 # opportunity to reuse an existing container (e.g., container reuse
187 # is not implemented yet, or we have already found that no existing
188 # containers are suitable).
189 def self.resolve_runtime_constraints(runtime_constraints)
191 runtime_constraints.each do |k, v|
198 if rc['keep_cache_ram'] == 0
199 rc['keep_cache_ram'] = Rails.configuration.Containers.DefaultKeepCacheRAM
201 if rc['keep_cache_disk'] == 0 and rc['keep_cache_ram'] == 0
202 rc['keep_cache_disk'] = bound_keep_cache_disk(rc['ram'])
207 # Return a mounts hash suitable for a Container, i.e., with every
208 # readonly collection UUID resolved to a PDH.
209 def self.resolve_mounts(mounts)
211 mounts.each do |k, mount|
214 if mount['kind'] != 'collection'
218 uuid = mount.delete 'uuid'
220 if mount['portable_data_hash'].nil? and !uuid.nil?
221 # PDH not supplied, try by UUID
223 readable_by(current_user).
225 select(:portable_data_hash).
228 raise ArvadosModel::UnresolvableContainerError.new "cannot mount collection #{uuid.inspect}: not found"
230 mount['portable_data_hash'] = c.portable_data_hash
236 # Return a container_image PDH suitable for a Container.
237 def self.resolve_container_image(container_image)
238 coll = Collection.for_latest_docker_image(container_image)
240 raise ArvadosModel::UnresolvableContainerError.new "docker image #{container_image.inspect} not found"
242 coll.portable_data_hash
245 def self.find_reusable(attrs)
246 log_reuse_info { "starting with #{Container.all.count} container records in database" }
247 candidates = Container.where_serialized(:command, attrs[:command], md5: true)
248 log_reuse_info(candidates) { "after filtering on command #{attrs[:command].inspect}" }
250 candidates = candidates.where('cwd = ?', attrs[:cwd])
251 log_reuse_info(candidates) { "after filtering on cwd #{attrs[:cwd].inspect}" }
253 candidates = candidates.where_serialized(:environment, attrs[:environment], md5: true)
254 log_reuse_info(candidates) { "after filtering on environment #{attrs[:environment].inspect}" }
256 candidates = candidates.where('output_path = ?', attrs[:output_path])
257 log_reuse_info(candidates) { "after filtering on output_path #{attrs[:output_path].inspect}" }
259 image = resolve_container_image(attrs[:container_image])
260 candidates = candidates.where('container_image = ?', image)
261 log_reuse_info(candidates) { "after filtering on container_image #{image.inspect} (resolved from #{attrs[:container_image].inspect})" }
263 candidates = candidates.where_serialized(:mounts, resolve_mounts(attrs[:mounts]), md5: true)
264 log_reuse_info(candidates) { "after filtering on mounts #{attrs[:mounts].inspect}" }
266 secret_mounts_md5 = Digest::MD5.hexdigest(SafeJSON.dump(self.deep_sort_hash(attrs[:secret_mounts])))
267 candidates = candidates.where('secret_mounts_md5 = ?', secret_mounts_md5)
268 log_reuse_info(candidates) { "after filtering on secret_mounts_md5 #{secret_mounts_md5.inspect}" }
270 resolved_runtime_constraints = resolve_runtime_constraints(attrs[:runtime_constraints])
271 # Ideally we would completely ignore Keep cache constraints when making
272 # reuse considerations, but our database structure makes that impractical.
273 # The best we can do is generate a search that matches on all likely values.
274 runtime_constraint_variations = {
276 # Check for constraints without keep_cache_disk
277 # (containers that predate the constraint)
279 # Containers that use keep_cache_ram instead
282 bound_keep_cache_disk(resolved_runtime_constraints['ram']),
283 # The minimum default bound
284 bound_keep_cache_disk(0),
285 # The maximum default bound (presumably)
286 bound_keep_cache_disk(1 << 60),
287 # The requested value
288 resolved_runtime_constraints.delete('keep_cache_disk'),
291 # Containers that use keep_cache_disk instead
294 Rails.configuration.Containers.DefaultKeepCacheRAM,
295 # The requested value
296 resolved_runtime_constraints.delete('keep_cache_ram'),
299 resolved_cuda = resolved_runtime_constraints['cuda']
300 if resolved_cuda.nil? or resolved_cuda['device_count'] == 0
301 runtime_constraint_variations[:cuda] = [
302 # Check for constraints without cuda
303 # (containers that predate the constraint)
305 # The default "don't need CUDA" value
308 'driver_version' => '',
309 'hardware_capability' => '',
311 # The requested value
312 resolved_runtime_constraints.delete('cuda')
315 reusable_runtime_constraints = hash_product(runtime_constraint_variations)
316 .map { |v| resolved_runtime_constraints.merge(v) }
318 candidates = candidates.where_serialized(:runtime_constraints, reusable_runtime_constraints, md5: true, multivalue: true)
319 log_reuse_info(candidates) { "after filtering on runtime_constraints #{attrs[:runtime_constraints].inspect}" }
321 log_reuse_info { "checking for state=Complete with readable output and log..." }
323 select_readable_pdh = Collection.
