3 # We want files written by crunch-dispatch to be writable by other processes
4 # with the same GID, see bug #7228
11 (ARGV.any? ? ARGV : ['--jobs', '--pipelines']).each do |arg|
14 $options[:jobs] = true
16 $options[:pipelines] = true
18 abort "Unrecognized command line option '#{arg}'"
21 if not ($options[:jobs] or $options[:pipelines])
22 abort "Nothing to do. Please specify at least one of: --jobs, --pipelines."
25 ARGV.reject! { |a| a =~ /--jobs|--pipelines/ }
29 %w{TERM INT}.each do |sig|
32 $stderr.puts "Received #{signame} signal"
37 if ENV["CRUNCH_DISPATCH_LOCKFILE"]
38 lockfilename = ENV.delete "CRUNCH_DISPATCH_LOCKFILE"
39 lockfile = File.open(lockfilename, File::RDWR|File::CREAT, 0644)
40 unless lockfile.flock File::LOCK_EX|File::LOCK_NB
41 abort "Lock unavailable on #{lockfilename} - exit"
45 ENV["RAILS_ENV"] = ARGV[0] || ENV["RAILS_ENV"] || "development"
47 require File.dirname(__FILE__) + '/../config/boot'
48 require File.dirname(__FILE__) + '/../config/environment'
53 self.utc.strftime "%Y-%m-%d_%H:%M:%S"
58 include ApplicationHelper
61 EXIT_RETRY_UNLOCKED = 93
62 RETRY_UNLOCKED_LIMIT = 3
65 @crunch_job_bin = (ENV['CRUNCH_JOB_BIN'] || `which arv-crunch-job`.strip)
66 if @crunch_job_bin.empty?
67 raise "No CRUNCH_JOB_BIN env var, and crunch-job not in path."
70 @docker_bin = ENV['CRUNCH_JOB_DOCKER_BIN']
72 @arvados_internal = Rails.configuration.git_internal_dir
73 if not File.exists? @arvados_internal
74 $stderr.puts `mkdir -p #{@arvados_internal.shellescape} && git init --bare #{@arvados_internal.shellescape}`
75 raise "No internal git repository available" unless ($? == 0)
78 @repo_root = Rails.configuration.git_repositories_dir
79 @arvados_repo_path = Repository.where(name: "arvados").first.server_path
85 @pipe_auth_tokens = {}
88 @todo_job_retries = {}
89 @job_retry_counts = Hash.new(0)
94 return act_as_system_user
99 @todo = @todo_job_retries.values + Job.queue.select(&:repository)
101 if $options[:pipelines]
102 @todo_pipelines = PipelineInstance.queue
106 def each_slurm_line(cmd, outfmt, max_fields=nil)
107 max_fields ||= outfmt.split(":").size
108 max_fields += 1 # To accommodate the node field we add
109 @@slurm_version ||= Gem::Version.new(`sinfo --version`.match(/\b[\d\.]+\b/)[0])
110 if Gem::Version.new('2.3') <= @@slurm_version
111 `#{cmd} --noheader -o '%n:#{outfmt}'`.each_line do |line|
112 yield line.chomp.split(":", max_fields)
115 # Expand rows with hostname ranges (like "foo[1-3,5,9-12]:idle")
116 # into multiple rows with one hostname each.
117 `#{cmd} --noheader -o '%N:#{outfmt}'`.each_line do |line|
118 tokens = line.chomp.split(":", max_fields)
119 if (re = tokens[0].match /^(.*?)\[([-,\d]+)\]$/)
121 re[2].split(",").each do |range|
122 range = range.split("-").collect(&:to_i)
123 (range[0]..range[-1]).each do |n|
124 yield [re[1] + n.to_s] + tokens
136 each_slurm_line("sinfo", "%t") do |hostname, state|
137 # Treat nodes in idle* state as down, because the * means that slurm
138 # hasn't been able to communicate with it recently.
