@@ -347,7 +347,13 @@ def _requeue_failed_jobs(self, job_list: List[V0040JobInfo]) -> bool:
347347 )
348348 logging .warning (
349349 f"Restarting { job .job_state } job: { job_id_to_retry } . "
350- f"Restart count: { restart_count } "
350+ f"Restart count: { restart_count } " ,
351+ extra = {
352+ "process_name" : "submit_slurm_job_array" ,
353+ "pipeline_name" : "airflow DAG" ,
354+ "job_id" : job_id_to_retry ,
355+ "job_state" : job .job_state ,
356+ },
351357 )
352358 sleep (1 )
353359 command = f"scontrol requeue { job_id_to_retry } "
@@ -463,11 +469,27 @@ def _monitor_job(
463469 last_backfill = V0040Uint64NoVal (),
464470 last_update = V0040Uint64NoVal (),
465471 )
466- logging .debug (f"Initialized placeholder: { job_response } " )
472+ logging .debug (
473+ f"Initialized placeholder: { job_response } " ,
474+ extra = {
475+ "process_name" : "monitor_slurm_job" ,
476+ "pipeline_name" : "airflow DAG" ,
477+ "acquisition_name" : job_name ,
478+ "job_id" : job_id ,
479+ },
480+ )
467481 try :
468482 job_response = self .slurm .slurm_v0040_get_job (job_id = str (job_id ))
469483 except NotFoundException :
470- logging .warning ("Looking for job info in database..." )
484+ logging .warning (
485+ "Looking for job info in database..." ,
486+ extra = {
487+ "process_name" : "monitor_slurm_job" ,
488+ "pipeline_name" : "airflow DAG" ,
489+ "acquisition_name" : job_name ,
490+ "job_id" : job_id ,
491+ },
492+ )
471493 slurm_db_api = SlurmdbApi (api_client = self .slurm .api_client )
472494 slurm_db_response = slurm_db_api .slurmdb_v0040_get_job (
473495 job_id = str (job_id )
@@ -509,7 +531,15 @@ def _monitor_job(
509531 "start_time" : start_time ,
510532 }
511533 )
512- logging .info (message )
534+ logging .info (
535+ message ,
536+ extra = {
537+ "process_name" : "monitor_slurm_job" ,
538+ "pipeline_name" : "airflow DAG" ,
539+ "acquisition_name" : job_name ,
540+ "job_id" : job_id ,
541+ },
542+ )
513543 job_status = dict ()
514544 is_finished , is_error = self ._check_job_status (
515545 job_response , job_status
@@ -543,7 +573,15 @@ def _monitor_job(
543573 "start_time" : start_time ,
544574 }
545575 )
546- logging .info (message )
576+ logging .info (
577+ message ,
578+ extra = {
579+ "process_name" : "monitor_slurm_job" ,
580+ "pipeline_name" : "airflow DAG" ,
581+ "acquisition_name" : job_name ,
582+ "job_id" : job_id ,
583+ },
584+ )
547585 is_finished , is_error = self ._check_job_status (
548586 job_response , job_status
549587 )
@@ -562,14 +600,30 @@ def _monitor_job(
562600 remote_mnt_dir = self .remote_mnt_dir ,
563601 local_mnt_dir = self .local_mnt_dir ,
564602 )
565- logging .exception (f"std_err:\n { std_err_msg } " )
603+ logging .exception (
604+ f"std_err:\n { std_err_msg } " ,
605+ extra = {
606+ "process_name" : "monitor_slurm_job" ,
607+ "pipeline_name" : "airflow DAG" ,
608+ "acquisition_name" : job_name ,
609+ "job_id" : job_id ,
610+ },
611+ )
566612 raise Exception (
567613 f"There were errors with the slurm job. "
568614 f"Job: { message } . "
569615 f"Errors: { errors } "
570616 )
571617 else :
572- logging .info ("Job is Finished!" )
618+ logging .info (
