def get_open_ports(self, node_group):
cluster = node_group.cluster
ports_map = {
'namenode': [8020, 50070, 50470],
'datanode': [50010, 1004, 50075, 1006, 50020],
'master': [
int(c_helper.get_config_value("Spark", "Master port",
cluster)),
int(
c_helper.get_config_value("Spark", "Master webui port",
cluster)),
],
'slave': [
int(
c_helper.get_config_value("Spark", "Worker webui port",
cluster))
]
}
ports = []
for process in node_group.node_processes:
if process in ports_map:
ports.extend(ports_map[process])
return ports
def __init__(self, cluster):
super(EdpEngine, self).__init__(cluster)
self.master = plugin_utils.get_instance(cluster, "master")
self.plugin_params["spark-user"] = ""
self.plugin_params["spark-submit"] = os.path.join(
c_helper.get_config_value("Spark",
"Spark home",
self.cluster),
"bin/spark-submit")
self.plugin_params["deploy-mode"] = "client"
port = c_helper.get_config_value("Spark", "Master port", cluster)
self.plugin_params["master"] = ('spark://%(host)s:%(port)s')
def get_driver_classpath(self):
cp = c_helper.get_config_value("Spark",
"Executor extra classpath",
self.cluster)
if cp:
cp = " --driver-class-path " + cp
return cp
def _validate_existing_ng_scaling(self, cluster, existing):
scalable_processes = self._get_scalable_processes()
dn_to_delete = 0
for ng in cluster.node_groups:
if ng.id in existing:
if ng.count > existing[ng.id] and ("datanode" in ng.node_processes):
dn_to_delete += ng.count - existing[ng.id]
if not set(ng.node_processes).issubset(scalable_processes):
raise ex.NodeGroupCannotBeScaled(
ng.name,
_("Spark plugin cannot scale nodegroup" " with processes: %s") % " ".join(ng.node_processes),
)
dn_amount = len(utils.get_instances(cluster, "datanode"))
rep_factor = c_helper.get_config_value("HDFS", "dfs.replication", cluster)
if dn_to_delete > 0 and dn_amount - dn_to_delete < rep_factor:
raise ex.ClusterCannotBeScaled(
cluster.name,
_(
"Spark plugin cannot shrink cluster because "
"there would be not enough nodes for HDFS "
"replicas (replication factor is %s)"
)
% rep_factor,
)
def validate(self, cluster):
nn_count = sum([ng.count for ng
in utils.get_node_groups(cluster, "namenode")])
if nn_count != 1:
raise ex.InvalidComponentCountException("namenode", 1, nn_count)
dn_count = sum([ng.count for ng
in utils.get_node_groups(cluster, "datanode")])
if dn_count < 1:
raise ex.InvalidComponentCountException("datanode", _("1 or more"),
nn_count)
rep_factor = c_helper.get_config_value('HDFS', "dfs.replication",
cluster)
if dn_count < rep_factor:
raise ex.InvalidComponentCountException(
'datanode', _('%s or more') % rep_factor, dn_count,
_('Number of %(dn)s instances should not be less '
'than %(replication)s')
% {'dn': 'datanode', 'replication': 'dfs.replication'})
# validate Spark Master Node and Spark Slaves
sm_count = sum([ng.count for ng
in utils.get_node_groups(cluster, "master")])
if sm_count != 1:
raise ex.RequiredServiceMissingException("Spark master")
sl_count = sum([ng.count for ng
in utils.get_node_groups(cluster, "slave")])
if sl_count < 1:
raise ex.InvalidComponentCountException("Spark slave",
_("1 or more"),
sl_count)
def _set_cluster_info(self, cluster):
nn = utils.get_instance(cluster, "namenode")
sp_master = utils.get_instance(cluster, "master")
info = {}
if nn:
address = c_helper.get_config_value("HDFS", "dfs.http.address", cluster)
port = address[address.rfind(":") + 1 :]
info["HDFS"] = {"Web UI": "http://%s:%s" % (nn.management_ip, port)}
info["HDFS"]["NameNode"] = "hdfs://%s:8020" % nn.hostname()
if sp_master:
port = c_helper.get_config_value("Spark", "Master webui port", cluster)
if port is not None:
info["Spark"] = {"Web UI": "http://%s:%s" % (sp_master.management_ip, port)}
ctx = context.ctx()
conductor.cluster_update(ctx, cluster, {"info": info})
def _set_cluster_info(self, cluster):