139 state.sub!(/^idle\*/, "down")
140 state.sub!(/\W+$/, "")
141 state = "down" unless %w(idle alloc down).include?(state)
142 slurm_nodes[hostname] = {state: state, job: nil}
144 each_slurm_line("squeue", "%j") do |hostname, job_uuid|
145 slurm_nodes[hostname][:job] = job_uuid if slurm_nodes[hostname]
150 def update_node_status
151 return unless Server::Application.config.crunch_job_wrapper.to_s.match /^slurm/
152 slurm_status.each_pair do |hostname, slurmdata|
153 next if @node_state[hostname] == slurmdata
155 node = Node.where('hostname=?', hostname).order(:last_ping_at).last
157 $stderr.puts "dispatch: update #{hostname} state to #{slurmdata}"
158 node.info["slurm_state"] = slurmdata[:state]
159 node.job_uuid = slurmdata[:job]
161 @node_state[hostname] = slurmdata
163 $stderr.puts "dispatch: failed to update #{node.uuid}: #{node.errors.messages}"
165 elsif slurmdata[:state] != 'down'
166 $stderr.puts "dispatch: SLURM reports '#{hostname}' is not down, but no node has that name"
169 $stderr.puts "dispatch: error updating #{hostname} node status: #{error}"
174 def positive_int(raw_value, default=nil)
175 value = begin raw_value.to_i rescue 0 end
183 NODE_CONSTRAINT_MAP = {
184 # Map Job runtime_constraints keys to the corresponding Node info key.
185 'min_ram_mb_per_node' => 'total_ram_mb',
186 'min_scratch_mb_per_node' => 'total_scratch_mb',
187 'min_cores_per_node' => 'total_cpu_cores',
190 def nodes_available_for_job_now(job)
191 # Find Nodes that satisfy a Job's runtime constraints (by building
192 # a list of Procs and using them to test each Node). If there
193 # enough to run the Job, return an array of their names.
194 # Otherwise, return nil.
195 need_procs = NODE_CONSTRAINT_MAP.each_pair.map do |job_key, node_key|
197 positive_int(node.properties[node_key], 0) >=
198 positive_int(job.runtime_constraints[job_key], 0)
201 min_node_count = positive_int(job.runtime_constraints['min_nodes'], 1)
203 Node.all.select do |node|
204 node.info['slurm_state'] == 'idle'
205 end.sort_by do |node|
206 # Prefer nodes with no price, then cheap nodes, then expensive nodes
207 node.properties['cloud_node']['price'].to_f rescue 0
209 if need_procs.select { |node_test| not node_test.call(node) }.any?
213 if usable_nodes.count >= min_node_count
214 return usable_nodes.map { |node| node.hostname }
220 def nodes_available_for_job(job)
221 # Check if there are enough idle nodes with the Job's minimum
222 # hardware requirements to run it. If so, return an array of
223 # their names. If not, up to once per hour, signal start_jobs to
224 # hold off launching Jobs. This delay is meant to give the Node
225 # Manager an opportunity to make new resources available for new
228 # The exact timing parameters here might need to be adjusted for
229 # the best balance between helping the longest-waiting Jobs run,
230 # and making efficient use of immediately available resources.
231 # These are all just first efforts until we have more data to work
233 nodelist = nodes_available_for_job_now(job)
234 if nodelist.nil? and not did_recently(:wait_for_available_nodes, 3600)
235 $stderr.puts "dispatch: waiting for nodes for #{job.uuid}"
236 @node_wait_deadline = Time.now + 5.minutes
241 def fail_job job, message
242 $stderr.puts "dispatch: #{job.uuid}: #{message}"
244 Log.new(object_uuid: job.uuid,
245 event_type: 'dispatch',
246 owner_uuid: job.owner_uuid,
248 properties: {"text" => message}).save!
