19981: Reuse containers with various keep_cache constraints
[arvados.git] / services / api / app / models / container.rb
index 3a04c56046416771a903714e99543e40c7d66f4e..e6643d4c70d7937f9509e57db420d1e98e9627e0 100644 (file)
@@ -83,6 +83,8 @@ class Container < ArvadosModel
     t.add :interactive_session_started
     t.add :output_storage_classes
     t.add :output_properties
+    t.add :cost
+    t.add :subrequests_cost
   end
 
   # Supported states for a container
@@ -225,6 +227,9 @@ class Container < ArvadosModel
     if rc['keep_cache_ram'] == 0
       rc['keep_cache_ram'] = Rails.configuration.Containers.DefaultKeepCacheRAM
     end
+    if rc['keep_cache_disk'] == 0 and rc['keep_cache_ram'] == 0
+      rc['keep_cache_disk'] = bound_keep_cache_disk(rc['ram'])
+    end
     rc
   end
 
@@ -291,21 +296,55 @@ class Container < ArvadosModel
     candidates = candidates.where('secret_mounts_md5 = ?', secret_mounts_md5)
     log_reuse_info(candidates) { "after filtering on secret_mounts_md5 #{secret_mounts_md5.inspect}" }
 
-    if attrs[:runtime_constraints]['cuda'].nil?
-      attrs[:runtime_constraints]['cuda'] = {
-        'device_count' => 0,
-        'driver_version' => '',
-        'hardware_capability' => '',
-      }
-    end
-    resolved_runtime_constraints = [resolve_runtime_constraints(attrs[:runtime_constraints])]
-    if resolved_runtime_constraints[0]['cuda']['device_count'] == 0
-      # If no CUDA requested, extend search to include older container
-      # records that don't have a 'cuda' section in runtime_constraints
-      resolved_runtime_constraints << resolved_runtime_constraints[0].except('cuda')
-    end
-
-    candidates = candidates.where_serialized(:runtime_constraints, resolved_runtime_constraints, md5: true, multivalue: true)
+    resolved_runtime_constraints = resolve_runtime_constraints(attrs[:runtime_constraints])
+    # Ideally we would completely ignore Keep cache constraints when making
+    # reuse considerations, but our database structure makes that impractical.
+    # The best we can do is generate a search that matches on all likely values.
+    runtime_constraint_variations = {
+      keep_cache_disk: [
+        # Check for constraints without keep_cache_disk
+        # (containers that predate the constraint)
+        nil,
+        # Containers that use keep_cache_ram instead
+        0,
+        # The default value
+        bound_keep_cache_disk(resolved_runtime_constraints['ram']),
+        # The minimum default bound
+        bound_keep_cache_disk(0),
+        # The maximum default bound (presumably)
+        bound_keep_cache_disk(1 << 60),
+        # The requested value
+        resolved_runtime_constraints.delete('keep_cache_disk'),
+      ].uniq,
+      keep_cache_ram: [
+        # Containers that use keep_cache_disk instead
+        0,
+        # The default value
+        Rails.configuration.Containers.DefaultKeepCacheRAM,
+        # The requested value
+        resolved_runtime_constraints.delete('keep_cache_ram'),
+      ].uniq,
+    }
+    resolved_cuda = resolved_runtime_constraints['cuda']
+    if resolved_cuda.nil? or resolved_cuda['device_count'] == 0
+      runtime_constraint_variations[:cuda] = [
+        # Check for constraints without cuda
+        # (containers that predate the constraint)
+        nil,
+        # The default "don't need CUDA" value
+        {
+          'device_count' => 0,
+          'driver_version' => '',
+          'hardware_capability' => '',
+        },
+        # The requested value
+        resolved_runtime_constraints.delete('cuda')
+      ].uniq
+    end
+    reusable_runtime_constraints = hash_product(runtime_constraint_variations)
+                                     .map { |v| resolved_runtime_constraints.merge(v) }
+
+    candidates = candidates.where_serialized(:runtime_constraints, reusable_runtime_constraints, md5: true, multivalue: true)
     log_reuse_info(candidates) { "after filtering on runtime_constraints #{attrs[:runtime_constraints].inspect}" }
 
