diff --git a/docs/source/conf.py b/docs/source/conf.py index b85ec81..d1b7d07 100644 --- a/docs/source/conf.py +++ b/docs/source/conf.py @@ -56,8 +56,8 @@ # source_suffix = ['.rst', '.md'] source_suffix = '.rst' -# The master toctree document. -master_doc = 'index' +# The main toctree document. +main_doc = 'index' # The language for content autogenerated by Sphinx. Refer to documentation # for a list of supported languages. @@ -138,7 +138,7 @@ # (source start file, target name, title, # author, documentclass [howto, manual, or own class]). latex_documents = [ - (master_doc, 'TensorFlowOnSpark.tex', 'TensorFlowOnSpark Documentation', + (main_doc, 'TensorFlowOnSpark.tex', 'TensorFlowOnSpark Documentation', 'Lee Yang', 'manual'), ] @@ -148,7 +148,7 @@ # One entry per manual page. List of tuples # (source start file, name, description, authors, manual section). man_pages = [ - (master_doc, 'tensorflowonspark', 'TensorFlowOnSpark Documentation', + (main_doc, 'tensorflowonspark', 'TensorFlowOnSpark Documentation', [author], 1) ] @@ -159,7 +159,7 @@ # (source start file, target name, title, author, # dir menu entry, description, category) texinfo_documents = [ - (master_doc, 'TensorFlowOnSpark', 'TensorFlowOnSpark Documentation', + (main_doc, 'TensorFlowOnSpark', 'TensorFlowOnSpark Documentation', author, 'TensorFlowOnSpark', 'One line description of project.', 'Miscellaneous'), ] diff --git a/examples/mnist/estimator/mnist_estimator.py b/examples/mnist/estimator/mnist_estimator.py index dddd4b9..634f71c 100644 --- a/examples/mnist/estimator/mnist_estimator.py +++ b/examples/mnist/estimator/mnist_estimator.py @@ -184,5 +184,5 @@ def main(args, ctx): args = parser.parse_args() print("args:", args) - cluster = TFCluster.run(sc, main, args, args.cluster_size, args.num_ps, tensorboard=args.tensorboard, input_mode=TFCluster.InputMode.TENSORFLOW, log_dir=args.model, master_node='master') + cluster = TFCluster.run(sc, main, args, args.cluster_size, args.num_ps, tensorboard=args.tensorboard, input_mode=TFCluster.InputMode.TENSORFLOW, log_dir=args.model, main_node='main') cluster.shutdown() diff --git a/examples/mnist/keras/mnist_mlp_estimator.py b/examples/mnist/keras/mnist_mlp_estimator.py index 972c6a7..669f11a 100644 --- a/examples/mnist/keras/mnist_mlp_estimator.py +++ b/examples/mnist/keras/mnist_mlp_estimator.py @@ -150,13 +150,13 @@ def train_input_fn(): if args.input_mode == 'tf': # for TENSORFLOW mode, each node will load/train/infer entire dataset in memory per original example - cluster = TFCluster.run(sc, main_fun, args, args.cluster_size, args.num_ps, args.tensorboard, TFCluster.InputMode.TENSORFLOW, log_dir=args.model_dir, master_node='master') + cluster = TFCluster.run(sc, main_fun, args, args.cluster_size, args.num_ps, args.tensorboard, TFCluster.InputMode.TENSORFLOW, log_dir=args.model_dir, main_node='main') cluster.shutdown() else: # 'spark' # for SPARK mode, just use CSV format as an example images = sc.textFile(args.images).map(lambda ln: [float(x) for x in ln.split(',')]) labels = sc.textFile(args.labels).map(lambda ln: [float(x) for x in ln.split(',')]) dataRDD = images.zip(labels) - cluster = TFCluster.run(sc, main_fun, args, args.cluster_size, args.num_ps, args.tensorboard, TFCluster.InputMode.SPARK, log_dir=args.model_dir, master_node='master') + cluster = TFCluster.run(sc, main_fun, args, args.cluster_size, args.num_ps, args.tensorboard, TFCluster.InputMode.SPARK, log_dir=args.model_dir, main_node='main') cluster.train(dataRDD, args.epochs) cluster.shutdown() diff --git a/examples/mnist/spark/mnist_dist.py b/examples/mnist/spark/mnist_dist.py index cdf0a8c..aa76039 