nn = utils.get_instance(cluster, "namenode")
sp_master = utils.get_instance(cluster, "master")
info = {}
if nn:
address = c_helper.get_config_value('HDFS', 'dfs.http.address',
cluster)
port = address[address.rfind(':') + 1:]
info['HDFS'] = {
'Web UI': 'http://%s:%s' % (nn.management_ip, port)
}
info['HDFS']['NameNode'] = 'hdfs://%s:8020' % nn.hostname()
if sp_master:
port = c_helper.get_config_value('Spark', 'Master webui port',
cluster)
if port is not None:
info['Spark'] = {
'Web UI': 'http://%s:%s' % (sp_master.management_ip, port)
}
ctx = context.ctx()
conductor.cluster_update(ctx, cluster, {'info': info})
def get_open_ports(self, node_group):
cluster = node_group.cluster
ports_map = {
'namenode': [8020, 50070, 50470],
'datanode': [50010, 1004, 50075, 1006, 50020],
'master': [
int(c_helper.get_config_value("Spark", "Master port",
cluster)),
int(c_helper.get_config_value("Spark", "Master webui port",
cluster)),
],
'slave': [
int(c_helper.get_config_value("Spark", "Worker webui port",
cluster))
]
}
ports = []
for process in node_group.node_processes:
if process in ports_map:
ports.extend(ports_map[process])
return ports
def _set_cluster_info(self, cluster):
nn = utils.get_instance(cluster, "namenode")
sp_master = utils.get_instance(cluster, "master")
info = {}
if nn:
address = c_helper.get_config_value(
'HDFS', 'dfs.http.address', cluster)
port = address[address.rfind(':') + 1:]
info['HDFS'] = {
'Web UI': 'http://%s:%s' % (nn.management_ip, port)
}
info['HDFS']['NameNode'] = 'hdfs://%s:8020' % nn.hostname()
if sp_master:
port = c_helper.get_config_value(
'Spark', 'Master webui port', cluster)
if port is not None:
info['Spark'] = {
'Web UI': 'http://%s:%s' % (sp_master.management_ip, port)
}
ctx = context.ctx()
conductor.cluster_update(ctx, cluster, {'info': info})
def validate(self, cluster):
nn_count = sum(
[ng.count for ng in utils.get_node_groups(cluster, "namenode")])
if nn_count != 1:
raise ex.InvalidComponentCountException("namenode", 1, nn_count)
dn_count = sum(
[ng.count for ng in utils.get_node_groups(cluster, "datanode")])
if dn_count < 1:
raise ex.InvalidComponentCountException("datanode", _("1 or more"),
nn_count)
rep_factor = c_helper.get_config_value('HDFS', "dfs.replication",
cluster)
if dn_count < rep_factor:
raise ex.InvalidComponentCountException(
'datanode',
_('%s or more') % rep_factor, dn_count,
_('Number of %(dn)s instances should not be less '
'than %(replication)s') % {
'dn': 'datanode',
'replication': 'dfs.replication'
})
# validate Spark Master Node and Spark Slaves
sm_count = sum(
[ng.count for ng in utils.get_node_groups(cluster, "master")])
if sm_count != 1:
raise ex.RequiredServiceMissingException("Spark master")
sl_count = sum(
[ng.count for ng in utils.get_node_groups(cluster, "slave")])
if sl_count < 1:
raise ex.InvalidComponentCountException("Spark slave",
_("1 or more"), sl_count)
def decommission_sl(master, inst_to_be_deleted, survived_inst):
if survived_inst is not None:
slavenames = []
for slave in survived_inst:
slavenames.append(slave.hostname())
slaves_content = c_helper.generate_spark_slaves_configs(slavenames)
else:
slaves_content = "\n"
cluster = master.cluster
sp_home = c_helper.get_config_value("Spark", "Spark home", cluster)
r_master = remote.get_remote(master)
run.stop_spark(r_master, sp_home)
# write new slave file to master
files = {os.path.join(sp_home, 'conf/slaves'): slaves_content}
r_master.write_files_to(files)
# write new slaves file to each survived slave as well
for i in survived_inst:
with remote.get_remote(i) as r:
r.write_files_to(files)
run.start_spark_master(r_master, sp_home)
def _validate_existing_ng_scaling(self, cluster, existing):
scalable_processes = self._get_scalable_processes()
dn_to_delete = 0
for ng in cluster.node_groups:
if ng.id in existing:
if ng.count > existing[ng.id] and ("datanode"
in ng.node_processes):
dn_to_delete += ng.count - existing[ng.id]
if not set(ng.node_processes).issubset(scalable_processes):
raise ex.NodeGroupCannotBeScaled(
ng.name,
_("Spark plugin cannot scale nodegroup"
" with processes: %s") % ' '.join(ng.node_processes))
dn_amount = len(utils.get_instances(cluster, "datanode"))
rep_factor = c_helper.get_config_value('HDFS', "dfs.replication",
cluster)
if dn_to_delete > 0 and dn_amount - dn_to_delete < rep_factor:
raise ex.ClusterCannotBeScaled(
cluster.name,
_("Spark plugin cannot shrink cluster because "
"there would be not enough nodes for HDFS "
"replicas (replication factor is %s)") % rep_factor)
def get_driver_classpath(self):
cp = c_helper.get_config_value("Spark", "Executor extra classpath",
self.cluster)
if cp:
cp = " --driver-class-path " + cp
return cp
def run_job(self, job_execution):
ctx = context.ctx()
job = conductor.job_get(ctx, job_execution.job_id)
# We'll always run the driver program on the master
master = plugin_utils.get_instance(self.cluster, "master")
# TODO(tmckay): wf_dir should probably be configurable.
# The only requirement is that the dir is writable by the image user
wf_dir = job_utils.create_workflow_dir(master, '/tmp/spark-edp', job,
job_execution.id)
paths = job_utils.upload_job_files(master, wf_dir, job,
libs_subdir=False)
# We can shorten the paths in this case since we'll run out of wf_dir
paths = [os.path.basename(p) for p in paths]
# TODO(tmckay): for now, paths[0] is always assumed to be the app
# jar and we generate paths in order (mains, then libs).
# When we have a Spark job type, we can require a "main" and set
# the app jar explicitly to be "main"
app_jar = paths.pop(0)
# The rest of the paths will be passed with --jars
additional_jars = ",".join(paths)
if additional_jars:
additional_jars = "--jars " + additional_jars
# Launch the spark job using spark-submit and deploy_mode = client
host = master.hostname()
port = c_helper.get_config_value("Spark", "Master port", self.cluster)
spark_submit = os.path.join(
c_helper.get_config_value("Spark",
"Spark home",
self.cluster),
"bin/spark-submit")
job_class = job_execution.job_configs.configs["edp.java.main_class"]
# TODO(tmckay): we need to clean up wf_dirs on long running clusters
# TODO(tmckay): probably allow for general options to spark-submit
args = " ".join(job_execution.job_configs.get('args', []))
# The redirects of stdout and stderr will preserve output in the wf_dir
cmd = "%s %s --class %s %s --master spark://%s:%s %s" % (
spark_submit,
app_jar,
job_class,
additional_jars,
host,
port,
args)
# If an exception is raised here, the job_manager will mark
# the job failed and log the exception
with remote.get_remote(master) as r:
# Upload the command launch script
launch = os.path.join(wf_dir, "launch_command")
r.write_file_to(launch, self._job_script())
r.execute_command("chmod +x %s" % launch)
ret, stdout = r.execute_command(
"cd %s; ./launch_command %s > /dev/null 2>&1 & echo $!"
% (wf_dir, cmd))
if ret == 0:
# Success, we'll add the wf_dir in job_execution.extra and store
# pid@instance_id as the job id
# We know the job is running so return "RUNNING"
return (stdout.strip() + "@" + master.id,
edp.JOB_STATUS_RUNNING,
{'spark-path': wf_dir})
# Hmm, no execption but something failed.
# Since we're using backgrounding with redirect, this is unlikely.
raise e.EDPError("Spark job execution failed. Exit status = %s, "
"stdout = %s" % (ret, stdout))
def run_job(self, job_execution):
ctx = context.ctx()
job = conductor.job_get(ctx, job_execution.job_id)
proxy_configs = job_execution.job_configs.get('proxy_configs')
# We'll always run the driver program on the master
master = plugin_utils.get_instance(self.cluster, "master")
