X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/274ca7a6c2f1f600de4242eee6cb4e8465d4440c..e910d13fc5fc63a86f20be3b758c08db3d429bc6:/sdk/cli/bin/crunch-job

diff --git a/sdk/cli/bin/crunch-job b/sdk/cli/bin/crunch-job
index 613be70ef1..f56099d1ad 100755
--- a/sdk/cli/bin/crunch-job
+++ b/sdk/cli/bin/crunch-job
@@ -161,6 +161,10 @@ if ($job_has_uuid)
       Log(undef, "Job is locked by " . $Job->{'is_locked_by_uuid'});
       exit EX_TEMPFAIL;
     }
+    if ($Job->{'state'} ne 'Queued') {
+      Log(undef, "Job state is " . $Job->{'state'} . ", but I can only start queued jobs.");
+      exit EX_TEMPFAIL;
+    }
     if ($Job->{'success'} ne undef) {
       Log(undef, "Job 'success' flag (" . $Job->{'success'} . ") is not null");
       exit EX_TEMPFAIL;
@@ -287,9 +291,7 @@ if ($job_has_uuid)
     Log(undef, "Error while updating / locking job, exiting ".EX_TEMPFAIL);
     exit EX_TEMPFAIL;
   }
-  $Job->update_attributes('started_at' => scalar gmtime,
-                          'running' => 1,
-                          'success' => undef,
+  $Job->update_attributes('state' => 'Running',
                           'tasks_summary' => { 'failed' => 0,
                                                'todo' => 1,
                                                'running' => 0,
@@ -876,12 +878,14 @@ Log (undef, "finish");
 save_meta();
 
 if ($job_has_uuid) {
-  $Job->update_attributes('running' => 0,
-                          'success' => $collated_output && $main::success,
-                          'finished_at' => scalar gmtime)
+  if ($collated_output && $main::success) {
+    $Job->update_attributes('state' => 'Complete')
+  } else {
+    $Job->update_attributes('state' => 'Failed')
+  }
 }
 
-exit ($Job->{'success'} ? 1 : 0);
+exit ($Job->{'state'} != 'Complete' ? 1 : 0);
 
 
 
@@ -1033,12 +1037,16 @@ sub check_refresh_wanted
       my $Job2 = $arv->{'jobs'}->{'get'}->execute('uuid' => $jobspec);
       for my $attr ('cancelled_at',
                     'cancelled_by_user_uuid',
-                    'cancelled_by_client_uuid') {
+                    'cancelled_by_client_uuid',
+                    'state') {
         $Job->{$attr} = $Job2->{$attr};
       }
-      if ($Job->{'cancelled_at'}) {
-        Log (undef, "Job cancelled at " . $Job->{cancelled_at} .
-             " by user " . $Job->{cancelled_by_user_uuid});
+      if ($Job->{'state'} ne "Running") {
+        if ($Job->{'state'} eq "Cancelled") {
+          Log (undef, "Job cancelled at " . $Job->{'cancelled_at'} . " by user " . $Job->{'cancelled_by_user_uuid'});
+        } else {
+          Log (undef, "Job state unexpectedly changed to " . $Job->{'state'});
+        }
         $main::success = 0;
         $main::please_freeze = 1;
       }
@@ -1207,7 +1215,8 @@ sub collate_output
   Log (undef, "collate");
 
   my ($child_out, $child_in);
-  my $pid = open2($child_out, $child_in, 'arv-put', '--raw');
+  my $pid = open2($child_out, $child_in, 'arv-put', '--raw',
+                  '--retries', put_retry_count());
   my $joboutput;
   for (@jobstep)
   {
@@ -1335,9 +1344,11 @@ sub croak
 sub cleanup
 {
   return if !$job_has_uuid;
-  $Job->update_attributes('running' => 0,
-                          'success' => 0,
-                          'finished_at' => scalar gmtime);
+  if ($Job->{'state'} eq 'Cancelled') {
+    $Job->update_attributes('finished_at' => scalar gmtime);
+  } else {
+    $Job->update_attributes('state' => 'Failed');
+  }
 }
 
 
@@ -1347,8 +1358,9 @@ sub save_meta
   return if $justcheckpoint;  # checkpointing is not relevant post-Warehouse.pm
 
   $local_logfile->flush;
-  my $cmd = "arv-put --portable-data-hash --filename ''\Q$keep_logfile\E "
-      . quotemeta($local_logfile->filename);
+  my $retry_count = put_retry_count();
+  my $cmd = "arv-put --portable-data-hash --retries $retry_count " .
+      "--filename ''\Q$keep_logfile\E " . quotemeta($local_logfile->filename);
   my $loglocator = `$cmd`;
   die "system $cmd failed: $?" if $?;
   chomp($loglocator);
@@ -1494,6 +1506,20 @@ sub find_docker_image {
   }
 }
 
+sub put_retry_count {
+  # Calculate a --retries argument for arv-put that will have it try
+  # approximately as long as this Job has been running.
+  my $stoptime = shift || time;
+  my $starttime = $jobstep[0]->{starttime};
+  my $timediff = defined($starttime) ? ($stoptime - $starttime) : 1;
+  my $retries = 0;
+  while ($timediff >= 2) {
+    $retries++;
+    $timediff /= 2;
+  }
+  return ($retries > 3) ? $retries : 3;
+}
+
 __DATA__
 #!/usr/bin/perl