100644 --- a/examples/mnist/spark/mnist_dist.py +++ b/examples/mnist/spark/mnist_dist.py @@ -101,7 +101,7 @@ def rdd_generator(): summary_writer = tf.summary.FileWriter("tensorboard_%d" % worker_num, graph=tf.get_default_graph()) hooks = [tf.train.StopAtStepHook(last_step=args.steps)] if args.mode == "train" else [] - with tf.train.MonitoredTrainingSession(master=server.target, + with tf.train.MonitoredTrainingSession(main=server.target, is_chief=(task_index == 0), scaffold=tf.train.Scaffold(init_op=init_op, summary_op=summary_op, saver=saver), checkpoint_dir=logdir, diff --git a/examples/mnist/tf/mnist_dist.py b/examples/mnist/tf/mnist_dist.py index cc95e93..6e87cfa 100644 --- a/examples/mnist/tf/mnist_dist.py +++ b/examples/mnist/tf/mnist_dist.py @@ -136,7 +136,7 @@ def build_model(graph, x): tf.gfile.MkDir(output_dir) output_file = tf.gfile.Open("{}/part-{:05d}".format(output_dir, task_index), mode='w') - with tf.train.MonitoredTrainingSession(master=server.target, + with tf.train.MonitoredTrainingSession(main=server.target, is_chief=(task_index == 0), scaffold=tf.train.Scaffold(init_op=init_op, summary_op=summary_op, saver=saver), checkpoint_dir=model_dir, diff --git a/examples/wide_deep/census_main.py b/examples/wide_deep/census_main.py index dfbde48..2e793fb 100644 --- a/examples/wide_deep/census_main.py +++ b/examples/wide_deep/census_main.py @@ -155,5 +155,5 @@ def main_fun(argv, ctx): num_workers = args.cluster_size - args.num_ps print("===== num_executors={}, num_workers={}, num_ps={}".format(args.cluster_size, num_workers, args.num_ps)) - cluster = TFCluster.run(sc, main_fun, remainder, args.cluster_size, args.num_ps, False, TFCluster.InputMode.TENSORFLOW, master_node='master') + cluster = TFCluster.run(sc, main_fun, remainder, args.cluster_size, args.num_ps, False, TFCluster.InputMode.TENSORFLOW, main_node='main') cluster.shutdown() diff --git a/scripts/spark_ec2.py b/scripts/spark_ec2.py index 2da4b0d..c818fd0 100644 --- a/scripts/spark_ec2.py +++ b/scripts/spark_ec2.py @@ -177,11 +177,11 @@ def parse_args(): prog="spark-ec2", version="%prog {v}".format(v=SPARK_EC2_VERSION), usage="%prog [options] \n\n" - + " can be: launch, destroy, login, stop, start, get-master, reboot-slaves") + + " can be: launch, destroy, login, stop, start, get-main, reboot-subordinates") parser.add_option( - "-s", "--slaves", type="int", default=1, - help="Number of slaves to launch (default: %default)") + "-s", "--subordinates", type="int", default=1, + help="Number of subordinates to launch (default: %default)") parser.add_option( "-w", "--wait", type="int", help="DEPRECATED (no longer necessary) - Seconds to wait for nodes to start") @@ -200,15 +200,15 @@ def parse_args(): help="Type of instance to launch (default: %default). " + "WARNING: must be 64-bit; small instances won't work") parser.add_option( - "-m", "--master-instance-type", default="", - help="Master instance type (leave empty for same as instance-type)") + "-m", "--main-instance-type", default="", + help="Main instance type (leave empty for same as instance-type)") parser.add_option( "-r", "--region", default="us-east-1", help="EC2 region used to launch instances in, or to find them in (default: %default)") parser.add_option( "-z", "--zone", default="", help="Availability zone to launch instances in, or 'all' to spread " + - "slaves across multiple (an additional $0.01/Gb for bandwidth" + + "subordinates across multiple (an additional $0.01/Gb for bandwidth" + "between zones applies) (default: a single zone chosen at random)") parser.add_option( "-a", "--ami", @@ -231,7 +231,7 @@ def