6 include ApplicationHelper
10 EXIT_RETRY_UNLOCKED = 93
11 RETRY_UNLOCKED_LIMIT = 3
15 self.utc.strftime "%Y-%m-%d_%H:%M:%S"
20 @crunch_job_bin = (ENV['CRUNCH_JOB_BIN'] || `which arv-crunch-job`.strip)
21 if @crunch_job_bin.empty?
22 raise "No CRUNCH_JOB_BIN env var, and crunch-job not in path."
25 @docker_bin = ENV['CRUNCH_JOB_DOCKER_BIN']
26 @docker_run_args = ENV['CRUNCH_JOB_DOCKER_RUN_ARGS']
28 @arvados_internal = Rails.configuration.git_internal_dir
29 if not File.exists? @arvados_internal
30 $stderr.puts `mkdir -p #{@arvados_internal.shellescape} && git init --bare #{@arvados_internal.shellescape}`
31 raise "No internal git repository available" unless ($? == 0)
34 @repo_root = Rails.configuration.git_repositories_dir
35 @arvados_repo_path = Repository.where(name: "arvados").first.server_path
41 @pipe_auth_tokens = {}
44 @todo_job_retries = {}
45 @job_retry_counts = Hash.new(0)
50 return act_as_system_user
55 @todo = @todo_job_retries.values + Job.queue.select(&:repository)
57 if @runoptions[:pipelines]
58 @todo_pipelines = PipelineInstance.queue
62 def each_slurm_line(cmd, outfmt, max_fields=nil)
63 max_fields ||= outfmt.split(":").size
64 max_fields += 1 # To accommodate the node field we add
65 @@slurm_version ||= Gem::Version.new(`sinfo --version`.match(/\b[\d\.]+\b/)[0])
66 if Gem::Version.new('2.3') <= @@slurm_version
67 `#{cmd} --noheader -o '%n:#{outfmt}'`.each_line do |line|
68 yield line.chomp.split(":", max_fields)
71 # Expand rows with hostname ranges (like "foo[1-3,5,9-12]:idle")
72 # into multiple rows with one hostname each.
73 `#{cmd} --noheader -o '%N:#{outfmt}'`.each_line do |line|
74 tokens = line.chomp.split(":", max_fields)
75 if (re = tokens[0].match /^(.*?)\[([-,\d]+)\]$/)
77 re[2].split(",").each do |range|
78 range = range.split("-").collect(&:to_i)
79 (range[0]..range[-1]).each do |n|
80 yield [re[1] + n.to_s] + tokens
92 each_slurm_line("sinfo", "%t") do |hostname, state|
93 # Treat nodes in idle* state as down, because the * means that slurm
94 # hasn't been able to communicate with it recently.
95 state.sub!(/^idle\*/, "down")
96 state.sub!(/\W+$/, "")
97 state = "down" unless %w(idle alloc down).include?(state)
98 slurm_nodes[hostname] = {state: state, job: nil}
100 each_slurm_line("squeue", "%j") do |hostname, job_uuid|
101 slurm_nodes[hostname][:job] = job_uuid if slurm_nodes[hostname]
106 def update_node_status
107 return unless Server::Application.config.crunch_job_wrapper.to_s.match /^slurm/
108 slurm_status.each_pair do |hostname, slurmdata|
109 next if @node_state[hostname] == slurmdata
111 node = Node.where('hostname=?', hostname).order(:last_ping_at).last
113 $stderr.puts "dispatch: update #{hostname} state to #{slurmdata}"
114 node.info["slurm_state"] = slurmdata[:state]
115 node.job_uuid = slurmdata[:job]
117 @node_state[hostname] = slurmdata
119 $stderr.puts "dispatch: failed to update #{node.uuid}: #{node.errors.messages}"
121 elsif slurmdata[:state] != 'down'
122 $stderr.puts "dispatch: SLURM reports '#{hostname}' is not down, but no node has that name"
125 $stderr.puts "dispatch: error updating #{hostname} node status: #{error}"
130 def positive_int(raw_value, default=nil)
131 value = begin raw_value.to_i rescue 0 end
139 NODE_CONSTRAINT_MAP = {
140 # Map Job runtime_constraints keys to the corresponding Node info key.