619+ "Job is Finished!" ,
620+ extra = {
621+ "process_name" : "monitor_slurm_job" ,
622+ "pipeline_name" : "airflow DAG" ,
623+ "acquisition_name" : job_name ,
624+ "job_id" : job_id ,
625+ },
626+ )
573627 return start_time , end_time
574628
575629 def _std_err_filepath (self , job_id : Union [int , str ]) -> str :
@@ -595,10 +649,34 @@ def run_job(self):
595649 submit_response = self ._submit_job ()
596650 job_id = submit_response .job_id
597651 job_name = self .job_properties .name
598- logging .info (f"Job Name: { job_name } " )
599- logging .info (f"Job ID: { job_id } " )
652+ logging .info (
653+ f"Job Name: { job_name } " ,
654+ extra = {
655+ "process_name" : "run_slurm_job" ,
656+ "pipeline_name" : "airflow DAG" ,
657+ "acquisition_name" : job_name ,
658+ "job_id" : job_id ,
659+ },
660+ )
661+ logging .info (
662+ f"Job ID: { job_id } " ,
663+ extra = {
664+ "process_name" : "run_slurm_job" ,
665+ "pipeline_name" : "airflow DAG" ,
666+ "acquisition_name" : job_name ,
667+ "job_id" : job_id ,
668+ },
669+ )
600670 std_err = self ._std_err_filepath (job_id = job_id )
601- logging .info (f"Please check { std_err } for additional logs." )
671+ logging .info (
672+ f"Please check { std_err } for additional logs." ,
673+ extra = {
674+ "process_name" : "run_slurm_job" ,
675+ "pipeline_name" : "airflow DAG" ,
676+ "acquisition_name" : job_name ,
677+ "job_id" : job_id ,
678+ },
679+ )
602680 self ._monitor_job (submit_response = submit_response )
603681
604682
@@ -657,7 +735,13 @@ def _requeue_failed_jobs(self, job_list: List[V0040JobInfo]) -> bool:
657735 )
658736 logging .warning (
659737 f"Restarting { job .job_state } job: { job_id_to_retry } . "
660- f"Restart count: { restart_count } "
738+ f"Restart count: { restart_count } " ,
739+ extra = {
740+ "process_name" : "slurm_job_sensor" ,
741+ "pipeline_name" : "airflow DAG" ,
742+ "job_id" : job_id_to_retry ,
743+ "job_state" : job .job_state ,
744+ },
661745 )
662746 sleep (1 )
663747 command = f"scontrol requeue { job_id_to_retry } "
@@ -751,15 +835,29 @@ def get_job_status(self) -> Tuple[bool, int, int]:
751835 remote_mnt_dir = self .remote_mnt_dir ,
752836 local_mnt_dir = self .local_mnt_dir ,
753837 )
754- logging .exception (f"std_err:\n { std_err_msg } " )
838+ logging .exception (
839+ f"std_err:\n { std_err_msg } " ,
840+ extra = {
841+ "process_name" : "slurm_job_sensor" ,
842+ "pipeline_name" : "airflow DAG" ,
843+ "job_id" : self .job_id ,
844+ },
845+ )
755846 raise Exception (
756847 f"There were errors with the slurm job. "
757848 f"Job: { self .job_id } . "
758849 f"Errors: { errors } "
759850 )
760851 return is_finished , start_time , end_time
761852 except NotFoundException :
762- logging .warning ("Looking for job info in database..." )
853+ logging .warning (
854+ "Looking for job info in database..." ,
855+ extra = {
856+ "process_name" : "slurm_job_sensor" ,
857+ "pipeline_name" : "airflow DAG" ,
858+ "job_id" : job_id ,
859+ },
860+ )
763861 slurm_db_api = SlurmdbApi (api_client = self .slurm .api_client )
764862 slurm_db_response = slurm_db_api .slurmdb_v0040_get_job (
765863 job_id = str (job_id )
0 commit comments