# TODO(tmckay): wf_dir should probably be configurable.
# The only requirement is that the dir is writable by the image user
wf_dir = job_utils.create_workflow_dir(master, '/tmp/spark-edp', job,
job_execution.id)
paths = job_utils.upload_job_files(master,
wf_dir,
job,
libs_subdir=False,
proxy_configs=proxy_configs)
# We can shorten the paths in this case since we'll run out of wf_dir
paths = [os.path.basename(p) for p in paths]
# TODO(tmckay): for now, paths[0] is always assumed to be the app
# jar and we generate paths in order (mains, then libs).
# When we have a Spark job type, we can require a "main" and set
# the app jar explicitly to be "main"
app_jar = paths.pop(0)
# The rest of the paths will be passed with --jars
additional_jars = ",".join(paths)
if additional_jars:
additional_jars = "--jars " + additional_jars
# Launch the spark job using spark-submit and deploy_mode = client
host = master.hostname()
port = c_helper.get_config_value("Spark", "Master port", self.cluster)
spark_submit = os.path.join(
c_helper.get_config_value("Spark", "Spark home", self.cluster),
"bin/spark-submit")
job_class = job_execution.job_configs.configs["edp.java.main_class"]
# TODO(tmckay): we need to clean up wf_dirs on long running clusters
# TODO(tmckay): probably allow for general options to spark-submit
args = " ".join(job_execution.job_configs.get('args', []))
# The redirects of stdout and stderr will preserve output in the wf_dir
cmd = "%s %s --class %s %s --master spark://%s:%s %s" % (
spark_submit, app_jar, job_class, additional_jars, host, port,
args)
# If an exception is raised here, the job_manager will mark
# the job failed and log the exception
with remote.get_remote(master) as r:
# Upload the command launch script
launch = os.path.join(wf_dir, "launch_command")
r.write_file_to(launch, self._job_script())
r.execute_command("chmod +x %s" % launch)
ret, stdout = r.execute_command(
"cd %s; ./launch_command %s > /dev/null 2>&1 & echo $!" %
(wf_dir, cmd))
if ret == 0:
# Success, we'll add the wf_dir in job_execution.extra and store
# pid@instance_id as the job id
# We know the job is running so return "RUNNING"
return (stdout.strip() + "@" + master.id, edp.JOB_STATUS_RUNNING, {
'spark-path': wf_dir
})
# Hmm, no execption but something failed.
# Since we're using backgrounding with redirect, this is unlikely.
raise e.EDPError(
_("Spark job execution failed. Exit status = "
"%(status)s, stdout = %(stdout)s") % {
'status': ret,
'stdout': stdout
})
def _spark_home(self, cluster):
return c_helper.get_config_value("Spark", "Spark home", cluster)
def run_job(self, job_execution):
ctx = context.ctx()
job = conductor.job_get(ctx, job_execution.job_id)
additional_sources, updated_job_configs = (
job_utils.resolve_data_source_references(
job_execution.job_configs))
# We'll always run the driver program on the master
master = plugin_utils.get_instance(self.cluster, "master")
# TODO(tmckay): wf_dir should probably be configurable.
# The only requirement is that the dir is writable by the image user
wf_dir = job_utils.create_workflow_dir(master, '/tmp/spark-edp', job,
job_execution.id, "700")
paths, builtin_paths = self._upload_job_files(master, wf_dir, job,
updated_job_configs)
# We can shorten the paths in this case since we'll run out of wf_dir
paths = [os.path.basename(p) for p in paths]
builtin_paths = [os.path.basename(p) for p in builtin_paths]