141 'min_ram_mb_per_node' => 'total_ram_mb',
142 'min_scratch_mb_per_node' => 'total_scratch_mb',
143 'min_cores_per_node' => 'total_cpu_cores',
146 def nodes_available_for_job_now(job)
147 # Find Nodes that satisfy a Job's runtime constraints (by building
148 # a list of Procs and using them to test each Node). If there
149 # enough to run the Job, return an array of their names.
150 # Otherwise, return nil.
151 need_procs = NODE_CONSTRAINT_MAP.each_pair.map do |job_key, node_key|
153 positive_int(node.properties[node_key], 0) >=
154 positive_int(job.runtime_constraints[job_key], 0)
157 min_node_count = positive_int(job.runtime_constraints['min_nodes'], 1)
159 Node.all.select do |node|
160 node.info['slurm_state'] == 'idle'
161 end.sort_by do |node|
162 # Prefer nodes with no price, then cheap nodes, then expensive nodes
163 node.properties['cloud_node']['price'].to_f rescue 0
165 if need_procs.select { |need_proc| not need_proc.call(node) }.any?
166 # At least one runtime constraint is not satisfied by this node
170 if usable_nodes.count >= min_node_count
171 return usable_nodes.map { |node| node.hostname }
177 def nodes_available_for_job(job)
178 # Check if there are enough idle nodes with the Job's minimum
179 # hardware requirements to run it. If so, return an array of
180 # their names. If not, up to once per hour, signal start_jobs to
181 # hold off launching Jobs. This delay is meant to give the Node
182 # Manager an opportunity to make new resources available for new
185 # The exact timing parameters here might need to be adjusted for
186 # the best balance between helping the longest-waiting Jobs run,
187 # and making efficient use of immediately available resources.
188 # These are all just first efforts until we have more data to work
190 nodelist = nodes_available_for_job_now(job)
191 if nodelist.nil? and not did_recently(:wait_for_available_nodes, 3600)
192 $stderr.puts "dispatch: waiting for nodes for #{job.uuid}"
193 @node_wait_deadline = Time.now + 5.minutes
198 def fail_job job, message, skip_lock: false
199 $stderr.puts "dispatch: #{job.uuid}: #{message}"
201 Log.new(object_uuid: job.uuid,
202 event_type: 'dispatch',
203 owner_uuid: job.owner_uuid,
205 properties: {"text" => message}).save!
207 $stderr.puts "dispatch: log.create failed"
210 if not skip_lock and not have_job_lock?(job)
212 job.lock @authorizations[job.uuid].user.uuid
213 rescue ArvadosModel::AlreadyLockedError
214 $stderr.puts "dispatch: tried to mark job #{job.uuid} as failed but it was already locked by someone else"
221 $stderr.puts "dispatch: save failed setting job #{job.uuid} to failed"
225 def stdout_s(cmd_a, opts={})
226 IO.popen(cmd_a, "r", opts) do |pipe|
227 return pipe.read.chomp
232 ["git", "--git-dir=#{@arvados_internal}"] + cmd_a
235 def get_authorization(job)
236 if @authorizations[job.uuid] and
237 @authorizations[job.uuid].user.uuid != job.modified_by_user_uuid
238 # We already made a token for this job, but we need a new one
239 # because modified_by_user_uuid has changed (the job will run
240 # as a different user).
241 @authorizations[job.uuid].update_attributes expires_at: Time.now
242 @authorizations[job.uuid] = nil
244 if not @authorizations[job.uuid]
245 auth = ApiClientAuthorization.
