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, :jsonbHash, default: {}
26 serialize :environment, Hash
27 serialize :mounts, Hash
28 serialize :runtime_constraints, Hash
29 serialize :command, Array
30 serialize :scheduling_parameters, Hash
32 after_find :fill_container_defaults_after_find
33 before_validation :fill_field_defaults, :if => :new_record?
34 before_validation :set_timestamps
35 before_validation :check_lock
36 before_validation :check_unlock
37 validates :command, :container_image, :output_path, :cwd, :priority, { presence: true }
38 validates :priority, numericality: { only_integer: true, greater_than_or_equal_to: 0 }
39 validate :validate_runtime_status
40 validate :validate_state_change
41 validate :validate_change
42 validate :validate_lock
43 validate :validate_output
44 after_validation :assign_auth
45 before_save :sort_serialized_attrs
46 before_save :update_secret_mounts_md5
47 before_save :scrub_secrets
48 before_save :clear_runtime_status_when_queued
49 after_save :update_cr_logs
50 after_save :handle_completed
51 after_save :propagate_priority
52 after_commit { UpdatePriority.run_update_thread }
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
84 # Supported states for a container
89 (Running = 'Running'),
90 (Complete = 'Complete'),
91 (Cancelled = 'Cancelled')
96 Queued => [Locked, Cancelled],
97 Locked => [Queued, Running, Cancelled],
98 Running => [Complete, Cancelled],
99 Complete => [Cancelled]
102 def self.limit_index_columns_read
106 def self.full_text_searchable_columns
107 super - ["secret_mounts", "secret_mounts_md5", "runtime_token", "gateway_address"]
110 def self.searchable_columns *args
111 super - ["secret_mounts_md5", "runtime_token", "gateway_address"]
114 def logged_attributes
115 super.except('secret_mounts', 'runtime_token')
118 def state_transitions
122 # Container priority is the highest "computed priority" of any
123 # matching request. The computed priority of a container-submitted
124 # request is the priority of the submitting container. The computed
125 # priority of a user-submitted request is a function of
126 # user-assigned priority and request creation time.
128 return if ![Queued, Locked, Running].include?(state)
129 p = ContainerRequest.
130 where('container_uuid=? and priority>0', uuid).
131 includes(:requesting_container).
134 if cr.requesting_container
135 cr.requesting_container.priority
137 (cr.priority << 50) - (cr.created_at.to_time.to_f * 1000).to_i
140 update_attributes!(priority: p)
143 def propagate_priority
144 return true unless saved_change_to_priority?
145 act_as_system_user do
146 # Update the priority of child container requests to match new
147 # priority of the parent container (ignoring requests with no
148 # container assigned, because their priority doesn't matter).
150 where(requesting_container_uuid: self.uuid,
151 state: ContainerRequest::Committed).
152 where('container_uuid is not null').
153 includes(:container).
155 map(&:update_priority!)
159 # Create a new container (or find an existing one) to satisfy the
160 # given container request.
161 def self.resolve(req)
162 if req.runtime_token.nil?
163 runtime_user = if req.modified_by_user_uuid.nil?
166 User.find_by_uuid(req.modified_by_user_uuid)
168 runtime_auth_scopes = ["all"]
170 auth = ApiClientAuthorization.validate(token: req.runtime_token)
172 raise ArgumentError.new "Invalid runtime token"
174 runtime_user = User.find_by_id(auth.user_id)
175 runtime_auth_scopes = auth.scopes
177 c_attrs = act_as_user runtime_user do
179 command: req.command,
181 environment: req.environment,
182 output_path: req.output_path,
183 container_image: resolve_container_image(req.container_image),
184 mounts: resolve_mounts(req.mounts),
185 runtime_constraints: resolve_runtime_constraints(req.runtime_constraints),
186 scheduling_parameters: req.scheduling_parameters,
187 secret_mounts: req.secret_mounts,
188 runtime_token: req.runtime_token,
189 runtime_user_uuid: runtime_user.uuid,
190 runtime_auth_scopes: runtime_auth_scopes
193 act_as_system_user do
194 if req.use_existing && (reusable = find_reusable(c_attrs))
197 Container.create!(c_attrs)
202 # Return a runtime_constraints hash that complies with requested but
203 # is suitable for saving in a container record, i.e., has specific
204 # values instead of ranges.