324 readable_by(current_user).
325 select(:portable_data_hash).
328 usable = candidates.where(state: Complete, exit_code: 0)
329 log_reuse_info(usable) { "with state=Complete, exit_code=0" }
331 usable = usable.where("log IN (#{select_readable_pdh})")
332 log_reuse_info(usable) { "with readable log" }
334 usable = usable.where("output IN (#{select_readable_pdh})")
335 log_reuse_info(usable) { "with readable output" }
337 usable = usable.order('finished_at ASC').limit(1).first
339 log_reuse_info { "done, reusing container #{usable.uuid} with state=Complete" }
343 # Check for non-failing Running candidates and return the most likely to finish sooner.
344 log_reuse_info { "checking for state=Running..." }
345 running = candidates.where(state: Running).
346 where("(runtime_status->'error') is null and priority > 0").
347 order('progress desc, started_at asc').
350 log_reuse_info { "done, reusing container #{running.uuid} with state=Running" }
353 log_reuse_info { "have no containers in Running state" }
356 # Check for Locked or Queued ones and return the most likely to start first.
357 locked_or_queued = candidates.
358 where("state IN (?)", [Locked, Queued]).
359 order('state asc, priority desc, created_at asc').
361 if !attrs[:scheduling_parameters]['preemptible']
362 locked_or_queued = locked_or_queued.
363 where("not ((scheduling_parameters::jsonb)->>'preemptible')::boolean")
365 chosen = locked_or_queued.first
367 log_reuse_info { "done, reusing container #{chosen.uuid} with state=#{chosen.state}" }
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 self.bound_keep_cache_disk(value)
454 elsif value > max_value
461 def self.hash_product(**kwargs)
462 # kwargs is a hash that maps parameters to an array of values.
463 # This function enumerates every possible hash where each key has one of
464 # the values from its array.
465 # The output keys are strings since that's what container hash attributes
467 # A nil value yields a hash without that key.
469 *kwargs.map { |(key, values)| [key.to_s].product(values) },
470 ).map { |param_pairs| Hash[param_pairs].compact }
473 def fill_field_defaults
474 self.state ||= Queued
475 self.environment ||= {}
476 self.runtime_constraints ||= {}
480 self.scheduling_parameters ||= {}
483 def permission_to_create
484 current_user.andand.is_admin
487 def permission_to_destroy
488 current_user.andand.is_admin
491 def ensure_owner_uuid_is_permitted
492 # validate_change ensures owner_uuid can't be changed at all --
493 # except during create, which requires admin privileges. Checking
494 # permission here would be superfluous.
499 if self.state_changed? and self.state == Running
500 self.started_at ||= db_current_time
503 if self.state_changed? and [Complete, Cancelled].include? self.state
504 self.finished_at ||= db_current_time
508 # Check that well-known runtime status keys have desired data types
509 def validate_runtime_status
511 'error', 'errorDetail', 'warning', 'warningDetail', 'activity'
513 if self.runtime_status.andand.include?(k) && !self.runtime_status[k].is_a?(String)
514 errors.add(:runtime_status, "'#{k}' value must be a string")
521 final_attrs = [:finished_at]
522 progress_attrs = [:progress, :runtime_status, :subrequests_cost, :cost,
523 :log, :output, :output_properties, :exit_code]
526 permitted.push(:owner_uuid, :command, :container_image, :cwd,
527 :environment, :mounts, :output_path, :priority,
528 :runtime_constraints, :scheduling_parameters,
529 :secret_mounts, :runtime_token,
530 :runtime_user_uuid, :runtime_auth_scopes,
531 :output_storage_classes)
536 permitted.push :priority, :runtime_status, :log, :lock_count
539 permitted.push :priority
542 permitted.push :priority, :output_properties, :gateway_address, *progress_attrs
543 if self.state_changed?