250 $stderr.puts "dispatch: log.create failed"
254 job.lock @authorizations[job.uuid].user.uuid
257 $stderr.puts "dispatch: save failed setting job #{job.uuid} to failed"
259 rescue ArvadosModel::AlreadyLockedError
260 $stderr.puts "dispatch: tried to mark job #{job.uuid} as failed but it was already locked by someone else"
264 def stdout_s(cmd_a, opts={})
265 IO.popen(cmd_a, "r", opts) do |pipe|
266 return pipe.read.chomp
271 ["git", "--git-dir=#{@arvados_internal}"] + cmd_a
274 def get_authorization(job)
275 if @authorizations[job.uuid] and
276 @authorizations[job.uuid].user.uuid != job.modified_by_user_uuid
277 # We already made a token for this job, but we need a new one
278 # because modified_by_user_uuid has changed (the job will run
279 # as a different user).
280 @authorizations[job.uuid].update_attributes expires_at: Time.now
281 @authorizations[job.uuid] = nil
283 if not @authorizations[job.uuid]
284 auth = ApiClientAuthorization.
285 new(user: User.where('uuid=?', job.modified_by_user_uuid).first,
288 $stderr.puts "dispatch: auth.save failed for #{job.uuid}"
290 @authorizations[job.uuid] = auth
293 @authorizations[job.uuid]
296 def internal_repo_has_commit? sha1
297 if (not @fetched_commits[sha1] and
298 sha1 == stdout_s(git_cmd("rev-list", "-n1", sha1), err: "/dev/null") and
300 @fetched_commits[sha1] = true
302 return @fetched_commits[sha1]
305 def get_commit src_repo, sha1
306 return true if internal_repo_has_commit? sha1
308 # commit does not exist in internal repository, so import the
309 # source repository using git fetch-pack
310 cmd = git_cmd("fetch-pack", "--no-progress", "--all", src_repo)
311 $stderr.puts "dispatch: #{cmd}"
312 $stderr.puts(stdout_s(cmd))
313 @fetched_commits[sha1] = ($? == 0)
316 def tag_commit(commit_hash, tag_name)
317 # @git_tags[T]==V if we know commit V has been tagged T in the
318 # arvados_internal repository.
319 if not @git_tags[tag_name]
320 cmd = git_cmd("tag", tag_name, commit_hash)
321 $stderr.puts "dispatch: #{cmd}"
322 $stderr.puts(stdout_s(cmd, err: "/dev/null"))
324 # git tag failed. This may be because the tag already exists, so check for that.
325 tag_rev = stdout_s(git_cmd("rev-list", "-n1", tag_name))
327 # We got a revision back
328 if tag_rev != commit_hash
329 # Uh oh, the tag doesn't point to the revision we were expecting.
330 # Someone has been monkeying with the job record and/or git.
331 fail_job job, "Existing tag #{tag_name} points to commit #{tag_rev} but expected commit #{commit_hash}"
334 # we're okay (fall through to setting @git_tags below)
336 # git rev-list failed for some reason.
337 fail_job job, "'git tag' for #{tag_name} failed but did not find any existing tag using 'git rev-list'"
341 # 'git tag' was successful, or there is an existing tag that points to the same revision.
342 @git_tags[tag_name] = commit_hash
343 elsif @git_tags[tag_name] != commit_hash
344 fail_job job, "Existing tag #{tag_name} points to commit #{@git_tags[tag_name]} but this job uses commit #{commit_hash}"
352 next if @running[job.uuid]
355 case Server::Application.config.crunch_job_wrapper
358 # Don't run more than one at a time.
