14323: Prioritize PDH over UUID for mount resolving when both provided.
[arvados.git] / services / api / app / models / container.rb
1 # Copyright (C) The Arvados Authors. All rights reserved.
2 #
3 # SPDX-License-Identifier: AGPL-3.0
4
5 require 'log_reuse_info'
6 require 'whitelist_update'
7 require 'safe_json'
8 require 'update_priority'
9
10 class Container < ArvadosModel
11   include ArvadosModelUpdates
12   include HasUuid
13   include KindAndEtag
14   include CommonApiTemplate
15   include WhitelistUpdate
16   extend CurrentApiClient
17   extend DbCurrentTime
18   extend LogReuseInfo
19
20   serialize :environment, Hash
21   serialize :mounts, Hash
22   serialize :runtime_constraints, Hash
23   serialize :command, Array
24   serialize :scheduling_parameters, Hash
25   serialize :secret_mounts, Hash
26   serialize :runtime_status, Hash
27
28   before_validation :fill_field_defaults, :if => :new_record?
29   before_validation :set_timestamps
30   validates :command, :container_image, :output_path, :cwd, :priority, { presence: true }
31   validates :priority, numericality: { only_integer: true, greater_than_or_equal_to: 0 }
32   validate :validate_runtime_status
33   validate :validate_state_change
34   validate :validate_change
35   validate :validate_lock
36   validate :validate_output
37   after_validation :assign_auth
38   before_save :sort_serialized_attrs
39   before_save :update_secret_mounts_md5
40   before_save :scrub_secrets
41   before_save :clear_runtime_status_when_queued
42   after_save :update_cr_logs
43   after_save :handle_completed
44   after_save :propagate_priority
45   after_commit { UpdatePriority.run_update_thread }
46
47   has_many :container_requests, :foreign_key => :container_uuid, :class_name => 'ContainerRequest', :primary_key => :uuid
48   belongs_to :auth, :class_name => 'ApiClientAuthorization', :foreign_key => :auth_uuid, :primary_key => :uuid
49
50   api_accessible :user, extend: :common do |t|
51     t.add :command
52     t.add :container_image
53     t.add :cwd
54     t.add :environment
55     t.add :exit_code
56     t.add :finished_at
57     t.add :locked_by_uuid
58     t.add :log
59     t.add :mounts
60     t.add :output
61     t.add :output_path
62     t.add :priority
63     t.add :progress
64     t.add :runtime_constraints
65     t.add :runtime_status
66     t.add :started_at
67     t.add :state
68     t.add :auth_uuid
69     t.add :scheduling_parameters
70     t.add :runtime_user_uuid
71     t.add :runtime_auth_scopes
72   end
73
74   # Supported states for a container
75   States =
76     [
77      (Queued = 'Queued'),
78      (Locked = 'Locked'),
79      (Running = 'Running'),
80      (Complete = 'Complete'),
81      (Cancelled = 'Cancelled')
82     ]
83
84   State_transitions = {
85     nil => [Queued],
86     Queued => [Locked, Cancelled],
87     Locked => [Queued, Running, Cancelled],
88     Running => [Complete, Cancelled]
89   }
90
91   def self.limit_index_columns_read
92     ["mounts"]
93   end
94
95   def self.full_text_searchable_columns
96     super - ["secret_mounts", "secret_mounts_md5", "runtime_token"]
97   end
98
99   def self.searchable_columns *args
100     super - ["secret_mounts_md5", "runtime_token"]
101   end
102
103   def logged_attributes
104     super.except('secret_mounts', 'runtime_token')
105   end
106
107   def state_transitions
108     State_transitions
109   end
110
111   # Container priority is the highest "computed priority" of any
112   # matching request. The computed priority of a container-submitted
113   # request is the priority of the submitting container. The computed
114   # priority of a user-submitted request is a function of
115   # user-assigned priority and request creation time.
116   def update_priority!
117     return if ![Queued, Locked, Running].include?(state)
118     p = ContainerRequest.
119         where('container_uuid=? and priority>0', uuid).
120         includes(:requesting_container).
121         lock(true).
122         map do |cr|
123       if cr.requesting_container
124         cr.requesting_container.priority
125       else
126         (cr.priority << 50) - (cr.created_at.to_time.to_f * 1000).to_i
127       end
128     end.max || 0
129     update_attributes!(priority: p)
130   end
131
132   def propagate_priority
133     return true unless priority_changed?