544 permitted.push :started_at
546 if !self.interactive_session_started_was
547 permitted.push :interactive_session_started
551 if self.state_was == Running
552 permitted.push *final_attrs, *progress_attrs
558 permitted.push :finished_at, *progress_attrs
560 permitted.push :finished_at, :log, :runtime_status, :cost
564 # The state_transitions check will add an error message for this
568 if self.state_was == Running &&
569 !current_api_client_authorization.nil? &&
570 (current_api_client_authorization.uuid == self.auth_uuid ||
571 current_api_client_authorization.token == self.runtime_token)
572 # The contained process itself can write final attrs but can't
573 # change priority or log.
574 permitted.push *final_attrs
575 permitted = permitted - [:log, :priority]
576 elsif !current_user.andand.is_admin
577 raise PermissionDeniedError
578 elsif self.locked_by_uuid && self.locked_by_uuid != current_api_client_authorization.andand.uuid
579 # When locked, progress fields cannot be updated by the wrong
580 # dispatcher, even though it has admin privileges.
581 permitted = permitted - progress_attrs
583 check_update_whitelist permitted
587 if [Locked, Running].include? self.state
588 # If the Container was already locked, locked_by_uuid must not
589 # changes. Otherwise, the current auth gets the lock.
590 need_lock = locked_by_uuid_was || current_api_client_authorization.andand.uuid
595 # The caller can provide a new value for locked_by_uuid, but only
596 # if it's exactly what we expect. This allows a caller to perform
597 # an update like {"state":"Unlocked","locked_by_uuid":null}.
598 if self.locked_by_uuid_changed?
599 if self.locked_by_uuid != need_lock
600 return errors.add :locked_by_uuid, "can only change to #{need_lock}"
603 self.locked_by_uuid = need_lock
607 # Output must exist and be readable by the current user. This is so
608 # that a container cannot "claim" a collection that it doesn't otherwise
609 # have access to just by setting the output field to the collection PDH.
612 readable_by(current_user, {include_trash: true}).
613 where(portable_data_hash: self.output).
616 errors.add :output, "collection must exist and be readable by current user."
622 # If self.final?, this update is superfluous: the final log/output
623 # update will be done when handle_completed calls finalize! on
624 # each requesting CR.
625 return if self.final? || !saved_change_to_log?
626 leave_modified_by_user_alone do
627 ContainerRequest.where(container_uuid: self.uuid, state: ContainerRequest::Committed).each do |cr|
628 cr.update_collections(container: self, collections: ['log'])
635 if self.auth_uuid_changed?
636 return errors.add :auth_uuid, 'is readonly'
638 if not [Locked, Running].include? self.state
639 # Don't need one. If auth already exists, expire it.
641 # We use db_transaction_time here (not db_current_time) to
642 # ensure the token doesn't validate later in the same
643 # transaction (e.g., in a test case) by satisfying expires_at >
644 # transaction timestamp.
645 self.auth.andand.update_attributes(expires_at: db_transaction_time)
652 if self.runtime_token.nil?
653 if self.runtime_user_uuid.nil?
654 # legacy behavior, we don't have a runtime_user_uuid so get
655 # the user from the highest priority container request, needed
656 # when performing an upgrade and there are queued containers,
658 cr = ContainerRequest.
659 where('container_uuid=? and priority>0', self.uuid).
660 order('priority desc').
663 return errors.add :auth_uuid, "cannot be assigned because priority <= 0"
665 self.runtime_user_uuid = cr.modified_by_user_uuid
666 self.runtime_auth_scopes = ["all"]
669 # Generate a new token. This runs with admin credentials as it's done by a
670 # dispatcher user, so expires_at isn't enforced by API.MaxTokenLifetime.
671 self.auth = ApiClientAuthorization.
672 create!(user_id: User.find_by_uuid(self.runtime_user_uuid).id,
674 scopes: self.runtime_auth_scopes)
678 def sort_serialized_attrs
679 if self.environment_changed?
680 self.environment = self.class.deep_sort_hash(self.environment)
682 if self.mounts_changed?
683 self.mounts = self.class.deep_sort_hash(self.mounts)
685 if self.runtime_constraints_changed?
686 self.runtime_constraints = self.class.deep_sort_hash(self.runtime_constraints)
688 if self.scheduling_parameters_changed?
689 self.scheduling_parameters = self.class.deep_sort_hash(self.scheduling_parameters)
691 if self.runtime_auth_scopes_changed?
692 self.runtime_auth_scopes = self.runtime_auth_scopes.sort
696 def update_secret_mounts_md5
697 if self.secret_mounts_changed?