362 when :slurm_immediate
363 nodelist = nodes_available_for_job(job)
365 if Time.now < @node_wait_deadline
371 cmd_args = ["salloc",
376 "--job-name=#{job.uuid}",
377 "--nodelist=#{nodelist.join(',')}"]
379 raise "Unknown crunch_job_wrapper: #{Server::Application.config.crunch_job_wrapper}"
382 if Server::Application.config.crunch_job_user
383 cmd_args.unshift("sudo", "-E", "-u",
384 Server::Application.config.crunch_job_user,
385 "LD_LIBRARY_PATH=#{ENV['LD_LIBRARY_PATH']}",
386 "PATH=#{ENV['PATH']}",
387 "PERLLIB=#{ENV['PERLLIB']}",
388 "PYTHONPATH=#{ENV['PYTHONPATH']}",
389 "RUBYLIB=#{ENV['RUBYLIB']}",
390 "GEM_PATH=#{ENV['GEM_PATH']}")
393 next unless get_authorization job
395 ready = internal_repo_has_commit? job.script_version
398 # Import the commit from the specified repository into the
399 # internal repository. This should have been done already when
400 # the job was created/updated; this code is obsolete except to
401 # avoid deployment races. Failing the job would be a
402 # reasonable thing to do at this point.
403 repo = Repository.where(name: job.repository).first
404 if repo.nil? or repo.server_path.nil?
405 fail_job "Repository #{job.repository} not found under #{@repo_root}"
408 ready &&= get_commit repo.server_path, job.script_version
409 ready &&= tag_commit job.script_version, job.uuid
412 # This should be unnecessary, because API server does it during
413 # job create/update, but it's still not a bad idea to verify the
414 # tag is correct before starting the job:
415 ready &&= tag_commit job.script_version, job.uuid
417 # The arvados_sdk_version doesn't support use of arbitrary
418 # remote URLs, so the requested version isn't necessarily copied
419 # into the internal repository yet.
420 if job.arvados_sdk_version
421 ready &&= get_commit @arvados_repo_path, job.arvados_sdk_version
422 ready &&= tag_commit job.arvados_sdk_version, "#{job.uuid}-arvados-sdk"
426 fail_job job, "commit not present in internal repository"
430 cmd_args += [@crunch_job_bin,
431 '--job-api-token', @authorizations[job.uuid].api_token,
433 '--git-dir', @arvados_internal]
436 cmd_args += ['--docker-bin', @docker_bin]
439 if @todo_job_retries.include?(job.uuid)
440 cmd_args << "--force-unlock"
443 $stderr.puts "dispatch: #{cmd_args.join ' '}"
446 i, o, e, t = Open3.popen3(*cmd_args)
448 $stderr.puts "dispatch: popen3: #{$!}"
453 $stderr.puts "dispatch: job #{job.uuid}"
454 start_banner = "dispatch: child #{t.pid} start #{LogTime.now}"
455 $stderr.puts start_banner
457 @running[job.uuid] = {
463 buf: {stderr: '', stdout: ''},
466 job_auth: @authorizations[job.uuid],
467 stderr_buf_to_flush: '',
468 stderr_flushed_at: Time.new(0),
471 log_throttle_is_open: true,
472 log_throttle_reset_time: Time.now + Rails.configuration.crunch_log_throttle_period,
473 log_throttle_bytes_so_far: 0,
474 log_throttle_lines_so_far: 0,
475 log_throttle_bytes_skipped: 0,
478 @todo_job_retries.delete(job.uuid)
483 # Test for hard cap on total output and for log throttling. Returns whether
484 # the log line should go to output or not. Modifies "line" in place to
485 # replace it with an error if a logging limit is tripped.
486 def rate_limit running_job, line
489 if running_job[:log_throttle_is_open]
490 running_job[:log_throttle_lines_so_far] += 1
491 running_job[:log_throttle_bytes_so_far] += linesize
492 running_job[:bytes_logged] += linesize
494 if (running_job[:bytes_logged] >
495 Rails.configuration.crunch_limit_log_bytes_per_job)
496 message = "Exceeded log limit #{Rails.configuration.crunch_limit_log_bytes_per_job} bytes (crunch_limit_log_bytes_per_job). Log will be truncated."