parse_args(): parser.add_option( "--deploy-root-dir", default=None, - help="A directory to copy into / on the first master. " + + help="A directory to copy into / on the first main. " + "Must be absolute. Note that a trailing slash is handled as per rsync: " + "If you omit it, the last directory of the --deploy-root-dir path will be created " + "in / before copying its contents. If you append the trailing slash, " + @@ -272,7 +272,7 @@ def parse_args(): help="Swap space to set up per node, in MB (default: %default)") parser.add_option( "--spot-price", metavar="PRICE", type="float", - help="If specified, launch slaves as spot instances with the given " + + help="If specified, launch subordinates as spot instances with the given " + "maximum price (in dollars)") parser.add_option( "--ganglia", action="store_true", default=True, @@ -288,15 +288,15 @@ def parse_args(): "--delete-groups", action="store_true", default=False, help="When destroying a cluster, delete the security groups that were created") parser.add_option( - "--use-existing-master", action="store_true", default=False, - help="Launch fresh slaves, but use an existing stopped master if possible") + "--use-existing-main", action="store_true", default=False, + help="Launch fresh subordinates, but use an existing stopped main if possible") parser.add_option( "--worker-instances", type="int", default=1, help="Number of instances per worker: variable SPARK_WORKER_INSTANCES. Not used if YARN " + "is used as Hadoop major version (default: %default)") parser.add_option( - "--master-opts", type="string", default="", - help="Extra options to give to master through SPARK_MASTER_OPTS variable " + + "--main-opts", type="string", default="", + help="Extra options to give to main through SPARK_MASTER_OPTS variable " + "(e.g -Dspark.worker.timeout=180)") parser.add_option( "--user-data", type="string", default="", @@ -483,7 +483,7 @@ def get_spark_ami(opts): # Launch a cluster of the given name, by setting up its security groups, # and then starting new instances in them. -# Returns a tuple of EC2 reservation objects for the master and slaves +# Returns a tuple of EC2 reservation objects for the main and subordinates # Fails if there already instances running in the cluster's groups. def launch_cluster(conn, opts, cluster_name): if opts.identity_file is None: @@ -500,75 +500,75 @@ def launch_cluster(conn, opts, cluster_name): user_data_content = user_data_file.read() print("Setting up security groups...") - master_group = get_or_make_group(conn, cluster_name + "-master", opts.vpc_id) - slave_group = get_or_make_group(conn, cluster_name + "-slaves", opts.vpc_id) + main_group = get_or_make_group(conn, cluster_name + "-main", opts.vpc_id) + subordinate_group = get_or_make_group(conn, cluster_name + "-subordinates", opts.vpc_id) authorized_address = opts.authorized_address - if master_group.rules == []: # Group was just now created + if main_group.rules == []: # Group was just now created if opts.vpc_id is None: - master_group.authorize(src_group=master_group) - master_group.authorize(src_group=slave_group) + main_group.authorize(src_group=main_group) + main_group.authorize(src_group=subordinate_group) else: - master_group.authorize(ip_protocol='icmp', from_port=-1, to_port=-1, - src_group=master_group) - master_group.authorize(ip_protocol='tcp', from_port=0, to_port=65535, - src_group=master_group) - master_group.authorize(ip_protocol='udp', from_port=0, to_port=65535, - src_group=master_group) - master_group.authorize(ip_protocol='icmp', from_port=-1, to_port=-1, - src_group=slave_group) - master_group.authorize(ip_protocol='tcp', from_port=0, to_port=65535, - src_group=slave_group) - master_group.authorize(ip_protocol='udp', from_port=0, to_port=65535, - src_group=slave_group) - master_group.authorize('tcp', 