Merge branch '12533-token-n-ip-logging'

[arvados.git] / sdk / cli / bin / crunch-job

diff --git a/sdk/cli/bin/crunch-job b/sdk/cli/bin/crunch-job
index 5a92176e7f02fba11525190ccee511339819e1d2..f2e9fc2878c4ded51ce3069e15620061eabf5910 100755 (executable)
--- a/sdk/cli/bin/crunch-job
+++ b/sdk/cli/bin/crunch-job
@@ -837,8 +837,8 @@ for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
       close($_);
     }
     fcntl ("writer", F_SETFL, 0) or croak ($!); # no close-on-exec
-    open(STDOUT,">&writer");
-    open(STDERR,">&writer");
+    open(STDOUT,">&writer") or croak ($!);
+    open(STDERR,">&writer") or croak ($!);
 
     undef $dbh;
     undef $sth;
@@ -1022,7 +1022,7 @@ for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
   delete $Jobstep->{tempfail};
 
   $Jobstep->{'arvados_task'}->{started_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{starttime});
-  $Jobstep->{'arvados_task'}->save;
+  retry_op(sub { $Jobstep->{'arvados_task'}->save; }, "job_tasks.update API");
 
   splice @jobstep_todo, $todo_ptr, 1;
   --$todo_ptr;
@@ -1205,7 +1205,7 @@ sub reapchildren
             "ERROR: Task process exited %s, but never updated its task record to indicate success and record its output.",
             exit_status_s($childstatus)));
       $Jobstep->{'arvados_task'}->{success} = 0;
-      $Jobstep->{'arvados_task'}->save;
+      retry_op(sub { $Jobstep->{'arvados_task'}->save; }, "job_tasks.update API");
       $task_success = 0;
     }
 
@@ -1258,7 +1258,7 @@ sub reapchildren
     $Jobstep->{exitcode} = $childstatus;
     $Jobstep->{finishtime} = time;
     $Jobstep->{'arvados_task'}->{finished_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{finishtime});
-    $Jobstep->{'arvados_task'}->save;
+    retry_op(sub { $Jobstep->{'arvados_task'}->save; }, "job_tasks.update API");
     process_stderr_final ($jobstepidx);
     Log ($jobstepidx, sprintf("task output (%d bytes): %s",
                               length($Jobstep->{'arvados_task'}->{output}),
@@ -1544,7 +1544,7 @@ sub preprocess_stderr
         $st->{node}->{fail_count}++;
       }
     }
-    elsif ($line =~ /srun: error: .*?\b(Node failure on|Aborting, .*?\bio error\b)/i) {
+    elsif ($line =~ /srun: error: .*?\b(Node failure on|Aborting, .*?\bio error\b|cannot communicate with node .* aborting job)/i) {
       $jobstep[$jobstepidx]->{tempfail} = 1;
       if (defined($job_slot_index)) {
         $slot[$job_slot_index]->{node}->{fail_count}++;
@@ -1996,8 +1996,8 @@ sub srun_sync
     close($stdout_r);
     fcntl($stderr_w, F_SETFL, 0) or croak($!); # no close-on-exec
     fcntl($stdout_w, F_SETFL, 0) or croak($!);
-    open(STDERR, ">&", $stderr_w);
-    open(STDOUT, ">&", $stdout_w);
+    open(STDERR, ">&", $stderr_w) or croak ($!);
+    open(STDOUT, ">&", $stdout_w) or croak ($!);
     srun ($srunargs, $execargs, $opts, $stdin);
     exit (1);
   }
@@ -2177,8 +2177,22 @@ sub retry_op {
   # that can be retried, the second function will be called with
   # the current try count (0-based), next try time, and error message.
   my $operation = shift;
-  my $retry_callback = shift;
+  my $op_text = shift;
   my $retries = retry_count();
+  my $retry_callback = sub {
+    my ($try_count, $next_try_at, $errmsg) = @_;
+    $errmsg =~ s/\s*\bat \Q$0\E line \d+\.?\s*//;
+    $errmsg =~ s/\s/ /g;
+    $errmsg =~ s/\s+$//;
+    my $retry_msg;
+    if ($next_try_at < time) {
+      $retry_msg = "Retrying.";
+    } else {
+      my $next_try_fmt = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($next_try_at);
+      $retry_msg = "Retrying at $next_try_fmt.";
+    }
+    Log(undef, "$op_text failed: $errmsg. $retry_msg");
+  };
   foreach my $try_count (0..$retries) {
     my $next_try = time + (2 ** $try_count);
     my $result = eval { $operation->(@_); };
@@ -2201,25 +2215,11 @@ sub api_call {
   # This function will call that method, retrying as needed until
   # the current retry_count is exhausted, with a log on the first failure.
   my $method_name = shift;
-  my $log_api_retry = sub {
-    my ($try_count, $next_try_at, $errmsg) = @_;
-    $errmsg =~ s/\s*\bat \Q$0\E line \d+\.?\s*//;
-    $errmsg =~ s/\s/ /g;
-    $errmsg =~ s/\s+$//;
-    my $retry_msg;
-    if ($next_try_at < time) {
-      $retry_msg = "Retrying.";
-    } else {
-      my $next_try_fmt = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($next_try_at);
-      $retry_msg = "Retrying at $next_try_fmt.";
-    }
-    Log(undef, "API method $method_name failed: $errmsg. $retry_msg");
-  };
   my $method = $arv;
   foreach my $key (split(/\//, $method_name)) {
     $method = $method->{$key};
   }
-  return retry_op(sub { $method->execute(@_); }, $log_api_retry, @_);
+  return retry_op(sub { $method->execute(@_); }, "API method $method_name", @_);
 }
 
 sub exit_status_s {
@@ -2493,8 +2493,8 @@ if ((-d $python_dir) and can_run("python2.7")) {
 
 # Hide messages from the install script (unless it fails: shell_or_die
 # will show $destdir.log in that case).
-open(STDOUT, ">>", "$destdir.log");
-open(STDERR, ">&", STDOUT);
+open(STDOUT, ">>", "$destdir.log") or die ($!);
+open(STDERR, ">&", STDOUT) or die ($!);
 
 if (-e "$destdir/crunch_scripts/install") {
     shell_or_die (undef, "$destdir/crunch_scripts/install", $install_dir);
@@ -2515,7 +2515,7 @@ close L;
 
 sub can_run {
   my $command_name = shift;
-  open(my $which, "-|", "which", $command_name);
+  open(my $which, "-|", "which", $command_name) or die ($!);
   while (<$which>) { }
   close($which);
   return ($? == 0);