Obtain job details from Arvados, run tasks on compute nodes (typically
invoked by scheduler on controller):
- crunch-job --job x-y-z
+ crunch-job --job x-y-z --git-dir /path/to/repo/.git
Obtain job details from command line, run tasks on local machine
(typically invoked by application or developer on VM):
- crunch-job --job '{"script_version":"/path/to/tree","script":"scriptname",...}'
+ crunch-job --job '{"script_version":"/path/to/working/tree","script":"scriptname",...}'
+
+ crunch-job --job '{"repository":"https://github.com/curoverse/arvados.git","script_version":"master","script":"scriptname",...}'
=head1 OPTIONS
=item --git-dir
-Path to .git directory where the specified commit is found.
+Path to a .git directory (or a git URL) where the commit given in the
+job's C<script_version> attribute is to be found. If this is I<not>
+given, the job's C<repository> attribute will be used.
=item --job-api-token
setup. This can speed up development and debugging when running jobs
locally.
+=item --job
+
+UUID of the job to run, or a JSON-encoded job resource without a
+UUID. If the latter is given, a new job object will be created.
+
=back
=head1 RUNNING JOBS LOCALLY
use POSIX qw(strftime);
use Fcntl qw(F_GETFL F_SETFL O_NONBLOCK);
use Arvados;
+use Data::Dumper;
use Digest::MD5 qw(md5_hex);
use Getopt::Long;
use IPC::Open2;
my $arv = Arvados->new('apiVersion' => 'v1');
-my $User = $arv->{'users'}->{'current'}->execute;
-
my $Job;
my $job_id;
my $dbh;
my $sth;
+my @jobstep;
+
+my $User = api_call("users/current");
+
if ($jobspec =~ /^[-a-z\d]+$/)
{
# $jobspec is an Arvados UUID, not a JSON job specification
- $Job = $arv->{'jobs'}->{'get'}->execute('uuid' => $jobspec);
+ $Job = api_call("jobs/get", uuid => $jobspec);
if (!$force_unlock) {
# Claim this job, and make sure nobody else does
- eval {
- # lock() sets is_locked_by_uuid and changes state to Running.
- $arv->{'jobs'}->{'lock'}->execute('uuid' => $Job->{'uuid'})
- };
+ eval { api_call("jobs/lock", uuid => $Job->{uuid}); };
if ($@) {
Log(undef, "Error while locking job, exiting ".EX_TEMPFAIL);
exit EX_TEMPFAIL;
$Job->{'started_at'} = gmtime;
$Job->{'state'} = 'Running';
- $Job = $arv->{'jobs'}->{'create'}->execute('job' => $Job);
+ $Job = api_call("jobs/create", job => $Job);
}
$job_id = $Job->{'uuid'};
$ENV{"JOB_UUID"} = $job_id;
-my @jobstep;
my @jobstep_todo = ();
my @jobstep_done = ();
my @jobstep_tomerge = ();
}
else
{
- my $first_task = $arv->{'job_tasks'}->{'create'}->execute('job_task' => {
+ my $first_task = api_call("job_tasks/create", job_task => {
'job_uuid' => $Job->{'uuid'},
'sequence' => 0,
'qsequence' => 0,
'parameters' => {},
- });
+ });
push @jobstep, { 'level' => 0,
'failures' => 0,
'arvados_task' => $first_task,
if ($cleanpid == 0)
{
srun (["srun", "--nodelist=$nodelist", "-D", $ENV{'TMPDIR'}],
- ['bash', '-c', 'if mount | grep -q $JOB_WORK/; then for i in $JOB_WORK/*keep; do /bin/fusermount -z -u $i; done; fi; sleep 1; rm -rf $JOB_WORK $CRUNCH_TMP/opt $CRUNCH_TMP/src*']);
+ ['bash', '-c', 'if mount | grep -q $JOB_WORK/; then for i in $JOB_WORK/*keep $CRUNCH_TMP/task/*.keep; do /bin/fusermount -z -u $i; done; fi; sleep 1; rm -rf $JOB_WORK $CRUNCH_INSTALL $CRUNCH_TMP/task $CRUNCH_TMP/src*']);
exit (1);
}
while (1)
my $git_archive;
if (!defined $git_dir && $Job->{'script_version'} =~ m{^/}) {
- # If we're in user-land (i.e., not called from crunch-dispatch)
- # script_version can be an absolute directory path, signifying we
- # should work straight out of that directory instead of using a git
- # commit.
