require 'log_reuse_info'
require 'whitelist_update'
require 'safe_json'
-require 'update_priority'
+require 'update_priorities'
class Container < ArvadosModel
include ArvadosModelUpdates
before_save :clear_runtime_status_when_queued
after_save :update_cr_logs
after_save :handle_completed
- after_save :propagate_priority
- after_commit { UpdatePriority.run_update_thread }
has_many :container_requests, :foreign_key => :container_uuid, :class_name => 'ContainerRequest', :primary_key => :uuid
belongs_to :auth, :class_name => 'ApiClientAuthorization', :foreign_key => :auth_uuid, :primary_key => :uuid
# priority of a user-submitted request is a function of
# user-assigned priority and request creation time.
def update_priority!
- return if ![Queued, Locked, Running].include?(state)
- p = ContainerRequest.
- where('container_uuid=? and priority>0', uuid).
- includes(:requesting_container).
- lock(true).
- map do |cr|
- if cr.requesting_container
- cr.requesting_container.priority
- else
- (cr.priority << 50) - (cr.created_at.to_time.to_f * 1000).to_i
- end
- end.max || 0
- update_attributes!(priority: p)
- end
-
- def propagate_priority
- return true unless saved_change_to_priority?
- act_as_system_user do
- # Update the priority of child container requests to match new
- # priority of the parent container (ignoring requests with no
- # container assigned, because their priority doesn't matter).
- ContainerRequest.
- where(requesting_container_uuid: self.uuid,
- state: ContainerRequest::Committed).
- where('container_uuid is not null').
- includes(:container).
- map(&:container).
- map(&:update_priority!)
- end
+ update_priorities uuid
+ reload
end
# Create a new container (or find an existing one) to satisfy the
# Check for non-failing Running candidates and return the most likely to finish sooner.
log_reuse_info { "checking for state=Running..." }
running = candidates.where(state: Running).
- where("(runtime_status->'error') is null").
+ where("(runtime_status->'error') is null and priority > 0").
order('progress desc, started_at asc').
limit(1).first
if running
locked_or_queued = candidates.
where("state IN (?)", [Locked, Queued]).
order('state asc, priority desc, created_at asc').
- limit(1).first
- if locked_or_queued
- log_reuse_info { "done, reusing container #{locked_or_queued.uuid} with state=#{locked_or_queued.state}" }
- return locked_or_queued
+ limit(1)
+ if !attrs[:scheduling_parameters]['preemptible']
+ locked_or_queued = locked_or_queued.
+ where("not ((scheduling_parameters::jsonb)->>'preemptible')::boolean")
+ end
+ chosen = locked_or_queued.first
+ if chosen
+ log_reuse_info { "done, reusing container #{chosen.uuid} with state=#{chosen.state}" }
+ return chosen
else
log_reuse_info { "have no containers in Locked or Queued state" }
end
# each requesting CR.
return if self.final? || !saved_change_to_log?
leave_modified_by_user_alone do
- ContainerRequest.where(container_uuid: self.uuid).each do |cr|
+ ContainerRequest.where(container_uuid: self.uuid, state: ContainerRequest::Committed).each do |cr|
cr.update_collections(container: self, collections: ['log'])
cr.save!
end
self.with_lock do
act_as_system_user do
if self.state == Cancelled
- retryable_requests = ContainerRequest.where("container_uuid = ? and priority > 0 and state = 'Committed' and container_count < container_count_max", uuid)
+ # Cancelled means the container didn't run to completion.
+ # This happens either because it was cancelled by the user
+ # or because there was an infrastructure failure. We want
+ # to retry infrastructure failures automatically.
+ #
+ # Seach for live container requests to determine if we
+ # should retry the container.
+ retryable_requests = ContainerRequest.
+ joins('left outer join containers as requesting_container on container_requests.requesting_container_uuid = requesting_container.uuid').
+ where("container_requests.container_uuid = ? and "+
+ "container_requests.priority > 0 and "+
+ "container_requests.owner_uuid not in (select group_uuid from trashed_groups) and "+
+ "(requesting_container.priority is null or (requesting_container.state = 'Running' and requesting_container.priority > 0)) and "+
+ "container_requests.state = 'Committed' and "+
+ "container_requests.container_count < container_requests.container_count_max", uuid).
+ order('container_requests.uuid asc')
else
retryable_requests = []
end
# Cancel outstanding container requests made by this container.
ContainerRequest.
- includes(:container).
where(requesting_container_uuid: uuid,
- state: ContainerRequest::Committed).each do |cr|
+ state: ContainerRequest::Committed).
+ in_batches(of: 15).each_record do |cr|
leave_modified_by_user_alone do
- cr.update_attributes!(priority: 0)
- cr.container.reload
- if cr.container.state == Container::Queued || cr.container.state == Container::Locked
+ cr.set_priority_zero
+ container_state = Container.where(uuid: cr.container_uuid).pluck(:state).first
+ if container_state == Container::Queued || container_state == Container::Locked
# If the child container hasn't started yet, finalize the
# child CR now instead of leaving it "on hold", i.e.,
# Queued with priority 0. (OTOH, if the child is already
projects / arvados.git / blobdiff