497 running_job[:log_throttle_reset_time] = Time.now + 100.years
498 running_job[:log_throttle_is_open] = false
500 elsif (running_job[:log_throttle_bytes_so_far] >
501 Rails.configuration.crunch_log_throttle_bytes)
502 remaining_time = running_job[:log_throttle_reset_time] - Time.now
503 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"
504 running_job[:log_throttle_is_open] = false
506 elsif (running_job[:log_throttle_lines_so_far] >
507 Rails.configuration.crunch_log_throttle_lines)
508 remaining_time = running_job[:log_throttle_reset_time] - Time.now
509 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"
510 running_job[:log_throttle_is_open] = false
514 if not running_job[:log_throttle_is_open]
515 # Don't log anything if any limit has been exceeded. Just count lossage.
516 running_job[:log_throttle_bytes_skipped] += linesize
520 # Yes, write to logs, but use our "rate exceeded" message
521 # instead of the log message that exceeded the limit.
525 running_job[:log_throttle_is_open]
530 @running.each do |job_uuid, j|
534 if now > j[:log_throttle_reset_time]
535 # It has been more than throttle_period seconds since the last
536 # checkpoint so reset the throttle
537 if j[:log_throttle_bytes_skipped] > 0
538 message = "#{job_uuid} ! Skipped #{j[:log_throttle_bytes_skipped]} bytes of log"
540 j[:stderr_buf_to_flush] << "#{LogTime.now} #{message}\n"
543 j[:log_throttle_reset_time] = now + Rails.configuration.crunch_log_throttle_period
544 j[:log_throttle_bytes_so_far] = 0
545 j[:log_throttle_lines_so_far] = 0
546 j[:log_throttle_bytes_skipped] = 0
547 j[:log_throttle_is_open] = true
550 j[:buf].each do |stream, streambuf|
551 # Read some data from the child stream
554 # It's important to use a big enough buffer here. When we're
555 # being flooded with logs, we must read and discard many
556 # bytes at once. Otherwise, we can easily peg a CPU with
557 # time-checking and other loop overhead. (Quick tests show a
558 # 1MiB buffer working 2.5x as fast as a 64 KiB buffer.)
560 # So don't reduce this buffer size!
561 buf = j[stream].read_nonblock(2**20)
562 rescue Errno::EAGAIN, EOFError
565 # Short circuit the counting code if we're just going to throw
566 # away the data anyway.
567 if not j[:log_throttle_is_open]
568 j[:log_throttle_bytes_skipped] += streambuf.size + buf.size
575 # Append to incomplete line from previous read, if any
579 streambuf.each_line do |line|
580 if not line.end_with? $/
581 if line.size > Rails.configuration.crunch_log_throttle_bytes
582 # Without a limit here, we'll use 2x an arbitrary amount
583 # of memory, and waste a lot of time copying strings
584 # around, all without providing any feedback to anyone
585 # about what's going on _or_ hitting any of our throttle
588 # Here we leave "line" alone, knowing it will never be
589 # sent anywhere: rate_limit() will reach
590 # crunch_log_throttle_bytes immediately. However, we'll
591 # leave [...] in bufend: if the trailing end of the long
592 # line does end up getting sent anywhere, it will have
593 # some indication that it is incomplete.
596 # If line length is sane, we'll wait for the rest of the
597 # line to appear in the next read_pipes() call.
602 # rate_limit returns true or false as to whether to actually log
603 # the line or not. It also modifies "line" in place to replace
604 # it with an error if a logging limit is tripped.
605 if rate_limit j, line
606 $stderr.print "#{job_uuid} ! " unless line.index(job_uuid)
608 pub_msg = "#{LogTime.now} #{line.strip}\n"
609 j[:stderr_buf_to_flush] << pub_msg
613 # Leave the trailing incomplete line (if any) in streambuf for
615 streambuf.replace bufend
617 # Flush buffered logs to the logs table, if appropriate. We have
618 # to do this even if we didn't collect any new logs this time:
619 # otherwise, buffered data older than seconds_between_events
620 # won't get flushed until new data arrives.