134     act_as_system_user do
135       # Update the priority of child container requests to match new
136       # priority of the parent container (ignoring requests with no
137       # container assigned, because their priority doesn't matter).
138       ContainerRequest.
139         where(requesting_container_uuid: self.uuid,
140               state: ContainerRequest::Committed).
141         where('container_uuid is not null').
142         includes(:container).
143         map(&:container).
144         map(&:update_priority!)
145     end
146   end
147
148   # Create a new container (or find an existing one) to satisfy the
149   # given container request.
150   def self.resolve(req)
151     if req.runtime_token.nil?
152       runtime_user = if req.modified_by_user_uuid.nil?
153                        current_user
154                      else
155                        User.find_by_uuid(req.modified_by_user_uuid)
156                      end
157       runtime_auth_scopes = ["all"]
158     else
159       auth = ApiClientAuthorization.validate(token: req.runtime_token)
160       if auth.nil?
161         raise ArgumentError.new "Invalid runtime token"
162       end
163       runtime_user = User.find_by_id(auth.user_id)
164       runtime_auth_scopes = auth.scopes
165     end
166     c_attrs = act_as_user runtime_user do
167       {
168         command: req.command,
169         cwd: req.cwd,
170         environment: req.environment,
171         output_path: req.output_path,
172         container_image: resolve_container_image(req.container_image),
173         mounts: resolve_mounts(req.mounts),
174         runtime_constraints: resolve_runtime_constraints(req.runtime_constraints),
175         scheduling_parameters: req.scheduling_parameters,
176         secret_mounts: req.secret_mounts,
177         runtime_token: req.runtime_token,
178         runtime_user_uuid: runtime_user.uuid,
179         runtime_auth_scopes: runtime_auth_scopes
180       }
181     end
182     act_as_system_user do
183       if req.use_existing && (reusable = find_reusable(c_attrs))
184         reusable
185       else
186         Container.create!(c_attrs)
187       end
188     end
189   end
190
191   # Return a runtime_constraints hash that complies with requested but
192   # is suitable for saving in a container record, i.e., has specific
193   # values instead of ranges.
194   #
195   # Doing this as a step separate from other resolutions, like "git
196   # revision range to commit hash", makes sense only when there is no
197   # opportunity to reuse an existing container (e.g., container reuse
198   # is not implemented yet, or we have already found that no existing
199   # containers are suitable).
200   def self.resolve_runtime_constraints(runtime_constraints)
201     rc = {}
202     defaults = {
203       'keep_cache_ram' =>
204       Rails.configuration.container_default_keep_cache_ram,
205     }
206     defaults.merge(runtime_constraints).each do |k, v|
207       if v.is_a? Array
208         rc[k] = v[0]
209       else
210         rc[k] = v
211       end
212     end
213     rc
214   end
215
216   # Return a mounts hash suitable for a Container, i.e., with every
217   # readonly collection UUID resolved to a PDH.
218   def self.resolve_mounts(mounts)
219     c_mounts = {}
220     mounts.each do |k, mount|
221       mount = mount.dup
222       c_mounts[k] = mount
223       if mount['kind'] != 'collection'
224         next
225       end
226
227       pdh = mount['portable_data_hash']
228       uuid = mount['uuid']
229
230       # Prioritize PDH over UUID for mount selection
231       if not pdh.nil?
232         c = Collection.
233           readable_by(current_user).
234           where(portable_data_hash: pdh)
235         if c.count > 0
236           if uuid && c.where(uuid: uuid).count == 0
237             raise ArgumentError.new "cannot mount collection #{uuid.inspect}: current portable_data_hash #{pdh.inspect} does not match #{c.first.portable_data_hash.inspect} in request"
238           end
239           # mount.delete 'uuid'
240           next
241         end
242         raise ArgumentError.new "cannot mount collection #{pdh.inspect}: not found"
243       end
244
245       # PDH not supplied, try by UUID
246       if uuid
247         c = Collection.
248           readable_by(current_user).
249           where(uuid: uuid).
250           select(:portable_data_hash).
