Merge branch '8784-dir-listings'
[arvados.git] / services / api / lib / crunch_dispatch.rb
index 05f85c7bb67f8c7863ed46fd793ad6a96fa2cf77..230f03e5777a56aec28b57bda56dcea4e82b065f 100644 (file)
@@ -1,3 +1,7 @@
+# Copyright (C) The Arvados Authors. All rights reserved.
+#
+# SPDX-License-Identifier: AGPL-3.0
+
 require 'open3'
 require 'shellwords'
 
@@ -24,9 +28,10 @@ class CrunchDispatch
 
     @docker_bin = ENV['CRUNCH_JOB_DOCKER_BIN']
     @docker_run_args = ENV['CRUNCH_JOB_DOCKER_RUN_ARGS']
+    @cgroup_root = ENV['CRUNCH_CGROUP_ROOT']
 
     @arvados_internal = Rails.configuration.git_internal_dir
-    if not File.exists? @arvados_internal
+    if not File.exist? @arvados_internal
       $stderr.puts `mkdir -p #{@arvados_internal.shellescape} && git init --bare #{@arvados_internal.shellescape}`
       raise "No internal git repository available" unless ($? == 0)
     end
@@ -72,7 +77,7 @@ class CrunchDispatch
       # into multiple rows with one hostname each.
       `#{cmd} --noheader -o '%N:#{outfmt}'`.each_line do |line|
         tokens = line.chomp.split(":", max_fields)
-        if (re = tokens[0].match /^(.*?)\[([-,\d]+)\]$/)
+        if (re = tokens[0].match(/^(.*?)\[([-,\d]+)\]$/))
           tokens.shift
           re[2].split(",").each do |range|
             range = range.split("-").collect(&:to_i)
@@ -94,7 +99,7 @@ class CrunchDispatch
       # hasn't been able to communicate with it recently.
       state.sub!(/^idle\*/, "down")
       state.sub!(/\W+$/, "")
-      state = "down" unless %w(idle alloc down).include?(state)
+      state = "down" unless %w(idle alloc comp mix drng down).include?(state)
       slurm_nodes[hostname] = {state: state, job: nil}
     end
     each_slurm_line("squeue", "%j") do |hostname, job_uuid|
@@ -104,7 +109,7 @@ class CrunchDispatch
   end
 
   def update_node_status
-    return unless Server::Application.config.crunch_job_wrapper.to_s.match /^slurm/
+    return unless Server::Application.config.crunch_job_wrapper.to_s.match(/^slurm/)
     slurm_status.each_pair do |hostname, slurmdata|
       next if @node_state[hostname] == slurmdata
       begin
@@ -168,7 +173,25 @@ class CrunchDispatch
       end
       usable_nodes << node
       if usable_nodes.count >= min_node_count
-        return usable_nodes.map { |node| node.hostname }
+        hostnames = usable_nodes.map(&:hostname)
+        log_nodes = usable_nodes.map do |n|
+          "#{n.hostname} #{n.uuid} #{n.properties.to_json}"
+        end
+        log_job = "#{job.uuid} #{job.runtime_constraints}"
+        log_text = "dispatching job #{log_job} to #{log_nodes.join(", ")}"
+        $stderr.puts log_text
+        begin
+          act_as_system_user do
+            Log.new(object_uuid: job.uuid,
+                    event_type: 'dispatch',
+                    owner_uuid: system_user_uuid,
+                    summary: "dispatching to #{hostnames.join(", ")}",
+                    properties: {'text' => log_text}).save!
+          end
+        rescue => e
+          $stderr.puts "dispatch: log.create failed: #{e}"
+        end
+        return hostnames
       end
     end
     nil
@@ -203,8 +226,8 @@ class CrunchDispatch
               owner_uuid: job.owner_uuid,
               summary: message,
               properties: {"text" => message}).save!
-    rescue
-      $stderr.puts "dispatch: log.create failed"
+    rescue => e
+      $stderr.puts "dispatch: log.create failed: #{e}"
     end
 
     if not skip_lock and not have_job_lock?(job)
@@ -384,6 +407,10 @@ class CrunchDispatch
                    '--job', job.uuid,
                    '--git-dir', @arvados_internal]
 
+      if @cgroup_root
+        cmd_args += ['--cgroup-root', @cgroup_root]
+      end
+
       if @docker_bin
         cmd_args += ['--docker-bin', @docker_bin]
       end
@@ -429,6 +456,8 @@ class CrunchDispatch
         log_throttle_bytes_so_far: 0,
         log_throttle_lines_so_far: 0,
         log_throttle_bytes_skipped: 0,
+        log_throttle_partial_line_last_at: Time.new(0),
+        log_throttle_first_partial_line: true,
       }
       i.close
       @todo_job_retries.delete(job.uuid)
@@ -443,9 +472,23 @@ class CrunchDispatch
     message = false
     linesize = line.size
     if running_job[:log_throttle_is_open]
-      running_job[:log_throttle_lines_so_far] += 1
-      running_job[:log_throttle_bytes_so_far] += linesize
-      running_job[:bytes_logged] += linesize
+      partial_line = false
+      skip_counts = false
+      matches = line.match(/^\S+ \S+ \d+ \d+ stderr (.*)/)
+      if matches and matches[1] and matches[1].start_with?('[...]') and matches[1].end_with?('[...]')
+        partial_line = true
+        if Time.now > running_job[:log_throttle_partial_line_last_at] + Rails.configuration.crunch_log_partial_line_throttle_period
+          running_job[:log_throttle_partial_line_last_at] = Time.now
+        else
+          skip_counts = true
+        end
+      end
+
+      if !skip_counts
+        running_job[:log_throttle_lines_so_far] += 1
+        running_job[:log_throttle_bytes_so_far] += linesize
+        running_job[:bytes_logged] += linesize
+      end
 