22, 22, authorized_address) - master_group.authorize('tcp', 8080, 8081, authorized_address) - master_group.authorize('tcp', 18080, 18080, authorized_address) - master_group.authorize('tcp', 19999, 19999, authorized_address) - master_group.authorize('tcp', 50030, 50030, authorized_address) - master_group.authorize('tcp', 50070, 50070, authorized_address) - master_group.authorize('tcp', 60070, 60070, authorized_address) - master_group.authorize('tcp', 4040, 4045, authorized_address) + main_group.authorize(ip_protocol='icmp', from_port=-1, to_port=-1, + src_group=main_group) + main_group.authorize(ip_protocol='tcp', from_port=0, to_port=65535, + src_group=main_group) + main_group.authorize(ip_protocol='udp', from_port=0, to_port=65535, + src_group=main_group) + main_group.authorize(ip_protocol='icmp', from_port=-1, to_port=-1, + src_group=subordinate_group) + main_group.authorize(ip_protocol='tcp', from_port=0, to_port=65535, + src_group=subordinate_group) + main_group.authorize(ip_protocol='udp', from_port=0, to_port=65535, + src_group=subordinate_group) + main_group.authorize('tcp', 22, 22, authorized_address) + main_group.authorize('tcp', 8080, 8081, authorized_address) + main_group.authorize('tcp', 18080, 18080, authorized_address) + main_group.authorize('tcp', 19999, 19999, authorized_address) + main_group.authorize('tcp', 50030, 50030, authorized_address) + main_group.authorize('tcp', 50070, 50070, authorized_address) + main_group.authorize('tcp', 60070, 60070, authorized_address) + main_group.authorize('tcp', 4040, 4045, authorized_address) # HDFS NFS gateway requires 111,2049,4242 for tcp & udp - master_group.authorize('tcp', 111, 111, authorized_address) - master_group.authorize('udp', 111, 111, authorized_address) - master_group.authorize('tcp', 2049, 2049, authorized_address) - master_group.authorize('udp', 2049, 2049, authorized_address) - master_group.authorize('tcp', 4242, 4242, authorized_address) - master_group.authorize('udp', 4242, 4242, authorized_address) + main_group.authorize('tcp', 111, 111, authorized_address) + main_group.authorize('udp', 111, 111, authorized_address) + main_group.authorize('tcp', 2049, 2049, authorized_address) + main_group.authorize('udp', 2049, 2049, authorized_address) + main_group.authorize('tcp', 4242, 4242, authorized_address) + main_group.authorize('udp', 4242, 4242, authorized_address) # RM in YARN mode uses 8088 - master_group.authorize('tcp', 8088, 8088, authorized_address) + main_group.authorize('tcp', 8088, 8088, authorized_address) if opts.ganglia: - master_group.authorize('tcp', 5080, 5080, authorized_address) - if slave_group.rules == []: # Group was just now created + main_group.authorize('tcp', 5080, 5080, authorized_address) + if subordinate_group.rules == []: # Group was just now created if opts.vpc_id is None: - slave_group.authorize(src_group=master_group) - slave_group.authorize(src_group=slave_group) + subordinate_group.authorize(src_group=main_group) + subordinate_group.authorize(src_group=subordinate_group) else: - slave_group.authorize(ip_protocol='icmp', from_port=-1, to_port=-1, - src_group=master_group) - slave_group.authorize(ip_protocol='tcp', from_port=0, to_port=65535, - src_group=master_group) - slave_group.authorize(ip_protocol='udp', from_port=0, to_port=65535, - src_group=master_group) - slave_group.authorize(ip_protocol='icmp', from_port=-1, to_port=-1, - src_group=slave_group) - slave_group.authorize(ip_protocol='tcp', from_port=0, to_port=65535, - src_group=slave_group) - slave_group.authorize(ip_protocol='udp', from_port=0, to_port=65535, - src_group=slave_group) - slave_group.authorize('tcp', 