246 new(user: User.where('uuid=?', job.modified_by_user_uuid).first,
249 $stderr.puts "dispatch: auth.save failed for #{job.uuid}"
251 @authorizations[job.uuid] = auth
254 @authorizations[job.uuid]
257 def internal_repo_has_commit? sha1
258 if (not @fetched_commits[sha1] and
259 sha1 == stdout_s(git_cmd("rev-list", "-n1", sha1), err: "/dev/null") and
261 @fetched_commits[sha1] = true
263 return @fetched_commits[sha1]
266 def get_commit src_repo, sha1
267 return true if internal_repo_has_commit? sha1
269 # commit does not exist in internal repository, so import the
270 # source repository using git fetch-pack
271 cmd = git_cmd("fetch-pack", "--no-progress", "--all", src_repo)
272 $stderr.puts "dispatch: #{cmd}"
273 $stderr.puts(stdout_s(cmd))
274 @fetched_commits[sha1] = ($? == 0)
277 def tag_commit(commit_hash, tag_name)
278 # @git_tags[T]==V if we know commit V has been tagged T in the
279 # arvados_internal repository.
280 if not @git_tags[tag_name]
281 cmd = git_cmd("tag", tag_name, commit_hash)
282 $stderr.puts "dispatch: #{cmd}"
283 $stderr.puts(stdout_s(cmd, err: "/dev/null"))
285 # git tag failed. This may be because the tag already exists, so check for that.
286 tag_rev = stdout_s(git_cmd("rev-list", "-n1", tag_name))
288 # We got a revision back
289 if tag_rev != commit_hash
290 # Uh oh, the tag doesn't point to the revision we were expecting.
291 # Someone has been monkeying with the job record and/or git.
292 fail_job job, "Existing tag #{tag_name} points to commit #{tag_rev} but expected commit #{commit_hash}"
295 # we're okay (fall through to setting @git_tags below)
297 # git rev-list failed for some reason.
298 fail_job job, "'git tag' for #{tag_name} failed but did not find any existing tag using 'git rev-list'"
302 # 'git tag' was successful, or there is an existing tag that points to the same revision.
303 @git_tags[tag_name] = commit_hash
304 elsif @git_tags[tag_name] != commit_hash
305 fail_job job, "Existing tag #{tag_name} points to commit #{@git_tags[tag_name]} but this job uses commit #{commit_hash}"
313 next if @running[job.uuid]
316 case Server::Application.config.crunch_job_wrapper
319 # Don't run more than one at a time.
323 when :slurm_immediate
324 nodelist = nodes_available_for_job(job)
326 if Time.now < @node_wait_deadline
332 cmd_args = ["salloc",
337 "--job-name=#{job.uuid}",
338 "--nodelist=#{nodelist.join(',')}"]
340 raise "Unknown crunch_job_wrapper: #{Server::Application.config.crunch_job_wrapper}"
343 cmd_args = sudo_preface + cmd_args
345 next unless get_authorization job
347 ready = internal_repo_has_commit? job.script_version
350 # Import the commit from the specified repository into the
351 # internal repository. This should have been done already when
352 # the job was created/updated; this code is obsolete except to
353 # avoid deployment races. Failing the job would be a
354 # reasonable thing to do at this point.
355 repo = Repository.where(name: job.repository).first
356 if repo.nil? or repo.server_path.nil?
357 fail_job job, "Repository #{job.repository} not found under #{@repo_root}"
360 ready &&= get_commit repo.server_path, job.script_version
361 ready &&= tag_commit job.script_version, job.uuid
364 # This should be unnecessary, because API server does it during
365 # job create/update, but it's still not a bad idea to verify the
366 # tag is correct before starting the job:
367 ready &&= tag_commit job.script_version, job.uuid
369 # The arvados_sdk_version doesn't support use of arbitrary
370 # remote URLs, so the requested version isn't necessarily copied
371 # into the internal repository yet.