206 # Doing this as a step separate from other resolutions, like "git
207 # revision range to commit hash", makes sense only when there is no
208 # opportunity to reuse an existing container (e.g., container reuse
209 # is not implemented yet, or we have already found that no existing
210 # containers are suitable).
211 def self.resolve_runtime_constraints(runtime_constraints)
213 runtime_constraints.each do |k, v|
220 if rc['keep_cache_ram'] == 0
221 rc['keep_cache_ram'] = Rails.configuration.Containers.DefaultKeepCacheRAM
226 # Return a mounts hash suitable for a Container, i.e., with every
227 # readonly collection UUID resolved to a PDH.
228 def self.resolve_mounts(mounts)
230 mounts.each do |k, mount|
233 if mount['kind'] != 'collection'
237 uuid = mount.delete 'uuid'
239 if mount['portable_data_hash'].nil? and !uuid.nil?
240 # PDH not supplied, try by UUID
242 readable_by(current_user).
244 select(:portable_data_hash).
247 raise ArvadosModel::UnresolvableContainerError.new "cannot mount collection #{uuid.inspect}: not found"
249 mount['portable_data_hash'] = c.portable_data_hash
255 # Return a container_image PDH suitable for a Container.
256 def self.resolve_container_image(container_image)
257 coll = Collection.for_latest_docker_image(container_image)
259 raise ArvadosModel::UnresolvableContainerError.new "docker image #{container_image.inspect} not found"
261 coll.portable_data_hash
264 def self.find_reusable(attrs)
265 log_reuse_info { "starting with #{Container.all.count} container records in database" }
266 candidates = Container.where_serialized(:command, attrs[:command], md5: true)
267 log_reuse_info(candidates) { "after filtering on command #{attrs[:command].inspect}" }
269 candidates = candidates.where('cwd = ?', attrs[:cwd])
270 log_reuse_info(candidates) { "after filtering on cwd #{attrs[:cwd].inspect}" }
272 candidates = candidates.where_serialized(:environment, attrs[:environment], md5: true)
273 log_reuse_info(candidates) { "after filtering on environment #{attrs[:environment].inspect}" }
275 candidates = candidates.where('output_path = ?', attrs[:output_path])
276 log_reuse_info(candidates) { "after filtering on output_path #{attrs[:output_path].inspect}" }
278 image = resolve_container_image(attrs[:container_image])
279 candidates = candidates.where('container_image = ?', image)
280 log_reuse_info(candidates) { "after filtering on container_image #{image.inspect} (resolved from #{attrs[:container_image].inspect})" }
282 candidates = candidates.where_serialized(:mounts, resolve_mounts(attrs[:mounts]), md5: true)
283 log_reuse_info(candidates) { "after filtering on mounts #{attrs[:mounts].inspect}" }
285 secret_mounts_md5 = Digest::MD5.hexdigest(SafeJSON.dump(self.deep_sort_hash(attrs[:secret_mounts])))
286 candidates = candidates.where('secret_mounts_md5 = ?', secret_mounts_md5)
287 log_reuse_info(candidates) { "after filtering on secret_mounts_md5 #{secret_mounts_md5.inspect}" }
289 candidates = candidates.where_serialized(:runtime_constraints, resolve_runtime_constraints(attrs[:runtime_constraints]), md5: true)
290 log_reuse_info(candidates) { "after filtering on runtime_constraints #{attrs[:runtime_constraints].inspect}" }
292 log_reuse_info { "checking for state=Complete with readable output and log..." }
294 select_readable_pdh = Collection.
295 readable_by(current_user).
296 select(:portable_data_hash).
