9945: Merge branch 'master' into 9945-make-python-package-dependency-free
[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       uuid = mount.delete 'uuid'
228
229       if mount['portable_data_hash'].nil? and !uuid.nil?
230         # PDH not supplied, try by UUID
231         c = Collection.
232           readable_by(current_user).
233           where(uuid: uuid).
234           select(:portable_data_hash).
235           first
236         if !c
237           raise ArvadosModel::UnresolvableContainerError.new "cannot mount collection #{uuid.inspect}: not found"
238         end
239         mount['portable_data_hash'] = c.portable_data_hash
240       end
241     end
242     return c_mounts
243   end
244
245   # Return a container_image PDH suitable for a Container.
246   def self.resolve_container_image(container_image)
247     coll = Collection.for_latest_docker_image(container_image)
248     if !coll
249       raise ArvadosModel::UnresolvableContainerError.new "docker image #{container_image.inspect} not found"
250     end
251     coll.portable_data_hash
252   end
253
254   def self.find_reusable(attrs)
255     log_reuse_info { "starting with #{Container.all.count} container records in database" }
256     candidates = Container.where_serialized(:command, attrs[:command], md5: true)
257     log_reuse_info(candidates) { "after filtering on command #{attrs[:command].inspect}" }
258
259     candidates = candidates.where('cwd = ?', attrs[:cwd])
260     log_reuse_info(candidates) { "after filtering on cwd #{attrs[:cwd].inspect}" }
261
262     candidates = candidates.where_serialized(:environment, attrs[:environment], md5: true)
263     log_reuse_info(candidates) { "after filtering on environment #{attrs[:environment].inspect}" }
264
265     candidates = candidates.where('output_path = ?', attrs[:output_path])
266     log_reuse_info(candidates) { "after filtering on output_path #{attrs[:output_path].inspect}" }
267
268     image = resolve_container_image(attrs[:container_image])
269     candidates = candidates.where('container_image = ?', image)
270     log_reuse_info(candidates) { "after filtering on container_image #{image.inspect} (resolved from #{attrs[:container_image].inspect})" }
271
272     candidates = candidates.where_serialized(:mounts, resolve_mounts(attrs[:mounts]), md5: true)
273     log_reuse_info(candidates) { "after filtering on mounts #{attrs[:mounts].inspect}" }
274
275     secret_mounts_md5 = Digest::MD5.hexdigest(SafeJSON.dump(self.deep_sort_hash(attrs[:secret_mounts])))
276     candidates = candidates.where('secret_mounts_md5 = ?', secret_mounts_md5)
277     log_reuse_info(candidates) { "after filtering on secret_mounts_md5 #{secret_mounts_md5.inspect}" }
278
279     candidates = candidates.where_serialized(:runtime_constraints, resolve_runtime_constraints(attrs[:runtime_constraints]), md5: true)
280     log_reuse_info(candidates) { "after filtering on runtime_constraints #{attrs[:runtime_constraints].inspect}" }
281
282     log_reuse_info { "checking for state=Complete with readable output and log..." }
283
284     select_readable_pdh = Collection.
285       readable_by(current_user).
286       select(:portable_data_hash).
287       to_sql
288
289     usable = candidates.where(state: Complete, exit_code: 0)
290     log_reuse_info(usable) { "with state=Complete, exit_code=0" }
291
292     usable = usable.where("log IN (#{select_readable_pdh})")
293     log_reuse_info(usable) { "with readable log" }
294
295     usable = usable.where("output IN (#{select_readable_pdh})")
296     log_reuse_info(usable) { "with readable output" }
297
298     usable = usable.order('finished_at ASC').limit(1).first
299     if usable
300       log_reuse_info { "done, reusing container #{usable.uuid} with state=Complete" }
301       return usable
302     end
303
304     # Check for non-failing Running candidates and return the most likely to finish sooner.
305     log_reuse_info { "checking for state=Running..." }
306     running = candidates.where(state: Running).
307               where("(runtime_status->'error') is null").
308               order('progress desc, started_at asc').
309               limit(1).first
310     if running
311       log_reuse_info { "done, reusing container #{running.uuid} with state=Running" }
312       return running
313     else
314       log_reuse_info { "have no containers in Running state" }
315     end
316
317     # Check for Locked or Queued ones and return the most likely to start first.