+ # If script_version looks like an absolute path, *and* the --git-dir
+ # argument was not given -- which implies we were not invoked by
+ # crunch-dispatch -- we will use the given path as a working
+ # directory instead of resolving script_version to a git commit (or
+ # doing anything else with git).
$ENV{"CRUNCH_SRC_COMMIT"} = $Job->{'script_version'};
$ENV{"CRUNCH_SRC"} = $Job->{'script_version'};
}
else {
+ # Resolve the given script_version to a git commit sha1. Also, if
+ # the repository is remote, clone it into our local filesystem: this
+ # ensures "git archive" will work, and is necessary to reliably
+ # resolve a symbolic script_version like "master^".
$ENV{"CRUNCH_SRC"} = "$ENV{CRUNCH_TMP}/src";
- # Install requested code version
Log (undef, "Looking for version ".$Job->{script_version}." from repository ".$Job->{repository});
$ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
} else {
# $repo is none of the above. It must be the name of a hosted
# repository.
- my $arv_repo_list = $arv->{'repositories'}->{'list'}->execute(
- 'filters' => [['name','=',$repo]]
- )->{'items'};
- my $n_found = scalar @{$arv_repo_list};
+ my $arv_repo_list = api_call("repositories/list",
+ 'filters' => [['name','=',$repo]]);
+ my @repos_found = @{$arv_repo_list->{'items'}};
+ my $n_found = $arv_repo_list->{'serverResponse'}->{'items_available'};
if ($n_found > 0) {
Log(undef, "Repository '$repo' -> "
- . join(", ", map { $_->{'uuid'} } @{$arv_repo_list}));
+ . join(", ", map { $_->{'uuid'} } @repos_found));
}
if ($n_found != 1) {
croak("Error: Found $n_found repositories with name '$repo'.");
}
- $repo = $arv_repo_list->[0]->{'fetch_url'};
+ $repo = $repos_found[0]->{'fetch_url'};
$repo_location = 'remote';
}
Log(undef, "Using $repo_location repository '$repo'");
# our local cache first, since that's cheaper. (We don't want to
# do that with tags/branches though -- those change over time, so
# they should always be resolved by the remote repo.)
- if ($treeish =~ /^[0-9a-f]{3,40}$/s) {
+ if ($treeish =~ /^[0-9a-f]{7,40}$/s) {
# Hide stderr because it's normal for this to fail:
my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E 2>/dev/null`;
if ($? == 0 &&
- $sha1 =~ /^$treeish/ && # Don't use commit 123 @ branch abc!
+ # Careful not to resolve a branch named abcdeff to commit 1234567:
+ $sha1 =~ /^$treeish/ &&
$sha1 =~ /^([0-9a-f]{40})$/s) {
$commit = $1;
Log(undef, "Commit $commit already present in $local_repo");
my @execargs = ("sh", "-c",
"mkdir -p $ENV{CRUNCH_INSTALL} && cd $ENV{CRUNCH_TMP} && perl -");
- # Note: this section is almost certainly unnecessary if we're
- # running tasks in docker containers.
my $installpid = fork();
if ($installpid == 0)
{
freeze_if_want_freeze ($installpid);
select (undef, undef, undef, 0.1);
}
- Log (undef, "Install script exited ".exit_status_s($?));
+ my $install_exited = $?;
+ Log (undef, "Install script exited ".exit_status_s($install_exited));
+ exit (1) if $install_exited != 0;
}
if (!$have_slurm)
}
$ENV{"TASK_SLOT_NODE"} = $slot[$childslot]->{node}->{name};
$ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
- $ENV{"TASK_WORK"} = $ENV{"JOB_WORK"}."/$id.$$";
+ $ENV{"TASK_WORK"} = $ENV{"CRUNCH_TMP"}."/task/$childslotname";
$ENV{"HOME"} = $ENV{"TASK_WORK"};
$ENV{"TASK_KEEPMOUNT"} = $ENV{"TASK_WORK"}.".keep";
$ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
qw(-n1 -c1 -N1 -D), $ENV{'TMPDIR'},
"--job-name=$job_id.$id.$$",
);
- my $build_script_to_send = "";
my $command =
"if [ -e $ENV{TASK_WORK} ]; then rm -rf $ENV{TASK_WORK}; fi; "
."mkdir -p $ENV{CRUNCH_TMP} $ENV{JOB_WORK} $ENV{TASK_WORK} $ENV{TASK_KEEPMOUNT} "
."&& cd $ENV{CRUNCH_TMP} ";
- if ($build_script)
- {
- $build_script_to_send = $build_script;
- $command .=
- "&& perl -";
- }
$command .= "&& exec arv-mount --by-id --allow-other $ENV{TASK_KEEPMOUNT} --exec ";
if ($docker_hash)
{
- $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -cgroup-parent=docker -cgroup-cid=$ENV{TASK_WORK}/docker.cid -poll=10000 ";
- $command .= "$docker_bin run --rm=true --attach=stdout --attach=stderr --user=crunch --cidfile=$ENV{TASK_WORK}/docker.cid ";
+ my $cidfile = "$ENV{CRUNCH_TMP}/$ENV{TASK_UUID}.cid";
+ $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -cgroup-parent=docker -cgroup-cid=$cidfile -poll=10000 ";
+ $command .= "$docker_bin run --rm=true --attach=stdout --attach=stderr --attach=stdin -i --user=crunch --cidfile=$cidfile --sig-proxy ";