251           first
252         if !c
253           raise ArvadosModel::UnresolvableContainerError.new "cannot mount collection #{uuid.inspect}: not found"
254         end
255         mount['portable_data_hash'] = c.portable_data_hash
256       end
257     end
258     return c_mounts
259   end
260
261   # Return a container_image PDH suitable for a Container.
262   def self.resolve_container_image(container_image)
263     coll = Collection.for_latest_docker_image(container_image)
264     if !coll
265       raise ArvadosModel::UnresolvableContainerError.new "docker image #{container_image.inspect} not found"
266     end
267     coll.portable_data_hash
268   end
269
270   def self.find_reusable(attrs)
271     log_reuse_info { "starting with #{Container.all.count} container records in database" }
272     candidates = Container.where_serialized(:command, attrs[:command], md5: true)
273     log_reuse_info(candidates) { "after filtering on command #{attrs[:command].inspect}" }
274
275     candidates = candidates.where('cwd = ?', attrs[:cwd])
276     log_reuse_info(candidates) { "after filtering on cwd #{attrs[:cwd].inspect}" }
277
278     candidates = candidates.where_serialized(:environment, attrs[:environment], md5: true)
279     log_reuse_info(candidates) { "after filtering on environment #{attrs[:environment].inspect}" }
280
281     candidates = candidates.where('output_path = ?', attrs[:output_path])
282     log_reuse_info(candidates) { "after filtering on output_path #{attrs[:output_path].inspect}" }
283
284     image = resolve_container_image(attrs[:container_image])
285     candidates = candidates.where('container_image = ?', image)
286     log_reuse_info(candidates) { "after filtering on container_image #{image.inspect} (resolved from #{attrs[:container_image].inspect})" }
287
288     candidates = candidates.where_serialized(:mounts, resolve_mounts(attrs[:mounts]), md5: true)
289     log_reuse_info(candidates) { "after filtering on mounts #{attrs[:mounts].inspect}" }
290
291     secret_mounts_md5 = Digest::MD5.hexdigest(SafeJSON.dump(self.deep_sort_hash(attrs[:secret_mounts])))
292     candidates = candidates.where('secret_mounts_md5 = ?', secret_mounts_md5)
293     log_reuse_info(candidates) { "after filtering on secret_mounts_md5 #{secret_mounts_md5.inspect}" }
294
295     candidates = candidates.where_serialized(:runtime_constraints, resolve_runtime_constraints(attrs[:runtime_constraints]), md5: true)
296     log_reuse_info(candidates) { "after filtering on runtime_constraints #{attrs[:runtime_constraints].inspect}" }
297
298     candidates = candidates.where('runtime_user_uuid = ? or (runtime_user_uuid is NULL and runtime_auth_scopes is NULL)',
299                                   attrs[:runtime_user_uuid])
300     log_reuse_info(candidates) { "after filtering on runtime_user_uuid #{attrs[:runtime_user_uuid].inspect}" }
301
302     candidates = candidates.where('runtime_auth_scopes = ? or (runtime_user_uuid is NULL and runtime_auth_scopes is NULL)',
303                                   SafeJSON.dump(attrs[:runtime_auth_scopes].sort))
304     log_reuse_info(candidates) { "after filtering on runtime_auth_scopes #{attrs[:runtime_auth_scopes].inspect}" }
305
306     log_reuse_info { "checking for state=Complete with readable output and log..." }
307
308     select_readable_pdh = Collection.
309       readable_by(current_user).
310       select(:portable_data_hash).
311       to_sql
312
313     usable = candidates.where(state: Complete, exit_code: 0)
314     log_reuse_info(usable) { "with state=Complete, exit_code=0" }
315
316     usable = usable.where("log IN (#{select_readable_pdh})")
317     log_reuse_info(usable) { "with readable log" }
318
319     usable = usable.where("output IN (#{select_readable_pdh})")
320     log_reuse_info(usable) { "with readable output" }
321
322     usable = usable.order('finished_at ASC').limit(1).first
323     if usable
324       log_reuse_info { "done, reusing container #{usable.uuid} with state=Complete" }
325       return usable
326     end
327
328     # Check for non-failing Running candidates and return the most likely to finish sooner.
329     log_reuse_info { "checking for state=Running..." }
330     running = candidates.where(state: Running).
331               where("(runtime_status->'error') is null").
332               order('progress desc, started_at asc').