22, 22, authorized_address) - slave_group.authorize('tcp', 8080, 8081, authorized_address) - slave_group.authorize('tcp', 50060, 50060, authorized_address) - slave_group.authorize('tcp', 50075, 50075, authorized_address) - slave_group.authorize('tcp', 60060, 60060, authorized_address) - slave_group.authorize('tcp', 60075, 60075, authorized_address) + subordinate_group.authorize(ip_protocol='icmp', from_port=-1, to_port=-1, + src_group=main_group) + subordinate_group.authorize(ip_protocol='tcp', from_port=0, to_port=65535, + src_group=main_group) + subordinate_group.authorize(ip_protocol='udp', from_port=0, to_port=65535, + src_group=main_group) + subordinate_group.authorize(ip_protocol='icmp', from_port=-1, to_port=-1, + src_group=subordinate_group) + subordinate_group.authorize(ip_protocol='tcp', from_port=0, to_port=65535, + src_group=subordinate_group) + subordinate_group.authorize(ip_protocol='udp', from_port=0, to_port=65535, + src_group=subordinate_group) + subordinate_group.authorize('tcp', 22, 22, authorized_address) + subordinate_group.authorize('tcp', 8080, 8081, authorized_address) + subordinate_group.authorize('tcp', 50060, 50060, authorized_address) + subordinate_group.authorize('tcp', 50075, 50075, authorized_address) + subordinate_group.authorize('tcp', 60060, 60060, authorized_address) + subordinate_group.authorize('tcp', 60075, 60075, authorized_address) # Check if instances are already running in our groups - existing_masters, existing_slaves = get_existing_cluster(conn, opts, cluster_name, + existing_mains, existing_subordinates = get_existing_cluster(conn, opts, cluster_name, die_on_error=False) - if existing_slaves or (existing_masters and not opts.use_existing_master): + if existing_subordinates or (existing_mains and not opts.use_existing_main): print("ERROR: There are already instances running in group %s or %s" % - (master_group.name, slave_group.name), file=stderr) + (main_group.name, subordinate_group.name), file=stderr) sys.exit(1) # Figure out Spark AMI @@ -609,32 +609,32 @@ def launch_cluster(conn, opts, cluster_name): name = '/dev/sd' + string.ascii_letters[i + 1] block_map[name] = dev - # Launch slaves + # Launch subordinates if opts.spot_price is not None: # Launch spot instances with the requested price - print("Requesting %d slaves as spot instances with price $%.3f" % - (opts.slaves, opts.spot_price)) + print("Requesting %d subordinates as spot instances with price $%.3f" % + (opts.subordinates, opts.spot_price)) zones = get_zones(conn, opts) num_zones = len(zones) i = 0 my_req_ids = [] for zone in zones: - num_slaves_this_zone = get_partition(opts.slaves, num_zones, i) - slave_reqs = conn.request_spot_instances( + num_subordinates_this_zone = get_partition(opts.subordinates, num_zones, i) + subordinate_reqs = conn.request_spot_instances( price=opts.spot_price, image_id=opts.ami, launch_group="launch-group-%s" % cluster_name, placement=zone, - count=num_slaves_this_zone, + count=num_subordinates_this_zone, key_name=opts.key_pair, - security_group_ids=[slave_group.id] + additional_group_ids, + security_group_ids=[subordinate_group.id] + additional_group_ids, instance_type=opts.instance_type, block_device_map=block_map, subnet_id=opts.subnet_id, placement_group=opts.placement_group, user_data=user_data_content, instance_profile_name=opts.instance_profile_name) - my_req_ids += [req.id for req in slave_reqs] + my_req_ids += [req.id for req in subordinate_reqs] i += 1 print("Waiting for spot