     log_reuse_info { "checking for state=Complete with readable output and log..." }
@@ -430,6 +469,31 @@ class Container < ArvadosModel
 
   protected
 
+  def self.bound_keep_cache_disk(value)
+    value ||= 0
+    min_value = 2 << 30
+    max_value = 32 << 30
+    if value < min_value
+      min_value
+    elsif value > max_value
+      max_value
+    else
+      value
+    end
+  end
+
+  def self.hash_product(**kwargs)
+    # kwargs is a hash that maps parameters to an array of values.
+    # This function enumerates every possible hash where each key has one of
+    # the values from its array.
+    # The output keys are strings since that's what container hash attributes
+    # want.
+    # A nil value yields a hash without that key.
+    [[:_, nil]].product(
+      *kwargs.map { |(key, values)| [key.to_s].product(values) },
+    ).map { |param_pairs| Hash[param_pairs].compact }
+  end
+
   def fill_field_defaults
     self.state ||= Queued
     self.environment ||= {}
@@ -478,8 +542,9 @@ class Container < ArvadosModel
 
   def validate_change
     permitted = [:state]
-    progress_attrs = [:progress, :runtime_status, :log, :output, :output_properties]
-    final_attrs = [:exit_code, :finished_at]
+    final_attrs = [:finished_at]
+    progress_attrs = [:progress, :runtime_status, :subrequests_cost, :cost,
+                      :log, :output, :output_properties, :exit_code]
 
     if self.new_record?
       permitted.push(:owner_uuid, :command, :container_image, :cwd,
@@ -498,9 +563,9 @@ class Container < ArvadosModel
       permitted.push :priority
 
     when Running
-      permitted.push :priority, :output_properties, *progress_attrs
+      permitted.push :priority, :output_properties, :gateway_address, *progress_attrs
       if self.state_changed?
-        permitted.push :started_at, :gateway_address
+        permitted.push :started_at
       end
       if !self.interactive_session_started_was
         permitted.push :interactive_session_started
@@ -516,7 +581,7 @@ class Container < ArvadosModel
       when Running
         permitted.push :finished_at, *progress_attrs
       when Queued, Locked
-        permitted.push :finished_at, :log, :runtime_status
+        permitted.push :finished_at, :log, :runtime_status, :cost
       end
 
     else
@@ -700,6 +765,36 @@ class Container < ArvadosModel
           end
 
           if retryable_requests.any?
+            scheduling_parameters = {
+              # partitions: empty if any are empty, else the union of all parameters
+              "partitions": retryable_requests
+                              .map { |req| req.scheduling_parameters["partitions"] || [] }
+                              .reduce { |cur, new| (cur.empty? or new.empty?) ? [] : (cur | new) },
+
+              # preemptible: true if all are true, else false
+              "preemptible": retryable_requests
+                               .map { |req| req.scheduling_parameters["preemptible"] }
+                               .all?,
+
+              # supervisor: true if all any true, else false
+              "supervisor": retryable_requests
+                               .map { |req| req.scheduling_parameters["supervisor"] }
+                               .any?,
+
+              # max_run_time: 0 if any are 0 (unlimited), else the maximum
+              "max_run_time": retryable_requests
+                                .map { |req| req.scheduling_parameters["max_run_time"] || 0 }
+                                .reduce do |cur, new|
+                if cur == 0 or new == 0
+                  0
+                elsif new > cur
+                  new
+                else
+                  cur
+                end
+              end,
+            }
+
             c_attrs = {
               command: self.command,
               cwd: self.cwd,
@@ -708,7 +803,7 @@ class Container < ArvadosModel
               container_image: self.container_image,
               mounts: self.mounts,
               runtime_constraints: self.runtime_constraints,
-              scheduling_parameters: self.scheduling_parameters,
+              scheduling_parameters: scheduling_parameters,
               secret_mounts: prev_secret_mounts,
               runtime_token: prev_runtime_token,
               runtime_user_uuid: self.runtime_user_uuid,
@@ -719,6 +814,7 @@ class Container < ArvadosModel
               cr.with_lock do
                 leave_modified_by_user_alone do
                   # Use row locking because this increments container_count
+                  cr.cumulative_cost += self.cost + self.subrequests_cost
                   cr.container_uuid = c.uuid
                   cr.save!
                 end