299 usable = candidates.where(state: Complete, exit_code: 0)
300 log_reuse_info(usable) { "with state=Complete, exit_code=0" }
302 usable = usable.where("log IN (#{select_readable_pdh})")
303 log_reuse_info(usable) { "with readable log" }
305 usable = usable.where("output IN (#{select_readable_pdh})")
306 log_reuse_info(usable) { "with readable output" }
308 usable = usable.order('finished_at ASC').limit(1).first
310 log_reuse_info { "done, reusing container #{usable.uuid} with state=Complete" }
314 # Check for non-failing Running candidates and return the most likely to finish sooner.
315 log_reuse_info { "checking for state=Running..." }
316 running = candidates.where(state: Running).
317 where("(runtime_status->'error') is null").
318 order('progress desc, started_at asc').
321 log_reuse_info { "done, reusing container #{running.uuid} with state=Running" }
324 log_reuse_info { "have no containers in Running state" }
327 # Check for Locked or Queued ones and return the most likely to start first.
328 locked_or_queued = candidates.
329 where("state IN (?)", [Locked, Queued]).
330 order('state asc, priority desc, created_at asc').
333 log_reuse_info { "done, reusing container #{locked_or_queued.uuid} with state=#{locked_or_queued.state}" }
334 return locked_or_queued
336 log_reuse_info { "have no containers in Locked or Queued state" }
339 log_reuse_info { "done, no reusable container found" }
345 if self.state != Queued
346 raise LockFailedError.new("cannot lock when #{self.state}")
348 self.update_attributes!(state: Locked)
353 if state_was == Queued and state == Locked
354 if self.priority <= 0
355 raise LockFailedError.new("cannot lock when priority<=0")
357 self.lock_count = self.lock_count+1
363 if self.state != Locked
364 raise InvalidStateTransitionError.new("cannot unlock when #{self.state}")
366 self.update_attributes!(state: Queued)
371 if state_was == Locked and state == Queued
372 if self.locked_by_uuid != current_api_client_authorization.uuid
373 raise ArvadosModel::PermissionDeniedError.new("locked by a different token")
375 if self.lock_count >= Rails.configuration.Containers.MaxDispatchAttempts
376 self.state = Cancelled
377 self.runtime_status = {error: "Failed to start container. Cancelled after exceeding 'Containers.MaxDispatchAttempts' (lock_count=#{self.lock_count})"}
382 def self.readable_by(*users_list)
383 # Load optional keyword arguments, if they exist.
384 if users_list.last.is_a? Hash
385 kwargs = users_list.pop
389 if users_list.select { |u| u.is_admin }.any?
392 Container.where(ContainerRequest.readable_by(*users_list).where("containers.uuid = container_requests.container_uuid").arel.exists)
396 [Complete, Cancelled].include?(self.state)
399 def self.for_current_token
400 return if !current_api_client_authorization
401 _, _, _, container_uuid = Thread.current[:token].split('/')
402 if container_uuid.nil?
403 Container.where(auth_uuid: current_api_client_authorization.uuid).first
405 Container.where('auth_uuid=? or (uuid=? and runtime_token=?)',
406 current_api_client_authorization.uuid,
408 current_api_client_authorization.token).first
414 def fill_field_defaults
415 self.state ||= Queued
416 self.environment ||= {}
417 self.runtime_constraints ||= {}
421 self.scheduling_parameters ||= {}
424 def permission_to_create
425 current_user.andand.is_admin
428 def permission_to_destroy
429 current_user.andand.is_admin
432 def ensure_owner_uuid_is_permitted
433 # validate_change ensures owner_uuid can't be changed at all --
434 # except during create, which requires admin privileges. Checking
435 # permission here would be superfluous.
440 if self.state_changed? and self.state == Running
441 self.started_at ||= db_current_time
444 if self.state_changed? and [Complete, Cancelled].include? self.state
445 self.finished_at ||= db_current_time
449 # Check that well-known runtime status keys have desired data types
450 def validate_runtime_status
452 'error', 'errorDetail', 'warning', 'warningDetail', 'activity'
454 if self.runtime_status.andand.include?(k) && !self.runtime_status[k].is_a?(String)
455 errors.add(:runtime_status, "'#{k}' value must be a string")
462 progress_attrs = [:progress, :runtime_status, :log, :output]
463 final_attrs = [:exit_code, :finished_at]
466 permitted.push(:owner_uuid, :command, :container_image, :cwd,
467 :environment, :mounts, :output_path, :priority,
468 :runtime_constraints, :scheduling_parameters,
469 :secret_mounts, :runtime_token,
470 :runtime_user_uuid, :runtime_auth_scopes)
475 permitted.push :priority, :runtime_status, :log, :lock_count
478 permitted.push :priority
481 permitted.push :priority, *progress_attrs
482 if self.state_changed?