333               limit(1).first
334     if running
335       log_reuse_info { "done, reusing container #{running.uuid} with state=Running" }
336       return running
337     else
338       log_reuse_info { "have no containers in Running state" }
339     end
340
341     # Check for Locked or Queued ones and return the most likely to start first.
342     locked_or_queued = candidates.
343                        where("state IN (?)", [Locked, Queued]).
344                        order('state asc, priority desc, created_at asc').
345                        limit(1).first
346     if locked_or_queued
347       log_reuse_info { "done, reusing container #{locked_or_queued.uuid} with state=#{locked_or_queued.state}" }
348       return locked_or_queued
349     else
350       log_reuse_info { "have no containers in Locked or Queued state" }
351     end
352
353     log_reuse_info { "done, no reusable container found" }
354     nil
355   end
356
357   def check_lock_fail
358     if self.state != Queued
359       raise LockFailedError.new("cannot lock when #{self.state}")
360     elsif self.priority <= 0
361       raise LockFailedError.new("cannot lock when priority<=0")
362     end
363   end
364
365   def lock
366     # Check invalid state transitions once before getting the lock
367     # (because it's cheaper that way) and once after getting the lock
368     # (because state might have changed while acquiring the lock).
369     check_lock_fail
370     transaction do
371       reload
372       check_lock_fail
373       update_attributes!(state: Locked)
374     end
375   end
376
377   def check_unlock_fail
378     if self.state != Locked
379       raise InvalidStateTransitionError.new("cannot unlock when #{self.state}")
380     elsif self.locked_by_uuid != current_api_client_authorization.uuid
381       raise InvalidStateTransitionError.new("locked by a different token")
382     end
383   end
384
385   def unlock
386     # Check invalid state transitions twice (see lock)
387     check_unlock_fail
388     transaction do
389       reload(lock: 'FOR UPDATE')
390       check_unlock_fail
391       update_attributes!(state: Queued)
392     end
393   end
394
395   def self.readable_by(*users_list)
396     # Load optional keyword arguments, if they exist.
397     if users_list.last.is_a? Hash
398       kwargs = users_list.pop
399     else
400       kwargs = {}
401     end
402     Container.where(ContainerRequest.readable_by(*users_list).where("containers.uuid = container_requests.container_uuid").exists)
403   end
404
405   def final?
406     [Complete, Cancelled].include?(self.state)
407   end
408
409   def self.for_current_token
410     return if !current_api_client_authorization
411     _, _, _, container_uuid = Thread.current[:token].split('/')
412     if container_uuid.nil?
413       Container.where(auth_uuid: current_api_client_authorization.uuid).first
414     else
415       Container.where('auth_uuid=? or (uuid=? and runtime_token=?)',
416                       current_api_client_authorization.uuid,
417                       container_uuid,
418                       current_api_client_authorization.token).first
419     end
420   end
421
422   protected
423
424   def fill_field_defaults
425     self.state ||= Queued
426     self.environment ||= {}
427     self.runtime_constraints ||= {}
428     self.mounts ||= {}
429     self.cwd ||= "."
430     self.priority ||= 0
431     self.scheduling_parameters ||= {}
432   end
433
434   def permission_to_create
435     current_user.andand.is_admin
436   end
437
438   def ensure_owner_uuid_is_permitted
439     # validate_change ensures owner_uuid can't be changed at all --
440     # except during create, which requires admin privileges. Checking
441     # permission here would be superfluous.
442     true
443   end
444
445   def set_timestamps
446     if self.state_changed? and self.state == Running
447       self.started_at ||= db_current_time
448     end
449
450     if self.state_changed? and [Complete, Cancelled].include? self.state
451       self.finished_at ||= db_current_time
452     end
453   end
454
455   # Check that well-known runtime status keys have desired data types
456   def validate_runtime_status
457     [
458       'error', 'errorDetail', 'warning', 'warningDetail', 'activity'
459     ].each do |k|
460       if self.runtime_status.andand.include?(k) && !self.runtime_status[k].is_a?(String)
461         errors.add(:runtime_status, "'#{k}' value must be a string")
462       end
463     end
464   end
465
466   def validate_change
467     permitted = [:state]
468     progress_attrs = [:progress, :runtime_status, :log, :output]
469     final_attrs = [:exit_code, :finished_at]
470
471     if self.new_record?