372 if job.arvados_sdk_version
373 ready &&= get_commit @arvados_repo_path, job.arvados_sdk_version
374 ready &&= tag_commit job.arvados_sdk_version, "#{job.uuid}-arvados-sdk"
378 fail_job job, "commit not present in internal repository"
382 cmd_args += [@crunch_job_bin,
383 '--job-api-token', @authorizations[job.uuid].api_token,
385 '--git-dir', @arvados_internal]
388 cmd_args += ['--docker-bin', @docker_bin]
392 cmd_args += ['--docker-run-args', @docker_run_args]
395 if have_job_lock?(job)
396 cmd_args << "--force-unlock"
399 $stderr.puts "dispatch: #{cmd_args.join ' '}"
402 i, o, e, t = Open3.popen3(*cmd_args)
404 $stderr.puts "dispatch: popen3: #{$!}"
409 $stderr.puts "dispatch: job #{job.uuid}"
410 start_banner = "dispatch: child #{t.pid} start #{LogTime.now}"
411 $stderr.puts start_banner
413 @running[job.uuid] = {
419 buf: {stderr: '', stdout: ''},
422 job_auth: @authorizations[job.uuid],
423 stderr_buf_to_flush: '',
424 stderr_flushed_at: Time.new(0),
427 log_throttle_is_open: true,
428 log_throttle_reset_time: Time.now + Rails.configuration.crunch_log_throttle_period,
429 log_throttle_bytes_so_far: 0,
430 log_throttle_lines_so_far: 0,
431 log_throttle_bytes_skipped: 0,
434 @todo_job_retries.delete(job.uuid)
439 # Test for hard cap on total output and for log throttling. Returns whether
440 # the log line should go to output or not. Modifies "line" in place to
441 # replace it with an error if a logging limit is tripped.
442 def rate_limit running_job, line
445 if running_job[:log_throttle_is_open]
446 running_job[:log_throttle_lines_so_far] += 1
447 running_job[:log_throttle_bytes_so_far] += linesize
448 running_job[:bytes_logged] += linesize
450 if (running_job[:bytes_logged] >
451 Rails.configuration.crunch_limit_log_bytes_per_job)
452 message = "Exceeded log limit #{Rails.configuration.crunch_limit_log_bytes_per_job} bytes (crunch_limit_log_bytes_per_job). Log will be truncated."
453 running_job[:log_throttle_reset_time] = Time.now + 100.years
454 running_job[:log_throttle_is_open] = false
456 elsif (running_job[:log_throttle_bytes_so_far] >
457 Rails.configuration.crunch_log_throttle_bytes)
458 remaining_time = running_job[:log_throttle_reset_time] - Time.now
459 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"
460 running_job[:log_throttle_is_open] = false
462 elsif (running_job[:log_throttle_lines_so_far] >
463 Rails.configuration.crunch_log_throttle_lines)
464 remaining_time = running_job[:log_throttle_reset_time] - Time.now
465 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"
466 running_job[:log_throttle_is_open] = false
470 if not running_job[:log_throttle_is_open]
471 # Don't log anything if any limit has been exceeded. Just count lossage.
472 running_job[:log_throttle_bytes_skipped] += linesize
476 # Yes, write to logs, but use our "rate exceeded" message
477 # instead of the log message that exceeded the limit.
481 running_job[:log_throttle_is_open]
486 @running.each do |job_uuid, j|
490 if now > j[:log_throttle_reset_time]
491 # It has been more than throttle_period seconds since the last
492 # checkpoint so reset the throttle
493 if j[:log_throttle_bytes_skipped] > 0
494 message = "#{job_uuid} ! Skipped #{j[:log_throttle_bytes_skipped]} bytes of log"
496 j[:stderr_buf_to_flush] << "#{LogTime.now} #{message}\n"
499 j[:log_throttle_reset_time] = now + Rails.configuration.crunch_log_throttle_period
500 j[:log_throttle_bytes_so_far] = 0
501 j[:log_throttle_lines_so_far] = 0
502 j[:log_throttle_bytes_skipped] = 0
503 j[:log_throttle_is_open] = true
506 j[:buf].each do |stream, streambuf|
507 # Read some data from the child stream
510 # It's important to use a big enough buffer here. When we're
511 # being flooded with logs, we must read and discard many
512 # bytes at once. Otherwise, we can easily peg a CPU with
513 # time-checking and other loop overhead. (Quick tests show a
514 # 1MiB buffer working 2.5x as fast as a 64 KiB buffer.)
516 # So don't reduce this buffer size!
517 buf = j[stream].read_nonblock(2**20)
518 rescue Errno::EAGAIN, EOFError
521 # Short circuit the counting code if we're just going to throw
522 # away the data anyway.