       if (running_job[:bytes_logged] >
           Rails.configuration.crunch_limit_log_bytes_per_job)
@@ -456,14 +499,18 @@ class CrunchDispatch
       elsif (running_job[:log_throttle_bytes_so_far] >
              Rails.configuration.crunch_log_throttle_bytes)
         remaining_time = running_job[:log_throttle_reset_time] - Time.now
-        message = "Exceeded rate #{Rails.configuration.crunch_log_throttle_bytes} bytes per #{Rails.configuration.crunch_log_throttle_period} seconds (crunch_log_throttle_bytes). Logging will be silenced for the next #{remaining_time.round} seconds.\n"
+        message = "Exceeded rate #{Rails.configuration.crunch_log_throttle_bytes} bytes per #{Rails.configuration.crunch_log_throttle_period} seconds (crunch_log_throttle_bytes). Logging will be silenced for the next #{remaining_time.round} seconds."
         running_job[:log_throttle_is_open] = false
 
       elsif (running_job[:log_throttle_lines_so_far] >
              Rails.configuration.crunch_log_throttle_lines)
         remaining_time = running_job[:log_throttle_reset_time] - Time.now
-        message = "Exceeded rate #{Rails.configuration.crunch_log_throttle_lines} lines per #{Rails.configuration.crunch_log_throttle_period} seconds (crunch_log_throttle_lines), logging will be silenced for the next #{remaining_time.round} seconds.\n"
+        message = "Exceeded rate #{Rails.configuration.crunch_log_throttle_lines} lines per #{Rails.configuration.crunch_log_throttle_period} seconds (crunch_log_throttle_lines), logging will be silenced for the next #{remaining_time.round} seconds."
         running_job[:log_throttle_is_open] = false
+
+      elsif partial_line and running_job[:log_throttle_first_partial_line]
+        running_job[:log_throttle_first_partial_line] = false
+        message = "Rate-limiting partial segments of long lines to one every #{Rails.configuration.crunch_log_partial_line_throttle_period} seconds."
       end
     end
 
@@ -475,8 +522,11 @@ class CrunchDispatch
     if message
       # Yes, write to logs, but use our "rate exceeded" message
       # instead of the log message that exceeded the limit.
+      message += " A complete log is still being written to Keep, and will be available when the job finishes.\n"
       line.replace message
       true
+    elsif partial_line
+      false
     else
       running_job[:log_throttle_is_open]
     end
@@ -484,8 +534,6 @@ class CrunchDispatch
 
   def read_pipes
     @running.each do |job_uuid, j|
-      job = j[:job]
-
       now = Time.now
       if now > j[:log_throttle_reset_time]
         # It has been more than throttle_period seconds since the last
@@ -501,6 +549,8 @@ class CrunchDispatch
         j[:log_throttle_lines_so_far] = 0
         j[:log_throttle_bytes_skipped] = 0
         j[:log_throttle_is_open] = true
+        j[:log_throttle_partial_line_last_at] = Time.new(0)
+        j[:log_throttle_first_partial_line] = true
       end
 
       j[:buf].each do |stream, streambuf|
@@ -637,19 +687,21 @@ class CrunchDispatch
 
     jobrecord = Job.find_by_uuid(job_done.uuid)
 
-    if exit_status == EXIT_RETRY_UNLOCKED
-      # The job failed because all of the nodes allocated to it
-      # failed.  Only this crunch-dispatch process can retry the job:
+    if exit_status == EXIT_RETRY_UNLOCKED or (exit_tempfail and @job_retry_counts.include? jobrecord.uuid)
+      $stderr.puts("dispatch: job #{jobrecord.uuid} was interrupted by node failure")
+      # Only this crunch-dispatch process can retry the job:
       # it's already locked, and there's no way to put it back in the
       # Queued state.  Put it in our internal todo list unless the job
       # has failed this way excessively.
       @job_retry_counts[jobrecord.uuid] += 1
       exit_tempfail = @job_retry_counts[jobrecord.uuid] <= RETRY_UNLOCKED_LIMIT
+      do_what_next = "give up now"
       if exit_tempfail
         @todo_job_retries[jobrecord.uuid] = jobrecord
-      else
-        $stderr.puts("dispatch: job #{jobrecord.uuid} exceeded node failure retry limit -- giving up")
+        do_what_next = "re-attempt"
       end
+      $stderr.puts("dispatch: job #{jobrecord.uuid} has been interrupted " +
+                   "#{@job_retry_counts[jobrecord.uuid]}x, will #{do_what_next}")
     end
 