# TODO(tmckay): for now, paths[0] is always assumed to be the app
# jar and we generate paths in order (mains, then libs).
# When we have a Spark job type, we can require a "main" and set
# the app jar explicitly to be "main"
app_jar = paths.pop(0)
job_class = updated_job_configs["configs"]["edp.java.main_class"]
# If we uploaded builtins then we are using a wrapper jar. It will
# be the first one on the builtin list and the original app_jar needs
# to be added to the 'additional' jars
if builtin_paths:
wrapper_jar = builtin_paths.pop(0)
wrapper_class = 'org.openstack.sahara.edp.SparkWrapper'
wrapper_xml = self._upload_wrapper_xml(master, wf_dir,
updated_job_configs)
wrapper_args = "%s %s" % (wrapper_xml, job_class)
additional_jars = ",".join([app_jar] + paths + builtin_paths)
else:
wrapper_jar = wrapper_class = wrapper_args = ""
additional_jars = ",".join(paths)
# All additional jars are passed with the --jars option
if additional_jars:
additional_jars = " --jars " + additional_jars
# Launch the spark job using spark-submit and deploy_mode = client
host = master.hostname()
port = c_helper.get_config_value("Spark", "Master port", self.cluster)
spark_submit = os.path.join(
c_helper.get_config_value("Spark", "Spark home", self.cluster),
"bin/spark-submit")
# TODO(tmckay): we need to clean up wf_dirs on long running clusters
# TODO(tmckay): probably allow for general options to spark-submit
args = updated_job_configs.get('args', [])
args = " ".join([su.inject_swift_url_suffix(arg) for arg in args])
if args:
args = " " + args
if wrapper_jar and wrapper_class:
# Substrings which may be empty have spaces
# embedded if they are non-empty
cmd = ('%(spark_submit)s%(driver_cp)s'
' --class %(wrapper_class)s%(addnl_jars)s'
' --master spark://%(host)s:%(port)s'
' %(wrapper_jar)s %(wrapper_args)s%(args)s') % (
{
"spark_submit": spark_submit,
"driver_cp": self.get_driver_classpath(),
"wrapper_class": wrapper_class,
"addnl_jars": additional_jars,
"host": host,
"port": port,
"wrapper_jar": wrapper_jar,
"wrapper_args": wrapper_args,
"args": args
})
else:
cmd = ('%(spark_submit)s --class %(job_class)s%(addnl_jars)s'
' --master spark://%(host)s:%(port)s %(app_jar)s%(args)s'
) % ({
"spark_submit": spark_submit,
"job_class": job_class,
"addnl_jars": additional_jars,
"host": host,
"port": port,
"app_jar": app_jar,
"args": args
})
job_execution = conductor.job_execution_get(ctx, job_execution.id)
if job_execution.info['status'] == edp.JOB_STATUS_TOBEKILLED:
return (None, edp.JOB_STATUS_KILLED, None)
# If an exception is raised here, the job_manager will mark
# the job failed and log the exception
# The redirects of stdout and stderr will preserve output in the wf_dir
with remote.get_remote(master) as r:
# Upload the command launch script
launch = os.path.join(wf_dir, "launch_command")
r.write_file_to(launch, self._job_script())
r.execute_command("chmod +x %s" % launch)
ret, stdout = r.execute_command(
"cd %s; ./launch_command %s > /dev/null 2>&1 & echo $!" %
(wf_dir, cmd))
if ret == 0:
# Success, we'll add the wf_dir in job_execution.extra and store
# pid@instance_id as the job id
# We know the job is running so return "RUNNING"
return (stdout.strip() + "@" + master.id, edp.JOB_STATUS_RUNNING, {
'spark-path': wf_dir
})
# Hmm, no execption but something failed.
# Since we're using backgrounding with redirect, this is unlikely.
raise e.EDPError(
_("Spark job execution failed. Exit status = "
"%(status)s, stdout = %(stdout)s") % {
'status': ret,
'stdout': stdout
})
def run_job(self, job_execution):
ctx = context.ctx()
job = conductor.job_get(ctx, job_execution.job_id)
additional_sources, updated_job_configs = (
job_utils.resolve_data_source_references(job_execution.job_configs)
)
for data_source in additional_sources:
if data_source and data_source.type == 'hdfs':
h.configure_cluster_for_hdfs(self.cluster, data_source)
break
# We'll always run the driver program on the master
master = plugin_utils.get_instance(self.cluster, "master")