318     locked_or_queued = candidates.
319                        where("state IN (?)", [Locked, Queued]).
320                        order('state asc, priority desc, created_at asc').
321                        limit(1).first
322     if locked_or_queued
323       log_reuse_info { "done, reusing container #{locked_or_queued.uuid} with state=#{locked_or_queued.state}" }
324       return locked_or_queued
325     else
326       log_reuse_info { "have no containers in Locked or Queued state" }
327     end
328
329     log_reuse_info { "done, no reusable container found" }
330     nil
331   end
332
333   def check_lock_fail
334     if self.state != Queued
335       raise LockFailedError.new("cannot lock when #{self.state}")
336     elsif self.priority <= 0
337       raise LockFailedError.new("cannot lock when priority<=0")
338     end
339   end
340
341   def lock
342     # Check invalid state transitions once before getting the lock
343     # (because it's cheaper that way) and once after getting the lock
344     # (because state might have changed while acquiring the lock).
345     check_lock_fail
346     transaction do
347       reload
348       check_lock_fail
349       update_attributes!(state: Locked)
350     end
351   end
352
353   def check_unlock_fail
354     if self.state != Locked
355       raise InvalidStateTransitionError.new("cannot unlock when #{self.state}")
356     elsif self.locked_by_uuid != current_api_client_authorization.uuid
357       raise InvalidStateTransitionError.new("locked by a different token")
358     end
359   end
360
361   def unlock
362     # Check invalid state transitions twice (see lock)
363     check_unlock_fail
364     transaction do
365       reload(lock: 'FOR UPDATE')
366       check_unlock_fail
367       update_attributes!(state: Queued)
368     end
369   end
370
371   def self.readable_by(*users_list)
372     # Load optional keyword arguments, if they exist.
373     if users_list.last.is_a? Hash
374       kwargs = users_list.pop
375     else
376       kwargs = {}
377     end
378     Container.where(ContainerRequest.readable_by(*users_list).where("containers.uuid = container_requests.container_uuid").exists)
379   end
380
381   def final?
382     [Complete, Cancelled].include?(self.state)
383   end
384
385   def self.for_current_token
386     return if !current_api_client_authorization
387     _, _, _, container_uuid = Thread.current[:token].split('/')
388     if container_uuid.nil?
389       Container.where(auth_uuid: current_api_client_authorization.uuid).first
390     else
391       Container.where('auth_uuid=? or (uuid=? and runtime_token=?)',
392                       current_api_client_authorization.uuid,
393                       container_uuid,
394                       current_api_client_authorization.token).first
395     end
396   end
397
398   protected
399
400   def fill_field_defaults
401     self.state ||= Queued
402     self.environment ||= {}
403     self.runtime_constraints ||= {}
404     self.mounts ||= {}
405     self.cwd ||= "."
406     self.priority ||= 0
407     self.scheduling_parameters ||= {}
408   end
409
410   def permission_to_create
411     current_user.andand.is_admin
412   end
413
414   def ensure_owner_uuid_is_permitted
415     # validate_change ensures owner_uuid can't be changed at all --
416     # except during create, which requires admin privileges. Checking
417     # permission here would be superfluous.
418     true
419   end
420
421   def set_timestamps
422     if self.state_changed? and self.state == Running
423       self.started_at ||= db_current_time
424     end
425
426     if self.state_changed? and [Complete, Cancelled].include? self.state
427       self.finished_at ||= db_current_time
428     end
429   end
430
431   # Check that well-known runtime status keys have desired data types
432   def validate_runtime_status
433     [
434       'error', 'errorDetail', 'warning', 'warningDetail', 'activity'
435     ].each do |k|
436       if self.runtime_status.andand.include?(k) && !self.runtime_status[k].is_a?(String)
437         errors.add(:runtime_status, "'#{k}' value must be a string")
438       end
439     end
440   end
441
442   def validate_change
443     permitted = [:state]
444     progress_attrs = [:progress, :runtime_status, :log, :output]
445     final_attrs = [:exit_code, :finished_at]
446
447     if self.new_record?