626 return if 0 == @running.size
632 pid_done = waitpid(-1, Process::WNOHANG | Process::WUNTRACED)
634 j_done = @running.values.
635 select { |j| j[:wait_thr].pid == pid_done }.
638 rescue SystemCallError
639 # I have @running processes but system reports I have no
640 # children. This is likely to happen repeatedly if it happens at
641 # all; I will log this no more than once per child process I
643 if 0 < @running.select { |uuid,j| j[:warned_waitpid_error].nil? }.size
644 children = @running.values.collect { |j| j[:wait_thr].pid }.join ' '
645 $stderr.puts "dispatch: IPC bug: waitpid() error (#{$!}), but I have children #{children}"
647 @running.each do |uuid,j| j[:warned_waitpid_error] = true end
650 @running.each do |uuid, j|
651 if j[:wait_thr].status == false
652 pid_done = j[:wait_thr].pid
660 job_done = j_done[:job]
662 # Ensure every last drop of stdout and stderr is consumed.
664 # Reset flush timestamp to make sure log gets written.
665 j_done[:stderr_flushed_at] = Time.new(0)
666 # Write any remaining logs.
669 j_done[:buf].each do |stream, streambuf|
671 $stderr.puts streambuf + "\n"
675 # Wait the thread (returns a Process::Status)
676 exit_status = j_done[:wait_thr].value.exitstatus
677 exit_tempfail = exit_status == EXIT_TEMPFAIL
679 $stderr.puts "dispatch: child #{pid_done} exit #{exit_status}"
680 $stderr.puts "dispatch: job #{job_done.uuid} end"
682 jobrecord = Job.find_by_uuid(job_done.uuid)
684 if exit_status == EXIT_RETRY_UNLOCKED
685 # The job failed because all of the nodes allocated to it
686 # failed. Only this crunch-dispatch process can retry the job:
687 # it's already locked, and there's no way to put it back in the
688 # Queued state. Put it in our internal todo list unless the job
689 # has failed this way excessively.
690 @job_retry_counts[jobrecord.uuid] += 1
691 exit_tempfail = @job_retry_counts[jobrecord.uuid] <= RETRY_UNLOCKED_LIMIT
693 @todo_job_retries[jobrecord.uuid] = jobrecord
695 $stderr.puts("dispatch: job #{jobrecord.uuid} exceeded node failure retry limit -- giving up")
700 @job_retry_counts.delete(jobrecord.uuid)
701 if jobrecord.state == "Running"
702 # Apparently there was an unhandled error. That could potentially
703 # include "all allocated nodes failed" when we don't to retry
704 # because the job has already been retried RETRY_UNLOCKED_LIMIT
705 # times. Fail the job.
706 jobrecord.state = "Failed"
707 if not jobrecord.save
708 $stderr.puts "dispatch: jobrecord.save failed"
712 # If the job failed to run due to an infrastructure
713 # issue with crunch-job or slurm, we want the job to stay in the
714 # queue. If crunch-job exited after losing a race to another
715 # crunch-job process, it exits 75 and we should leave the job
716 # record alone so the winner of the race can do its thing.
717 # If crunch-job exited after all of its allocated nodes failed,
718 # it exits 93, and we want to retry it later (see the
719 # EXIT_RETRY_UNLOCKED `if` block).
721 # There is still an unhandled race condition: If our crunch-job
722 # process is about to lose a race with another crunch-job
723 # process, but crashes before getting to its "exit 75" (for
724 # example, "cannot fork" or "cannot reach API server") then we
725 # will assume incorrectly that it's our process's fault
726 # jobrecord.started_at is non-nil, and mark the job as failed
727 # even though the winner of the race is probably still doing
731 # Invalidate the per-job auth token, unless the job is still queued and we
732 # might want to try it again.