+
# Dynamically configure the container to use the host system as its
# DNS server. Get the host's global addresses from the ip command,
# and turn them into docker --dns options using gawk.
$command .=
q{$(ip -o address show scope global |
gawk 'match($4, /^([0-9\.:]+)\//, x){print "--dns", x[1]}') };
- $command .= "--volume=\Q$ENV{CRUNCH_SRC}:/tmp/crunch-src:ro\E ";
+
+ # The source tree and $destdir directory (which we have
+ # installed on the worker host) are available in the container,
+ # under the same path.
+ $command .= "--volume=\Q$ENV{CRUNCH_SRC}:$ENV{CRUNCH_SRC}:ro\E ";
+ $command .= "--volume=\Q$ENV{CRUNCH_INSTALL}:$ENV{CRUNCH_INSTALL}:ro\E ";
+
+ # Currently, we make arv-mount's mount point appear at /keep
+ # inside the container (instead of using the same path as the
+ # host like we do with CRUNCH_SRC and CRUNCH_INSTALL). However,
+ # crunch scripts and utilities must not rely on this. They must
+ # use $TASK_KEEPMOUNT.
$command .= "--volume=\Q$ENV{TASK_KEEPMOUNT}:/keep:ro\E ";
- $command .= "--env=\QHOME=/home/crunch\E ";
+ $ENV{TASK_KEEPMOUNT} = "/keep";
+
+ # TASK_WORK is almost exactly like a docker data volume: it
+ # starts out empty, is writable, and persists until no
+ # containers use it any more. We don't use --volumes-from to
+ # share it with other containers: it is only accessible to this
+ # task, and it goes away when this task stops.
+ #
+ # However, a docker data volume is writable only by root unless
+ # the mount point already happens to exist in the container with
+ # different permissions. Therefore, we [1] assume /tmp already
+ # exists in the image and is writable by the crunch user; [2]
+ # avoid putting TASK_WORK inside CRUNCH_TMP (which won't be
+ # writable if they are created by docker while setting up the
+ # other --volumes); and [3] create $TASK_WORK inside the
+ # container using $build_script.
+ $command .= "--volume=/tmp ";
+ $ENV{"TASK_WORK"} = "/tmp/crunch-job-task-work/$childslotname";
+ $ENV{"HOME"} = $ENV{"TASK_WORK"};
+ $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
+
+ # TODO: Share a single JOB_WORK volume across all task
+ # containers on a given worker node, and delete it when the job
+ # ends (and, in case that doesn't work, when the next job
+ # starts).
+ #
+ # For now, use the same approach as TASK_WORK above.
+ $ENV{"JOB_WORK"} = "/tmp/crunch-job-work";
+
while (my ($env_key, $env_val) = each %ENV)
{
- if ($env_key =~ /^(ARVADOS|JOB|TASK)_/) {
- if ($env_key eq "TASK_WORK") {
- $command .= "--env=\QTASK_WORK=/tmp/crunch-job\E ";
- }
- elsif ($env_key eq "TASK_KEEPMOUNT") {
- $command .= "--env=\QTASK_KEEPMOUNT=/keep\E ";
- }
- else {
- $command .= "--env=\Q$env_key=$env_val\E ";
- }
+ if ($env_key =~ /^(ARVADOS|CRUNCH|JOB|TASK)_/) {
+ $command .= "--env=\Q$env_key=$env_val\E ";
}
}
- $command .= "--env=\QCRUNCH_NODE_SLOTS=$ENV{CRUNCH_NODE_SLOTS}\E ";
- $command .= "--env=\QCRUNCH_SRC=/tmp/crunch-src\E ";
+ $command .= "--env=\QHOME=$ENV{HOME}\E ";
$command .= "\Q$docker_hash\E ";
$command .= "stdbuf --output=0 --error=0 ";
- $command .= "/tmp/crunch-src/crunch_scripts/" . $Job->{"script"};
+ $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
} else {
# Non-docker run
$command .= "crunchstat -cgroup-root=/sys/fs/cgroup -poll=10000 ";
$command .= "stdbuf --output=0 --error=0 ";
- $command .= "$ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
+ $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
}
my @execargs = ('bash', '-c', $command);
- srun (\@srunargs, \@execargs, undef, $build_script_to_send);
+ srun (\@srunargs, \@execargs, undef, $build_script);