698 self.secret_mounts_md5 = Digest::MD5.hexdigest(
699 SafeJSON.dump(self.class.deep_sort_hash(self.secret_mounts)))
704 # this runs after update_secret_mounts_md5, so the
705 # secret_mounts_md5 will still reflect the secrets that are being
707 if self.state_changed? && self.final?
708 self.secret_mounts = {}
709 self.runtime_token = nil
713 def clear_runtime_status_when_queued
714 # Avoid leaking status messages between different dispatch attempts
715 if self.state_was == Locked && self.state == Queued
716 self.runtime_status = {}
721 # This container is finished so finalize any associated container requests
722 # that are associated with this container.
723 if saved_change_to_state? and self.final?
724 # These get wiped out by with_lock (which reloads the record),
725 # so record them now in case we need to schedule a retry.
726 prev_secret_mounts = secret_mounts_before_last_save
727 prev_runtime_token = runtime_token_before_last_save
729 # Need to take a lock on the container to ensure that any
730 # concurrent container requests that might try to reuse this
731 # container will block until the container completion
732 # transaction finishes. This ensure that concurrent container
733 # requests that try to reuse this container are finalized (on
734 # Complete) or don't reuse it (on Cancelled).
736 act_as_system_user do
737 if self.state == Cancelled
738 # Cancelled means the container didn't run to completion.
739 # This happens either because it was cancelled by the user
740 # or because there was an infrastructure failure. We want
741 # to retry infrastructure failures automatically.
743 # Seach for live container requests to determine if we
744 # should retry the container.
745 retryable_requests = ContainerRequest.
746 joins('left outer join containers as requesting_container on container_requests.requesting_container_uuid = requesting_container.uuid').
747 where("container_requests.container_uuid = ? and "+
748 "container_requests.priority > 0 and "+
749 "container_requests.owner_uuid not in (select group_uuid from trashed_groups) and "+
750 "(requesting_container.priority is null or (requesting_container.state = 'Running' and requesting_container.priority > 0)) and "+
751 "container_requests.state = 'Committed' and "+
752 "container_requests.container_count < container_requests.container_count_max", uuid).
753 order('container_requests.uuid asc')
755 retryable_requests = []
758 if retryable_requests.any?
759 scheduling_parameters = {
760 # partitions: empty if any are empty, else the union of all parameters
761 "partitions": retryable_requests
762 .map { |req| req.scheduling_parameters["partitions"] || [] }
763 .reduce { |cur, new| (cur.empty? or new.empty?) ? [] : (cur | new) },
765 # preemptible: true if all are true, else false
766 "preemptible": retryable_requests
767 .map { |req| req.scheduling_parameters["preemptible"] }
770 # supervisor: true if all any true, else false
771 "supervisor": retryable_requests
772 .map { |req| req.scheduling_parameters["supervisor"] }
775 # max_run_time: 0 if any are 0 (unlimited), else the maximum
776 "max_run_time": retryable_requests
777 .map { |req| req.scheduling_parameters["max_run_time"] || 0 }
778 .reduce do |cur, new|
779 if cur == 0 or new == 0
790 command: self.command,
792 environment: self.environment,
793 output_path: self.output_path,
794 container_image: self.container_image,
796 runtime_constraints: self.runtime_constraints,
797 scheduling_parameters: scheduling_parameters,
798 secret_mounts: prev_secret_mounts,
799 runtime_token: prev_runtime_token,
800 runtime_user_uuid: self.runtime_user_uuid,
801 runtime_auth_scopes: self.runtime_auth_scopes
803 c = Container.create! c_attrs
804 retryable_requests.each do |cr|
806 leave_modified_by_user_alone do
807 # Use row locking because this increments container_count
808 cr.cumulative_cost += self.cost + self.subrequests_cost
809 cr.container_uuid = c.uuid
816 # Notify container requests associated with this container
817 ContainerRequest.where(container_uuid: uuid,
818 state: ContainerRequest::Committed).each do |cr|
819 leave_modified_by_user_alone do
824 # Cancel outstanding container requests made by this container.
826 where(requesting_container_uuid: uuid,
827 state: ContainerRequest::Committed).
828 in_batches(of: 15).each_record do |cr|
829 leave_modified_by_user_alone do
831 container_state = Container.where(uuid: cr.container_uuid).pluck(:state).first
832 if container_state == Container::Queued || container_state == Container::Locked
833 # If the child container hasn't started yet, finalize the
834 # child CR now instead of leaving it "on hold", i.e.,
835 # Queued with priority 0. (OTOH, if the child is already
836 # running, leave it alone so it can get cancelled the
837 # usual way, get a copy of the log collection, etc.)
838 cr.update_attributes!(state: ContainerRequest::Final)