523 if not j[:log_throttle_is_open]
524 j[:log_throttle_bytes_skipped] += streambuf.size + buf.size
531 # Append to incomplete line from previous read, if any
535 streambuf.each_line do |line|
536 if not line.end_with? $/
537 if line.size > Rails.configuration.crunch_log_throttle_bytes
538 # Without a limit here, we'll use 2x an arbitrary amount
539 # of memory, and waste a lot of time copying strings
540 # around, all without providing any feedback to anyone
541 # about what's going on _or_ hitting any of our throttle
544 # Here we leave "line" alone, knowing it will never be
545 # sent anywhere: rate_limit() will reach
546 # crunch_log_throttle_bytes immediately. However, we'll
547 # leave [...] in bufend: if the trailing end of the long
548 # line does end up getting sent anywhere, it will have
549 # some indication that it is incomplete.
552 # If line length is sane, we'll wait for the rest of the
553 # line to appear in the next read_pipes() call.
558 # rate_limit returns true or false as to whether to actually log
559 # the line or not. It also modifies "line" in place to replace
560 # it with an error if a logging limit is tripped.
561 if rate_limit j, line
562 $stderr.print "#{job_uuid} ! " unless line.index(job_uuid)
564 pub_msg = "#{LogTime.now} #{line.strip}\n"
565 j[:stderr_buf_to_flush] << pub_msg
569 # Leave the trailing incomplete line (if any) in streambuf for
571 streambuf.replace bufend
573 # Flush buffered logs to the logs table, if appropriate. We have
574 # to do this even if we didn't collect any new logs this time:
575 # otherwise, buffered data older than seconds_between_events
576 # won't get flushed until new data arrives.
582 return if 0 == @running.size
588 pid_done = waitpid(-1, Process::WNOHANG | Process::WUNTRACED)
590 j_done = @running.values.
591 select { |j| j[:wait_thr].pid == pid_done }.
594 rescue SystemCallError
595 # I have @running processes but system reports I have no
596 # children. This is likely to happen repeatedly if it happens at
597 # all; I will log this no more than once per child process I
599 if 0 < @running.select { |uuid,j| j[:warned_waitpid_error].nil? }.size
600 children = @running.values.collect { |j| j[:wait_thr].pid }.join ' '
601 $stderr.puts "dispatch: IPC bug: waitpid() error (#{$!}), but I have children #{children}"
603 @running.each do |uuid,j| j[:warned_waitpid_error] = true end
606 @running.each do |uuid, j|
607 if j[:wait_thr].status == false
608 pid_done = j[:wait_thr].pid
616 job_done = j_done[:job]
618 # Ensure every last drop of stdout and stderr is consumed.
620 # Reset flush timestamp to make sure log gets written.
621 j_done[:stderr_flushed_at] = Time.new(0)
622 # Write any remaining logs.
625 j_done[:buf].each do |stream, streambuf|
627 $stderr.puts streambuf + "\n"
631 # Wait the thread (returns a Process::Status)
632 exit_status = j_done[:wait_thr].value.exitstatus
633 exit_tempfail = exit_status == EXIT_TEMPFAIL
635 $stderr.puts "dispatch: child #{pid_done} exit #{exit_status}"
636 $stderr.puts "dispatch: job #{job_done.uuid} end"
638 jobrecord = Job.find_by_uuid(job_done.uuid)
640 if exit_status == EXIT_RETRY_UNLOCKED
641 # The job failed because all of the nodes allocated to it
642 # failed. Only this crunch-dispatch process can retry the job:
643 # it's already locked, and there's no way to put it back in the
644 # Queued state. Put it in our internal todo list unless the job
645 # has failed this way excessively.
646 @job_retry_counts[jobrecord.uuid] += 1
647 exit_tempfail = @job_retry_counts[jobrecord.uuid] <= RETRY_UNLOCKED_LIMIT
649 @todo_job_retries[jobrecord.uuid] = jobrecord
651 $stderr.puts("dispatch: job #{jobrecord.uuid} exceeded node failure retry limit -- giving up")
656 @job_retry_counts.delete(jobrecord.uuid)
657 if jobrecord.state == "Running"
658 # Apparently there was an unhandled error. That could potentially
659 # include "all allocated nodes failed" when we don't to retry
660 # because the job has already been retried RETRY_UNLOCKED_LIMIT
661 # times. Fail the job.