448       permitted.push(:owner_uuid, :command, :container_image, :cwd,
449                      :environment, :mounts, :output_path, :priority,
450                      :runtime_constraints, :scheduling_parameters,
451                      :secret_mounts, :runtime_token,
452                      :runtime_user_uuid, :runtime_auth_scopes)
453     end
454
455     case self.state
456     when Locked
457       permitted.push :priority, :runtime_status, :log
458
459     when Queued
460       permitted.push :priority
461
462     when Running
463       permitted.push :priority, *progress_attrs
464       if self.state_changed?
465         permitted.push :started_at
466       end
467
468     when Complete
469       if self.state_was == Running
470         permitted.push *final_attrs, *progress_attrs
471       end
472
473     when Cancelled
474       case self.state_was
475       when Running
476         permitted.push :finished_at, *progress_attrs
477       when Queued, Locked
478         permitted.push :finished_at, :log
479       end
480
481     else
482       # The state_transitions check will add an error message for this
483       return false
484     end
485
486     if self.state == Running &&
487        !current_api_client_authorization.nil? &&
488        (current_api_client_authorization.uuid == self.auth_uuid ||
489         current_api_client_authorization.token == self.runtime_token)
490       # The contained process itself can write final attrs but can't
491       # change priority or log.
492       permitted.push *final_attrs
493       permitted = permitted - [:log, :priority]
494     elsif self.locked_by_uuid && self.locked_by_uuid != current_api_client_authorization.andand.uuid
495       # When locked, progress fields cannot be updated by the wrong
496       # dispatcher, even though it has admin privileges.
497       permitted = permitted - progress_attrs
498     end
499     check_update_whitelist permitted
500   end
501
502   def validate_lock
503     if [Locked, Running].include? self.state
504       # If the Container was already locked, locked_by_uuid must not
505       # changes. Otherwise, the current auth gets the lock.
506       need_lock = locked_by_uuid_was || current_api_client_authorization.andand.uuid
507     else
508       need_lock = nil
509     end
510
511     # The caller can provide a new value for locked_by_uuid, but only
512     # if it's exactly what we expect. This allows a caller to perform
513     # an update like {"state":"Unlocked","locked_by_uuid":null}.
514     if self.locked_by_uuid_changed?
515       if self.locked_by_uuid != need_lock
516         return errors.add :locked_by_uuid, "can only change to #{need_lock}"
517       end
518     end
519     self.locked_by_uuid = need_lock
520   end
521
522   def validate_output
523     # Output must exist and be readable by the current user.  This is so
524     # that a container cannot "claim" a collection that it doesn't otherwise
525     # have access to just by setting the output field to the collection PDH.
526     if output_changed?
527       c = Collection.
528             readable_by(current_user, {include_trash: true}).
529             where(portable_data_hash: self.output).
530             first
531       if !c
532         errors.add :output, "collection must exist and be readable by current user."
533       end
534     end
535   end
536
537   def update_cr_logs
538     # If self.final?, this update is superfluous: the final log/output
539     # update will be done when handle_completed calls finalize! on
540     # each requesting CR.
541     return if self.final? || !self.log_changed?
542     leave_modified_by_user_alone do
543       ContainerRequest.where(container_uuid: self.uuid).each do |cr|
544         cr.update_collections(container: self, collections: ['log'])
545         cr.save!
546       end
547     end
548   end
549
550   def assign_auth
551     if self.auth_uuid_changed?
552          return errors.add :auth_uuid, 'is readonly'
553     end
554     if not [Locked, Running].include? self.state
555       # don't need one
556       self.auth.andand.update_attributes(expires_at: db_current_time)
557       self.auth = nil
558       return
559     elsif self.auth
560       # already have one
561       return
562     end
563     if self.runtime_token.nil?
564       if self.runtime_user_uuid.nil?
565         # legacy behavior, we don't have a runtime_user_uuid so get
566         # the user from the highest priority container request, needed
567         # when performing an upgrade and there are queued containers,
568         # and some tests.
569         cr = ContainerRequest.
570                where('container_uuid=? and priority>0', self.uuid).
571                order('priority desc').
572                first
573         if !cr
574           return errors.add :auth_uuid, "cannot be assigned because priority <= 0"
575         end
576         self.runtime_user_uuid = cr.modified_by_user_uuid
577         self.runtime_auth_scopes = ["all"]
578       end
579
580       # generate a new token
581       self.auth = ApiClientAuthorization.