# exec() failed, we assume nothing happened.
- Log(undef, "srun() failed on build script");
- die;
+ die "srun() failed on build script\n";
}
close("writer");
if (!defined $childpid)
while (my $manifest_line = <$orig_manifest>) {
$orig_manifest_text .= $manifest_line;
}
- my $output = $arv->{'collections'}->{'create'}->execute('collection' => {
- 'manifest_text' => $orig_manifest_text,
- });
+ my $output = api_call("collections/create", collection => {
+ 'manifest_text' => $orig_manifest_text});
Log(undef, "output uuid " . $output->{uuid});
Log(undef, "output hash " . $output->{portable_data_hash});
$Job->update_attributes('output' => $output->{portable_data_hash});
my $newtask_list = [];
my $newtask_results;
do {
- $newtask_results = $arv->{'job_tasks'}->{'list'}->execute(
+ $newtask_results = api_call(
+ "job_tasks/list",
'where' => {
'created_by_job_task_uuid' => $Jobstep->{'arvados_task'}->{uuid}
},
my @stat = stat $ENV{"CRUNCH_REFRESH_TRIGGER"};
if (@stat && $stat[9] > $latest_refresh) {
$latest_refresh = scalar time;
- my $Job2 = $arv->{'jobs'}->{'get'}->execute('uuid' => $jobspec);
+ my $Job2 = api_call("jobs/get", uuid => $jobspec);
for my $attr ('cancelled_at',
'cancelled_by_user_uuid',
'cancelled_by_client_uuid',
my ($child_out, $child_in);
my $pid = open2($child_out, $child_in, 'arv-put', '--raw',
- '--retries', put_retry_count());
+ '--retries', retry_count());
my $joboutput;
for (@jobstep)
{
Log (undef, $message);
freeze() if @jobstep_todo;
collate_output() if @jobstep_todo;
- cleanup() if $Job;
- save_meta() if log_writer_is_active();
+ cleanup();
+ save_meta();
die;
}
sub cleanup
{
+ return unless $Job;
if ($Job->{'state'} eq 'Cancelled') {
$Job->update_attributes('finished_at' => scalar gmtime);
} else {
{
my $justcheckpoint = shift; # false if this will be the last meta saved
return if $justcheckpoint; # checkpointing is not relevant post-Warehouse.pm
+ return unless log_writer_is_active();
my $loglocator = log_writer_finish();
Log (undef, "log manifest is $loglocator");
my $opts = shift || {};
my $stdin = shift;
my $args = $have_slurm ? [@$srunargs, @$execargs] : $execargs;
- print STDERR (join (" ",
- map { / / ? "'$_'" : $_ }
- (@$args)),
- "\n")
- if $ENV{CRUNCH_DEBUG};
+
+ $Data::Dumper::Terse = 1;
+ $Data::Dumper::Indent = 0;
+ my $show_cmd = Dumper($args);
+ $show_cmd =~ s/(TOKEN\\*=)[^\s\']+/${1}[...]/g;
+ $show_cmd =~ s/\n/ /g;
+ warn "starting: $show_cmd\n";
if (defined $stdin) {
my $child = open STDIN, "-|";
# If not, return undef for both values.
my $locator = shift;
my ($streamname, $filename);
- if (my $image = $arv->{collections}->{get}->execute(uuid => $locator)) {
+ my $image = api_call("collections/get", uuid => $locator);
+ if ($image) {
foreach my $line (split(/\n/, $image->{manifest_text})) {
my @tokens = split(/\s+/, $line);
next if (!@tokens);
}
}
-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++;
- $timediff /= 2;
+sub retry_count {
+ # Calculate the number of times an operation should be retried,
+ # assuming exponential backoff, and that we're willing to retry as
+ # long as tasks have been running. Enforce a minimum of 3 retries.