733 if jobrecord.state != "Queued" and !@todo_job_retries.include?(jobrecord.uuid)
734 j_done[:job_auth].update_attributes expires_at: Time.now
737 @running.delete job_done.uuid
741 expire_tokens = @pipe_auth_tokens.dup
742 @todo_pipelines.each do |p|
743 pipe_auth = (@pipe_auth_tokens[p.uuid] ||= ApiClientAuthorization.
744 create(user: User.where('uuid=?', p.modified_by_user_uuid).first,
746 puts `export ARVADOS_API_TOKEN=#{pipe_auth.api_token} && arv-run-pipeline-instance --run-pipeline-here --no-wait --instance #{p.uuid}`
747 expire_tokens.delete p.uuid
750 expire_tokens.each do |k, v|
751 v.update_attributes expires_at: Time.now
752 @pipe_auth_tokens.delete k
758 User.first.group_permissions
759 $stderr.puts "dispatch: ready"
760 while !$signal[:term] or @running.size > 0
763 @running.each do |uuid, j|
764 if !j[:started] and j[:sent_int] < 2
766 Process.kill 'INT', j[:wait_thr].pid
768 # No such pid = race condition + desired result is
775 refresh_todo unless did_recently(:refresh_todo, 1.0)
776 update_node_status unless did_recently(:update_node_status, 1.0)
777 unless @todo.empty? or did_recently(:start_jobs, 1.0) or $signal[:term]
780 unless (@todo_pipelines.empty? and @pipe_auth_tokens.empty?) or did_recently(:update_pipelines, 5.0)
785 select(@running.values.collect { |j| [j[:stdout], j[:stderr]] }.flatten,
788 # If there are jobs we wanted to retry, we have to mark them as failed now.
789 # Other dispatchers can't pick them up because we hold their lock.
790 @todo_job_retries.each_key do |job_uuid|
791 job = Job.find_by_uuid(job_uuid)
792 if job.state == "Running"
793 fail_job(job, "crunch-dispatch was stopped during job's tempfail retry loop")
800 def did_recently(thing, min_interval)
801 if !@did_recently[thing] or @did_recently[thing] < Time.now - min_interval
802 @did_recently[thing] = Time.now
809 # send message to log table. we want these records to be transient
810 def write_log running_job
811 return if running_job[:stderr_buf_to_flush] == ''
813 # Send out to log event if buffer size exceeds the bytes per event or if
814 # it has been at least crunch_log_seconds_between_events seconds since
816 if running_job[:stderr_buf_to_flush].size > Rails.configuration.crunch_log_bytes_per_event or
817 (Time.now - running_job[:stderr_flushed_at]) >= Rails.configuration.crunch_log_seconds_between_events
819 log = Log.new(object_uuid: running_job[:job].uuid,
820 event_type: 'stderr',
821 owner_uuid: running_job[:job].owner_uuid,
822 properties: {"text" => running_job[:stderr_buf_to_flush]})
824 running_job[:events_logged] += 1
826 $stderr.puts "Failed to write logs"
827 $stderr.puts exception.backtrace
829 running_job[:stderr_buf_to_flush] = ''
830 running_job[:stderr_flushed_at] = Time.now
835 # This is how crunch-job child procs know where the "refresh" trigger file is
836 ENV["CRUNCH_REFRESH_TRIGGER"] = Rails.configuration.crunch_refresh_trigger
838 # If salloc can't allocate resources immediately, make it use our temporary
839 # failure exit code. This ensures crunch-dispatch won't mark a job failed
840 # because of an issue with node allocation. This often happens when
841 # another dispatcher wins the race to allocate nodes.
842 ENV["SLURM_EXIT_IMMEDIATE"] = Dispatcher::EXIT_TEMPFAIL.to_s