483 permitted.push :started_at, :gateway_address
485 if !self.interactive_session_started_was
486 permitted.push :interactive_session_started
490 if self.state_was == Running
491 permitted.push *final_attrs, *progress_attrs
497 permitted.push :finished_at, *progress_attrs
499 permitted.push :finished_at, :log, :runtime_status
503 # The state_transitions check will add an error message for this
507 if self.state_was == Running &&
508 !current_api_client_authorization.nil? &&
509 (current_api_client_authorization.uuid == self.auth_uuid ||
510 current_api_client_authorization.token == self.runtime_token)
511 # The contained process itself can write final attrs but can't
512 # change priority or log.
513 permitted.push *final_attrs
514 permitted = permitted - [:log, :priority]
515 elsif !current_user.andand.is_admin
516 raise PermissionDeniedError
517 elsif self.locked_by_uuid && self.locked_by_uuid != current_api_client_authorization.andand.uuid
518 # When locked, progress fields cannot be updated by the wrong
519 # dispatcher, even though it has admin privileges.
520 permitted = permitted - progress_attrs
522 check_update_whitelist permitted
526 if [Locked, Running].include? self.state
527 # If the Container was already locked, locked_by_uuid must not
528 # changes. Otherwise, the current auth gets the lock.
529 need_lock = locked_by_uuid_was || current_api_client_authorization.andand.uuid
534 # The caller can provide a new value for locked_by_uuid, but only
535 # if it's exactly what we expect. This allows a caller to perform
536 # an update like {"state":"Unlocked","locked_by_uuid":null}.
537 if self.locked_by_uuid_changed?
538 if self.locked_by_uuid != need_lock
539 return errors.add :locked_by_uuid, "can only change to #{need_lock}"
542 self.locked_by_uuid = need_lock
546 # Output must exist and be readable by the current user. This is so
547 # that a container cannot "claim" a collection that it doesn't otherwise
548 # have access to just by setting the output field to the collection PDH.
551 readable_by(current_user, {include_trash: true}).
552 where(portable_data_hash: self.output).
555 errors.add :output, "collection must exist and be readable by current user."
561 # If self.final?, this update is superfluous: the final log/output
562 # update will be done when handle_completed calls finalize! on
563 # each requesting CR.
564 return if self.final? || !saved_change_to_log?
565 leave_modified_by_user_alone do
566 ContainerRequest.where(container_uuid: self.uuid).each do |cr|
567 cr.update_collections(container: self, collections: ['log'])
574 if self.auth_uuid_changed?
575 return errors.add :auth_uuid, 'is readonly'
577 if not [Locked, Running].include? self.state
578 # Don't need one. If auth already exists, expire it.
580 # We use db_transaction_time here (not db_current_time) to
581 # ensure the token doesn't validate later in the same
582 # transaction (e.g., in a test case) by satisfying expires_at >
583 # transaction timestamp.
584 self.auth.andand.update_attributes(expires_at: db_transaction_time)
591 if self.runtime_token.nil?
592 if self.runtime_user_uuid.nil?
593 # legacy behavior, we don't have a runtime_user_uuid so get
594 # the user from the highest priority container request, needed
595 # when performing an upgrade and there are queued containers,
597 cr = ContainerRequest.
598 where('container_uuid=? and priority>0', self.uuid).
599 order('priority desc').
602 return errors.add :auth_uuid, "cannot be assigned because priority <= 0"
604 self.runtime_user_uuid = cr.modified_by_user_uuid
605 self.runtime_auth_scopes = ["all"]
608 # Generate a new token. This runs with admin credentials as it's done by a
609 # dispatcher user, so expires_at isn't enforced by API.MaxTokenLifetime.