     if !exit_tempfail
@@ -787,6 +839,9 @@ class CrunchDispatch
         unless (@todo_pipelines.empty? and @pipe_auth_tokens.empty?) or did_recently(:update_pipelines, 5.0)
           update_pipelines
         end
+        unless did_recently('check_orphaned_slurm_jobs', 60)
+          check_orphaned_slurm_jobs
+        end
       end
       reap_children
       select(@running.values.collect { |j| [j[:stdout], j[:stderr]] }.flatten,
@@ -823,22 +878,14 @@ class CrunchDispatch
       end
       Rails.logger.info "fail_jobs: threshold is #{threshold}"
 
-      if Rails.configuration.crunch_job_wrapper == :slurm_immediate
-        # [["slurm_job_id", "slurm_job_name"], ...]
-        squeue = File.popen(['squeue', '-h', '-o', '%i %j']).readlines.map do |line|
-          line.strip.split(' ', 2)
-        end
-      else
-        squeue = []
-      end
-
+      squeue = squeue_jobs
       Job.where('state = ? and started_at < ?', Job::Running, threshold).
         each do |job|
         Rails.logger.debug "fail_jobs: #{job.uuid} started #{job.started_at}"
-        squeue.each do |slurm_id, slurm_name|
+        squeue.each do |slurm_name|
           if slurm_name == job.uuid
-            Rails.logger.info "fail_jobs: scancel #{slurm_id} for #{job.uuid}"
-            scancel slurm_id
+            Rails.logger.info "fail_jobs: scancel #{job.uuid}"
+            scancel slurm_name
           end
         end
         fail_job(job, "cleaned up stale job: started before #{threshold}",
@@ -847,6 +894,37 @@ class CrunchDispatch
     end
   end
 
+  def check_orphaned_slurm_jobs
+    act_as_system_user do
+      squeue_uuids = squeue_jobs.select{|uuid| uuid.match(/^[0-9a-z]{5}-8i9sb-[0-9a-z]{15}$/)}.
+                                  select{|uuid| !@running.has_key?(uuid)}
+
+      return if squeue_uuids.size == 0
+
+      scancel_uuids = squeue_uuids - Job.where('uuid in (?) and (state in (?) or modified_at>?)',
+                                               squeue_uuids,
+                                               ['Running', 'Queued'],
+                                               (Time.now - 60)).
+                                         collect(&:uuid)
+      scancel_uuids.each do |uuid|
+        Rails.logger.info "orphaned job: scancel #{uuid}"
+        scancel uuid
+      end
+    end
+  end
+
+  def sudo_preface
+    return [] if not Server::Application.config.crunch_job_user
+    ["sudo", "-E", "-u",
+     Server::Application.config.crunch_job_user,
+     "LD_LIBRARY_PATH=#{ENV['LD_LIBRARY_PATH']}",
+     "PATH=#{ENV['PATH']}",
+     "PERLLIB=#{ENV['PERLLIB']}",
+     "PYTHONPATH=#{ENV['PYTHONPATH']}",
+     "RUBYLIB=#{ENV['RUBYLIB']}",
+     "GEM_PATH=#{ENV['GEM_PATH']}"]
+  end
+
   protected
 
   def have_job_lock?(job)
@@ -889,23 +967,27 @@ class CrunchDispatch
     end
   end
 
-  def scancel slurm_id
-    cmd = sudo_preface + ['scancel', slurm_id]
-    puts File.popen(cmd).read
-    if not $?.success?
-      Rails.logger.error "scancel #{slurm_id.shellescape}: $?"
+  # An array of job_uuids in squeue
+  def squeue_jobs
+    if Rails.configuration.crunch_job_wrapper == :slurm_immediate
+      p = IO.popen(['squeue', '-a', '-h', '-o', '%j'])
+      begin
+        p.readlines.map {|line| line.strip}
+      ensure
+        p.close
+      end
+    else
+      []
     end
   end
 
-  def sudo_preface
-    return [] if not Server::Application.config.crunch_job_user
-    ["sudo", "-E", "-u",
-     Server::Application.config.crunch_job_user,
-     "LD_LIBRARY_PATH=#{ENV['LD_LIBRARY_PATH']}",
-     "PATH=#{ENV['PATH']}",
-     "PERLLIB=#{ENV['PERLLIB']}",
-     "PYTHONPATH=#{ENV['PYTHONPATH']}",
-     "RUBYLIB=#{ENV['RUBYLIB']}",
-     "GEM_PATH=#{ENV['GEM_PATH']}"]
+  def scancel slurm_name
+    cmd = sudo_preface + ['scancel', '-n', slurm_name]
+    IO.popen(cmd) do |scancel_pipe|
+      puts scancel_pipe.read
+    end
+    if not $?.success?
+      Rails.logger.error "scancel #{slurm_name.shellescape}: $?"
+    end
   end
 end