+ my ($starttime, $endtime, $timediff, $retries);
+ if (@jobstep) {
+ $starttime = $jobstep[0]->{starttime};
+ $endtime = $jobstep[-1]->{finishtime};
+ }
+ if (!defined($starttime)) {
+ $timediff = 0;
+ } elsif (!defined($endtime)) {
+ $timediff = time - $starttime;
+ } else {
+ $timediff = ($endtime - $starttime) - (time - $endtime);
+ }
+ if ($timediff > 0) {
+ $retries = int(log($timediff) / log(2));
+ } else {
+ $retries = 1; # Use the minimum.
}
return ($retries > 3) ? $retries : 3;
}
+sub retry_op {
+ # Pass in two function references.
+ # This method will be called with the remaining arguments.
+ # If it dies, retry it with exponential backoff until it succeeds,
+ # or until the current retry_count is exhausted. After each failure
+ # 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 $retries = retry_count();
+ foreach my $try_count (0..$retries) {
+ my $next_try = time + (2 ** $try_count);
+ my $result = eval { $operation->(@_); };
+ if (!$@) {
+ return $result;
+ } elsif ($try_count < $retries) {
+ $retry_callback->($try_count, $next_try, $@);
+ my $sleep_time = $next_try - time;
+ sleep($sleep_time) if ($sleep_time > 0);
+ }
+ }
+ # Ensure the error message ends in a newline, so Perl doesn't add
+ # retry_op's line number to it.
+ chomp($@);
+ die($@ . "\n");
+}
+
+sub api_call {
+ # Pass in a /-separated API method name, and arguments for it.
+ # 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-%d %H:%M:%S", $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, @_);
+}
+
sub exit_status_s {
# Given a $?, return a human-readable exit code string like "0" or
# "1" or "0 with signal 1" or "1 with signal 11".
# checkout-and-build
use Fcntl ':flock';
-use File::Path qw( make_path );
+use File::Path qw( make_path remove_tree );
my $destdir = $ENV{"CRUNCH_SRC"};
my $commit = $ENV{"CRUNCH_SRC_COMMIT"};
my $repo = $ENV{"CRUNCH_SRC_URL"};
+my $job_work = $ENV{"JOB_WORK"};
my $task_work = $ENV{"TASK_WORK"};
-for my $dir ($destdir, $task_work) {
- if ($dir) {
- make_path $dir;
- -e $dir or die "Failed to create temporary directory ($dir): $!";
- }
+for my $dir ($destdir, $job_work, $task_work) {
+ if ($dir) {
+ make_path $dir;
+ -e $dir or die "Failed to create temporary directory ($dir): $!";
+ }
}
+if ($task_work) {
+ remove_tree($task_work, {keep_root => 1});
+}
+
+
open L, ">", "$destdir.lock" or die "$destdir.lock: $!";
flock L, LOCK_EX;
if (readlink ("$destdir.commit") eq $commit && -d $destdir) {
- if (@ARGV) {
- exec(@ARGV);
- die "Cannot exec `@ARGV`: $!";
- } else {
- exit 0;
- }
+ # This version already installed -> nothing to do.
+ run_argv_and_exit();
}
unlink "$destdir.commit";
+open STDERR_ORIG, ">&STDERR";
open STDOUT, ">", "$destdir.log";
open STDERR, ">&STDOUT";
mkdir $destdir;
my @git_archive_data = <DATA>;
-if (@git_archive_data) {
- open TARX, "|-", "tar", "-C", $destdir, "-xf", "-";
- print TARX @git_archive_data;
- if(!close(TARX)) {
- die "'tar -C $destdir -xf -' exited $?: $!";
- }
+if (!@git_archive_data) {
+ # Nothing to extract -> nothing to install.
+ run_argv_and_exit();
+}
+
+open TARX, "|-", "tar", "-C", $destdir, "-xf", "-";
+print TARX @git_archive_data;
+if(!close(TARX)) {
+ die "'tar -C $destdir -xf -' exited $?: $!";
}
my $pwd;
close L;
-if (@ARGV) {
+run_argv_and_exit();
+
+sub run_argv_and_exit
+{
+ if (@ARGV) {
exec(@ARGV);
die "Cannot exec `@ARGV`: $!";
-} else {
+ } else {
exit 0;
+ }
}
sub shell_or_die
if ($ENV{"DEBUG"}) {
print STDERR "@_\n";
}
- system (@_) == 0
- or die "@_ failed: $! exit 0x".sprintf("%x",$?);
+ if (system (@_) != 0) {
+ my $err = $!;
+ my $exitstatus = sprintf("exit %d signal %d", $? >> 8, $? & 0x7f);
+ open STDERR, ">&STDERR_ORIG";
+ system ("cat $destdir.log >&2");
+ die "@_ failed ($err): $exitstatus";
+ }
}
__DATA__
projects / arvados.git / blobdiff