582                     create!(user_id: User.find_by_uuid(self.runtime_user_uuid).id,
583                             api_client_id: 0,
584                             scopes: self.runtime_auth_scopes)
585     end
586   end
587
588   def sort_serialized_attrs
589     if self.environment_changed?
590       self.environment = self.class.deep_sort_hash(self.environment)
591     end
592     if self.mounts_changed?
593       self.mounts = self.class.deep_sort_hash(self.mounts)
594     end
595     if self.runtime_constraints_changed?
596       self.runtime_constraints = self.class.deep_sort_hash(self.runtime_constraints)
597     end
598     if self.scheduling_parameters_changed?
599       self.scheduling_parameters = self.class.deep_sort_hash(self.scheduling_parameters)
600     end
601     if self.runtime_auth_scopes_changed?
602       self.runtime_auth_scopes = self.runtime_auth_scopes.sort
603     end
604   end
605
606   def update_secret_mounts_md5
607     if self.secret_mounts_changed?
608       self.secret_mounts_md5 = Digest::MD5.hexdigest(
609         SafeJSON.dump(self.class.deep_sort_hash(self.secret_mounts)))
610     end
611   end
612
613   def scrub_secrets
614     # this runs after update_secret_mounts_md5, so the
615     # secret_mounts_md5 will still reflect the secrets that are being
616     # scrubbed here.
617     if self.state_changed? && self.final?
618       self.secret_mounts = {}
619       self.runtime_token = nil
620     end
621   end
622
623   def clear_runtime_status_when_queued
624     # Avoid leaking status messages between different dispatch attempts
625     if self.state_was == Locked && self.state == Queued
626       self.runtime_status = {}
627     end
628   end
629
630   def handle_completed
631     # This container is finished so finalize any associated container requests
632     # that are associated with this container.
633     if self.state_changed? and self.final?
634       act_as_system_user do
635
636         if self.state == Cancelled
637           retryable_requests = ContainerRequest.where("container_uuid = ? and priority > 0 and state = 'Committed' and container_count < container_count_max", uuid)
638         else
639           retryable_requests = []
640         end
641
642         if retryable_requests.any?
643           c_attrs = {
644             command: self.command,
645             cwd: self.cwd,
646             environment: self.environment,
647             output_path: self.output_path,
648             container_image: self.container_image,
649             mounts: self.mounts,
650             runtime_constraints: self.runtime_constraints,
651             scheduling_parameters: self.scheduling_parameters,
652             secret_mounts: self.secret_mounts_was,
653             runtime_token: self.runtime_token_was,
654             runtime_user_uuid: self.runtime_user_uuid,
655             runtime_auth_scopes: self.runtime_auth_scopes
656           }
657           c = Container.create! c_attrs
658           retryable_requests.each do |cr|
659             cr.with_lock do
660               leave_modified_by_user_alone do
661                 # Use row locking because this increments container_count
662                 cr.container_uuid = c.uuid
663                 cr.save!
664               end
665             end
666           end
667         end
668
669         # Notify container requests associated with this container
670         ContainerRequest.where(container_uuid: uuid,
671                                state: ContainerRequest::Committed).each do |cr|
672           leave_modified_by_user_alone do
673             cr.finalize!
674           end
675         end
676
677         # Cancel outstanding container requests made by this container.
678         ContainerRequest.
679           includes(:container).
680           where(requesting_container_uuid: uuid,
681                 state: ContainerRequest::Committed).each do |cr|
682           leave_modified_by_user_alone do
683             cr.update_attributes!(priority: 0)
684             cr.container.reload
685             if cr.container.state == Container::Queued || cr.container.state == Container::Locked
686               # If the child container hasn't started yet, finalize the
687               # child CR now instead of leaving it "on hold", i.e.,
688               # Queued with priority 0.  (OTOH, if the child is already
689               # running, leave it alone so it can get cancelled the
690               # usual way, get a copy of the log collection, etc.)
691               cr.update_attributes!(state: ContainerRequest::Final)
692             end
693           end
694         end
695       end
696     end
697   end
698 end