# TODO(tmckay): wf_dir should probably be configurable.
# The only requirement is that the dir is writable by the image user
wf_dir = job_utils.create_workflow_dir(master, '/tmp/spark-edp', job,
job_execution.id, "700")
paths, builtin_paths = self._upload_job_files(
master, wf_dir, job, updated_job_configs)
# We can shorten the paths in this case since we'll run out of wf_dir
paths = [os.path.basename(p) for p in paths]
builtin_paths = [os.path.basename(p) for p in builtin_paths]
# TODO(tmckay): for now, paths[0] is always assumed to be the app
# jar and we generate paths in order (mains, then libs).
# When we have a Spark job type, we can require a "main" and set
# the app jar explicitly to be "main"
app_jar = paths.pop(0)
job_class = updated_job_configs["configs"]["edp.java.main_class"]
# If we uploaded builtins then we are using a wrapper jar. It will
# be the first one on the builtin list and the original app_jar needs
# to be added to the 'additional' jars
if builtin_paths:
wrapper_jar = builtin_paths.pop(0)
wrapper_class = 'org.openstack.sahara.edp.SparkWrapper'
wrapper_xml = self._upload_wrapper_xml(master,
wf_dir,
updated_job_configs)
wrapper_args = "%s %s" % (wrapper_xml, job_class)
additional_jars = ",".join([app_jar] + paths + builtin_paths)
else:
wrapper_jar = wrapper_class = wrapper_args = ""
additional_jars = ",".join(paths)
# All additional jars are passed with the --jars option
if additional_jars:
additional_jars = " --jars " + additional_jars
# Launch the spark job using spark-submit and deploy_mode = client
host = master.hostname()
port = c_helper.get_config_value("Spark", "Master port", self.cluster)
spark_submit = os.path.join(
c_helper.get_config_value("Spark",
"Spark home",
self.cluster),
"bin/spark-submit")
# TODO(tmckay): we need to clean up wf_dirs on long running clusters
# TODO(tmckay): probably allow for general options to spark-submit
args = updated_job_configs.get('args', [])
args = " ".join([su.inject_swift_url_suffix(arg) for arg in args])
if args:
args = " " + args
if wrapper_jar and wrapper_class:
# Substrings which may be empty have spaces
# embedded if they are non-empty
cmd = (
'%(spark_submit)s%(driver_cp)s'
' --class %(wrapper_class)s%(addnl_jars)s'
' --master spark://%(host)s:%(port)s'
' %(wrapper_jar)s %(wrapper_args)s%(args)s') % (
{
"spark_submit": spark_submit,
"driver_cp": self.get_driver_classpath(),
"wrapper_class": wrapper_class,
"addnl_jars": additional_jars,
"host": host,
"port": port,
"wrapper_jar": wrapper_jar,
"wrapper_args": wrapper_args,
"args": args
})
else:
cmd = (
'%(spark_submit)s --class %(job_class)s%(addnl_jars)s'
' --master spark://%(host)s:%(port)s %(app_jar)s%(args)s') % (
{
"spark_submit": spark_submit,
"job_class": job_class,
"addnl_jars": additional_jars,
"host": host,
"port": port,
"app_jar": app_jar,
"args": args
})
job_execution = conductor.job_execution_get(ctx, job_execution.id)
if job_execution.info['status'] == edp.JOB_STATUS_TOBEKILLED:
return (None, edp.JOB_STATUS_KILLED, None)
# If an exception is raised here, the job_manager will mark
# the job failed and log the exception
# The redirects of stdout and stderr will preserve output in the wf_dir
with remote.get_remote(master) as r:
# Upload the command launch script
launch = os.path.join(wf_dir, "launch_command")
r.write_file_to(launch, self._job_script())
r.execute_command("chmod +x %s" % launch)
ret, stdout = r.execute_command(
"cd %s; ./launch_command %s > /dev/null 2>&1 & echo $!"
% (wf_dir, cmd))
if ret == 0:
# Success, we'll add the wf_dir in job_execution.extra and store
# pid@instance_id as the job id
# We know the job is running so return "RUNNING"
return (stdout.strip() + "@" + master.id,
edp.JOB_STATUS_RUNNING,
{'spark-path': wf_dir})
# Hmm, no execption but something failed.
# Since we're using backgrounding with redirect, this is unlikely.
raise e.EDPError(_("Spark job execution failed. Exit status = "
"%(status)s, stdout = %(stdout)s") %
{'status': ret, 'stdout': stdout})
def _spark_home(self, cluster):
return c_helper.get_config_value("Spark", "Spark home", cluster)