662 jobrecord.state = "Failed"
663 if not jobrecord.save
664 $stderr.puts "dispatch: jobrecord.save failed"
668 # If the job failed to run due to an infrastructure
669 # issue with crunch-job or slurm, we want the job to stay in the
670 # queue. If crunch-job exited after losing a race to another
671 # crunch-job process, it exits 75 and we should leave the job
672 # record alone so the winner of the race can do its thing.
673 # If crunch-job exited after all of its allocated nodes failed,
674 # it exits 93, and we want to retry it later (see the
675 # EXIT_RETRY_UNLOCKED `if` block).
677 # There is still an unhandled race condition: If our crunch-job
678 # process is about to lose a race with another crunch-job
679 # process, but crashes before getting to its "exit 75" (for
680 # example, "cannot fork" or "cannot reach API server") then we
681 # will assume incorrectly that it's our process's fault
682 # jobrecord.started_at is non-nil, and mark the job as failed
683 # even though the winner of the race is probably still doing
687 # Invalidate the per-job auth token, unless the job is still queued and we
688 # might want to try it again.
689 if jobrecord.state != "Queued" and !@todo_job_retries.include?(jobrecord.uuid)
690 j_done[:job_auth].update_attributes expires_at: Time.now
693 @running.delete job_done.uuid
697 expire_tokens = @pipe_auth_tokens.dup
698 @todo_pipelines.each do |p|
699 pipe_auth = (@pipe_auth_tokens[p.uuid] ||= ApiClientAuthorization.
700 create(user: User.where('uuid=?', p.modified_by_user_uuid).first,
702 puts `export ARVADOS_API_TOKEN=#{pipe_auth.api_token} && arv-run-pipeline-instance --run-pipeline-here --no-wait --instance #{p.uuid}`
703 expire_tokens.delete p.uuid
706 expire_tokens.each do |k, v|
707 v.update_attributes expires_at: Time.now
708 @pipe_auth_tokens.delete k
714 (argv.any? ? argv : ['--jobs', '--pipelines']).each do |arg|
717 @runoptions[:jobs] = true
719 @runoptions[:pipelines] = true
721 abort "Unrecognized command line option '#{arg}'"
724 if not (@runoptions[:jobs] or @runoptions[:pipelines])
725 abort "Nothing to do. Please specify at least one of: --jobs, --pipelines."
732 # We want files written by crunch-dispatch to be writable by other
733 # processes with the same GID, see bug #7228
736 # This is how crunch-job child procs know where the "refresh"
738 ENV["CRUNCH_REFRESH_TRIGGER"] = Rails.configuration.crunch_refresh_trigger
740 # If salloc can't allocate resources immediately, make it use our
741 # temporary failure exit code. This ensures crunch-dispatch won't
742 # mark a job failed because of an issue with node allocation.
743 # This often happens when another dispatcher wins the race to
745 ENV["SLURM_EXIT_IMMEDIATE"] = CrunchDispatch::EXIT_TEMPFAIL.to_s
747 if ENV["CRUNCH_DISPATCH_LOCKFILE"]
748 lockfilename = ENV.delete "CRUNCH_DISPATCH_LOCKFILE"
749 lockfile = File.open(lockfilename, File::RDWR|File::CREAT, 0644)
750 unless lockfile.flock File::LOCK_EX|File::LOCK_NB
751 abort "Lock unavailable on #{lockfilename} - exit"
756 %w{TERM INT}.each do |sig|
759 $stderr.puts "Received #{signame} signal"
760 @signal[:term] = true
765 User.first.group_permissions
766 $stderr.puts "dispatch: ready"
767 while !@signal[:term] or @running.size > 0
770 @running.each do |uuid, j|
771 if !j[:started] and j[:sent_int] < 2
773 Process.kill 'INT', j[:wait_thr].pid
775 # No such pid = race condition + desired result is
782 refresh_todo unless did_recently(:refresh_todo, 1.0)
783 update_node_status unless did_recently(:update_node_status, 1.0)
784 unless @todo.empty? or did_recently(:start_jobs, 1.0) or @signal[:term]
787 unless (@todo_pipelines.empty? and @pipe_auth_tokens.empty?) or did_recently(:update_pipelines, 5.0)
792 select(@running.values.collect { |j| [j[:stdout], j[:stderr]] }.flatten,
795 # If there are jobs we wanted to retry, we have to mark them as failed now.