610 self.auth = ApiClientAuthorization.
611 create!(user_id: User.find_by_uuid(self.runtime_user_uuid).id,
613 scopes: self.runtime_auth_scopes)
617 def sort_serialized_attrs
618 if self.environment_changed?
619 self.environment = self.class.deep_sort_hash(self.environment)
621 if self.mounts_changed?
622 self.mounts = self.class.deep_sort_hash(self.mounts)
624 if self.runtime_constraints_changed?
625 self.runtime_constraints = self.class.deep_sort_hash(self.runtime_constraints)
627 if self.scheduling_parameters_changed?
628 self.scheduling_parameters = self.class.deep_sort_hash(self.scheduling_parameters)
630 if self.runtime_auth_scopes_changed?
631 self.runtime_auth_scopes = self.runtime_auth_scopes.sort
635 def update_secret_mounts_md5
636 if self.secret_mounts_changed?
637 self.secret_mounts_md5 = Digest::MD5.hexdigest(
638 SafeJSON.dump(self.class.deep_sort_hash(self.secret_mounts)))
643 # this runs after update_secret_mounts_md5, so the
644 # secret_mounts_md5 will still reflect the secrets that are being
646 if self.state_changed? && self.final?
647 self.secret_mounts = {}
648 self.runtime_token = nil
652 def clear_runtime_status_when_queued
653 # Avoid leaking status messages between different dispatch attempts
654 if self.state_was == Locked && self.state == Queued
655 self.runtime_status = {}
660 # This container is finished so finalize any associated container requests
661 # that are associated with this container.
662 if saved_change_to_state? and self.final?
663 # These get wiped out by with_lock (which reloads the record),
664 # so record them now in case we need to schedule a retry.
665 prev_secret_mounts = secret_mounts_before_last_save
666 prev_runtime_token = runtime_token_before_last_save
668 # Need to take a lock on the container to ensure that any
669 # concurrent container requests that might try to reuse this
670 # container will block until the container completion
671 # transaction finishes. This ensure that concurrent container
672 # requests that try to reuse this container are finalized (on
673 # Complete) or don't reuse it (on Cancelled).
675 act_as_system_user do
676 if self.state == Cancelled
677 retryable_requests = ContainerRequest.where("container_uuid = ? and priority > 0 and state = 'Committed' and container_count < container_count_max", uuid)
679 retryable_requests = []
682 if retryable_requests.any?
684 command: self.command,
686 environment: self.environment,
687 output_path: self.output_path,
688 container_image: self.container_image,
690 runtime_constraints: self.runtime_constraints,
691 scheduling_parameters: self.scheduling_parameters,
692 secret_mounts: prev_secret_mounts,
693 runtime_token: prev_runtime_token,
694 runtime_user_uuid: self.runtime_user_uuid,
695 runtime_auth_scopes: self.runtime_auth_scopes
697 c = Container.create! c_attrs
698 retryable_requests.each do |cr|
700 leave_modified_by_user_alone do
701 # Use row locking because this increments container_count
702 cr.container_uuid = c.uuid
709 # Notify container requests associated with this container
710 ContainerRequest.where(container_uuid: uuid,
711 state: ContainerRequest::Committed).each do |cr|
712 leave_modified_by_user_alone do
717 # Cancel outstanding container requests made by this container.
719 includes(:container).
720 where(requesting_container_uuid: uuid,
721 state: ContainerRequest::Committed).each do |cr|
722 leave_modified_by_user_alone do
723 cr.update_attributes!(priority: 0)
725 if cr.container.state == Container::Queued || cr.container.state == Container::Locked
726 # If the child container hasn't started yet, finalize the
727 # child CR now instead of leaving it "on hold", i.e.,
728 # Queued with priority 0. (OTOH, if the child is already
729 # running, leave it alone so it can get cancelled the
730 # usual way, get a copy of the log collection, etc.)
731 cr.update_attributes!(state: ContainerRequest::Final)