472       permitted.push(:owner_uuid, :command, :container_image, :cwd,
473                      :environment, :mounts, :output_path, :priority,
474                      :runtime_constraints, :scheduling_parameters,
475                      :secret_mounts, :runtime_token,
476                      :runtime_user_uuid, :runtime_auth_scopes)
477     end
478
479     case self.state
480     when Locked
481       permitted.push :priority, :runtime_status, :log
482
483     when Queued
484       permitted.push :priority
485
486     when Running
487       permitted.push :priority, *progress_attrs
488       if self.state_changed?
489         permitted.push :started_at
490       end
491
492     when Complete
493       if self.state_was == Running
494         permitted.push *final_attrs, *progress_attrs
495       end
496
497     when Cancelled
498       case self.state_was
499       when Running
500         permitted.push :finished_at, *progress_attrs
501       when Queued, Locked
502         permitted.push :finished_at, :log
503       end
504
505     else
506       # The state_transitions check will add an error message for this
507       return false
508     end
509
510     if self.state == Running &&
511        !current_api_client_authorization.nil? &&
512        (current_api_client_authorization.uuid == self.auth_uuid ||
513         current_api_client_authorization.token == self.runtime_token)
514       # The contained process itself can write final attrs but can't
515       # change priority or log.
516       permitted.push *final_attrs
517       permitted = permitted - [:log, :priority]
518     elsif self.locked_by_uuid && self.locked_by_uuid != current_api_client_authorization.andand.uuid
519       # When locked, progress fields cannot be updated by the wrong
520       # dispatcher, even though it has admin privileges.
521       permitted = permitted - progress_attrs
522     end
523     check_update_whitelist permitted
524   end
525
526   def validate_lock
527     if [Locked, Running].include? self.state
528       # If the Container was already locked, locked_by_uuid must not
529       # changes. Otherwise, the current auth gets the lock.
530       need_lock = locked_by_uuid_was || current_api_client_authorization.andand.uuid
531     else
532       need_lock = nil
533     end
534
535     # The caller can provide a new value for locked_by_uuid, but only
536     # if it's exactly what we expect. This allows a caller to perform
537     # an update like {"state":"Unlocked","locked_by_uuid":null}.
538     if self.locked_by_uuid_changed?
539       if self.locked_by_uuid != need_lock
540         return errors.add :locked_by_uuid, "can only change to #{need_lock}"
541       end
542     end
543     self.locked_by_uuid = need_lock
544   end
545
546   def validate_output
547     # Output must exist and be readable by the current user.  This is so
548     # that a container cannot "claim" a collection that it doesn't otherwise
549     # have access to just by setting the output field to the collection PDH.
550     if output_changed?
551       c = Collection.
552             readable_by(current_user, {include_trash: true}).
553             where(portable_data_hash: self.output).
554             first
555       if !c
556         errors.add :output, "collection must exist and be readable by current user."
557       end
558     end
559   end
560
561   def update_cr_logs
562     # If self.final?, this update is superfluous: the final log/output
563     # update will be done when handle_completed calls finalize! on
564     # each requesting CR.
565     return if self.final? || !self.log_changed?
566     leave_modified_by_user_alone do
567       ContainerRequest.where(container_uuid: self.uuid).each do |cr|
568         cr.update_collections(container: self, collections: ['log'])
569         cr.save!
570       end
571     end
572   end
573
574   def assign_auth
575     if self.auth_uuid_changed?
576          return errors.add :auth_uuid, 'is readonly'
577     end
578     if not [Locked, Running].include? self.state
579       # don't need one
580       self.auth.andand.update_attributes(expires_at: db_current_time)
581       self.auth = nil
582       return
583     elsif self.auth
584       # already have one
585       return
586     end
587     if self.runtime_token.nil?
588       if self.runtime_user_uuid.nil?
589         # legacy behavior, we don't have a runtime_user_uuid so get
590         # the user from the highest priority container request, needed
591         # when performing an upgrade and there are queued containers,
592         # and some tests.
593         cr = ContainerRequest.
594                where('container_uuid=? and priority>0', self.uuid).
595                order('priority desc').