796 # Other dispatchers can't pick them up because we hold their lock.
797 @todo_job_retries.each_key do |job_uuid|
798 job = Job.find_by_uuid(job_uuid)
799 if job.state == "Running"
800 fail_job(job, "crunch-dispatch was stopped during job's tempfail retry loop")
805 def fail_jobs before: nil
806 act_as_system_user do
808 if before == 'reboot'
810 open('/proc/stat').map(&:split).each do |stat, t|
816 raise "Could not find btime in /proc/stat"
818 threshold = Time.at(boottime.to_i)
820 threshold = Time.parse(before, Time.now)
822 threshold = db_current_time
824 Rails.logger.info "fail_jobs: threshold is #{threshold}"
826 if Rails.configuration.crunch_job_wrapper == :slurm_immediate
827 # [["slurm_job_id", "slurm_job_name"], ...]
828 squeue = File.popen(['squeue', '-h', '-o', '%i %j']).readlines.map do |line|
829 line.strip.split(' ', 2)
835 Job.where('state = ? and started_at < ?', Job::Running, threshold).
837 Rails.logger.debug "fail_jobs: #{job.uuid} started #{job.started_at}"
838 squeue.each do |slurm_id, slurm_name|
839 if slurm_name == job.uuid
840 Rails.logger.info "fail_jobs: scancel #{slurm_id} for #{job.uuid}"
844 fail_job(job, "cleaned up stale job: started before #{threshold}",
852 def have_job_lock?(job)
853 # Return true if the given job is locked by this crunch-dispatch, normally
854 # because we've run crunch-job for it.
855 @todo_job_retries.include?(job.uuid)
858 def did_recently(thing, min_interval)
859 if !@did_recently[thing] or @did_recently[thing] < Time.now - min_interval
860 @did_recently[thing] = Time.now
867 # send message to log table. we want these records to be transient
868 def write_log running_job
869 return if running_job[:stderr_buf_to_flush] == ''
871 # Send out to log event if buffer size exceeds the bytes per event or if
872 # it has been at least crunch_log_seconds_between_events seconds since
874 if running_job[:stderr_buf_to_flush].size > Rails.configuration.crunch_log_bytes_per_event or
875 (Time.now - running_job[:stderr_flushed_at]) >= Rails.configuration.crunch_log_seconds_between_events
877 log = Log.new(object_uuid: running_job[:job].uuid,
878 event_type: 'stderr',
879 owner_uuid: running_job[:job].owner_uuid,
880 properties: {"text" => running_job[:stderr_buf_to_flush]})
882 running_job[:events_logged] += 1
884 $stderr.puts "Failed to write logs"
885 $stderr.puts exception.backtrace
887 running_job[:stderr_buf_to_flush] = ''
888 running_job[:stderr_flushed_at] = Time.now
893 cmd = sudo_preface + ['scancel', slurm_id]
894 puts File.popen(cmd).read
896 Rails.logger.error "scancel #{slurm_id.shellescape}: $?"
901 return [] if not Server::Application.config.crunch_job_user
903 Server::Application.config.crunch_job_user,
904 "LD_LIBRARY_PATH=#{ENV['LD_LIBRARY_PATH']}",
905 "PATH=#{ENV['PATH']}",
906 "PERLLIB=#{ENV['PERLLIB']}",
907 "PYTHONPATH=#{ENV['PYTHONPATH']}",
908 "RUBYLIB=#{ENV['RUBYLIB']}",
909 "GEM_PATH=#{ENV['GEM_PATH']}"]