596                first
597         if !cr
598           return errors.add :auth_uuid, "cannot be assigned because priority <= 0"
599         end
600         self.runtime_user_uuid = cr.modified_by_user_uuid
601         self.runtime_auth_scopes = ["all"]
602       end
603
604       # generate a new token
605       self.auth = ApiClientAuthorization.
606                     create!(user_id: User.find_by_uuid(self.runtime_user_uuid).id,
607                             api_client_id: 0,
608                             scopes: self.runtime_auth_scopes)
609     end
610   end
611
612   def sort_serialized_attrs
613     if self.environment_changed?
614       self.environment = self.class.deep_sort_hash(self.environment)
615     end
616     if self.mounts_changed?
617       self.mounts = self.class.deep_sort_hash(self.mounts)
618     end
619     if self.runtime_constraints_changed?
620       self.runtime_constraints = self.class.deep_sort_hash(self.runtime_constraints)
621     end
622     if self.scheduling_parameters_changed?
623       self.scheduling_parameters = self.class.deep_sort_hash(self.scheduling_parameters)
624     end
625     if self.runtime_auth_scopes_changed?
626       self.runtime_auth_scopes = self.runtime_auth_scopes.sort
627     end
628   end
629
630   def update_secret_mounts_md5
631     if self.secret_mounts_changed?
632       self.secret_mounts_md5 = Digest::MD5.hexdigest(
633         SafeJSON.dump(self.class.deep_sort_hash(self.secret_mounts)))
634     end
635   end
636
637   def scrub_secrets
638     # this runs after update_secret_mounts_md5, so the
639     # secret_mounts_md5 will still reflect the secrets that are being
640     # scrubbed here.
641     if self.state_changed? && self.final?
642       self.secret_mounts = {}
643       self.runtime_token = nil
644     end
645   end
646
647   def clear_runtime_status_when_queued
648     # Avoid leaking status messages between different dispatch attempts
649     if self.state_was == Locked && self.state == Queued
650       self.runtime_status = {}
651     end
652   end
653
654   def handle_completed
655     # This container is finished so finalize any associated container requests
656     # that are associated with this container.
657     if self.state_changed? and self.final?
658       act_as_system_user do
659
660         if self.state == Cancelled
661           retryable_requests = ContainerRequest.where("container_uuid = ? and priority > 0 and state = 'Committed' and container_count < container_count_max", uuid)
662         else
663           retryable_requests = []
664         end
665
666         if retryable_requests.any?
667           c_attrs = {
668             command: self.command,
669             cwd: self.cwd,
670             environment: self.environment,
671             output_path: self.output_path,
672             container_image: self.container_image,
673             mounts: self.mounts,
674             runtime_constraints: self.runtime_constraints,
675             scheduling_parameters: self.scheduling_parameters,
676             secret_mounts: self.secret_mounts_was,
677             runtime_token: self.runtime_token_was,
678             runtime_user_uuid: self.runtime_user_uuid,
679             runtime_auth_scopes: self.runtime_auth_scopes
680           }
681           c = Container.create! c_attrs
682           retryable_requests.each do |cr|
683             cr.with_lock do
684               leave_modified_by_user_alone do
685                 # Use row locking because this increments container_count
686                 cr.container_uuid = c.uuid
687                 cr.save!
688               end
689             end
690           end
691         end
692
693         # Notify container requests associated with this container
694         ContainerRequest.where(container_uuid: uuid,
695                                state: ContainerRequest::Committed).each do |cr|
696           leave_modified_by_user_alone do
697             cr.finalize!
698           end
699         end
700
701         # Cancel outstanding container requests made by this container.
702         ContainerRequest.
703           includes(:container).
704           where(requesting_container_uuid: uuid,
705                 state: ContainerRequest::Committed).each do |cr|
706           leave_modified_by_user_alone do
707             cr.update_attributes!(priority: 0)
708             cr.container.reload
709             if cr.container.state == Container::Queued || cr.container.state == Container::Locked
710               # If the child container hasn't started yet, finalize the
711               # child CR now instead of leaving it "on hold", i.e.,
712               # Queued with priority 0.  (OTOH, if the child is already
713               # running, leave it alone so it can get cancelled the
714               # usual way, get a copy of the log collection, etc.)
715               cr.update_attributes!(state: ContainerRequest::Final)
716             end
717           end
718         end
719       end
720     end
721   end
722 end