instances to be granted...") @@ -649,23 +649,23 @@ def launch_cluster(conn, opts, cluster_name): for i in my_req_ids: if i in id_to_req and id_to_req[i].state == "active": active_instance_ids.append(id_to_req[i].instance_id) - if len(active_instance_ids) == opts.slaves: - print("All %d slaves granted" % opts.slaves) + if len(active_instance_ids) == opts.subordinates: + print("All %d subordinates granted" % opts.subordinates) reservations = conn.get_all_reservations(active_instance_ids) - slave_nodes = [] + subordinate_nodes = [] for r in reservations: - slave_nodes += r.instances + subordinate_nodes += r.instances break else: - print("%d of %d slaves granted, waiting longer" % ( - len(active_instance_ids), opts.slaves)) + print("%d of %d subordinates granted, waiting longer" % ( + len(active_instance_ids), opts.subordinates)) except: print("Canceling spot instance requests") conn.cancel_spot_instance_requests(my_req_ids) # Log a warning if any of these requests actually launched instances: - (master_nodes, slave_nodes) = get_existing_cluster( + (main_nodes, subordinate_nodes) = get_existing_cluster( conn, opts, cluster_name, die_on_error=False) - running = len(master_nodes) + len(slave_nodes) + running = len(main_nodes) + len(subordinate_nodes) if running: print(("WARNING: %d instances are still running" % running), file=stderr) sys.exit(0) @@ -674,48 +674,48 @@ def launch_cluster(conn, opts, cluster_name): zones = get_zones(conn, opts) num_zones = len(zones) i = 0 - slave_nodes = [] + subordinate_nodes = [] for zone in zones: - num_slaves_this_zone = get_partition(opts.slaves, num_zones, i) - if num_slaves_this_zone > 0: - slave_res = image.run( + num_subordinates_this_zone = get_partition(opts.subordinates, num_zones, i) + if num_subordinates_this_zone > 0: + subordinate_res = image.run( key_name=opts.key_pair, - security_group_ids=[slave_group.id] + additional_group_ids, + security_group_ids=[subordinate_group.id] + additional_group_ids, instance_type=opts.instance_type, placement=zone, - min_count=num_slaves_this_zone, - max_count=num_slaves_this_zone, + min_count=num_subordinates_this_zone, + max_count=num_subordinates_this_zone, block_device_map=block_map, subnet_id=opts.subnet_id, placement_group=opts.placement_group, user_data=user_data_content, instance_initiated_shutdown_behavior=opts.instance_initiated_shutdown_behavior, instance_profile_name=opts.instance_profile_name) - slave_nodes += slave_res.instances - print("Launched {s} slave{plural_s} in {z}, regid = {r}".format( - s=num_slaves_this_zone, - plural_s=('' if num_slaves_this_zone == 1 else 's'), + subordinate_nodes += subordinate_res.instances + print("Launched {s} subordinate{plural_s} in {z}, regid = {r}".format( + s=num_subordinates_this_zone, + plural_s=('' if num_subordinates_this_zone == 1 else 's'), z=zone, - r=slave_res.id)) + r=subordinate_res.id)) i += 1 - # Launch or resume masters - if existing_masters: - print("Starting master...") - for inst in existing_masters: + # Launch or resume mains + if existing_mains: + print("Starting main...") + for inst in existing_mains: if inst.state not in ["shutting-down", "terminated"]: inst.start() - master_nodes = existing_masters + main_nodes = existing_mains else: - master_type = opts.master_instance_type - if master_type == "": - master_type = opts.instance_type + main_type = opts.main_instance_type + if main_type == "": + main_type = opts.instance_type if opts.zone == 'all': opts.zone = random.choice(conn.get_all_zones()).name - master_res = image.run( + main_res = image.run( key_name=opts.key_pair, - security_group_ids=[master_group.id] + additional_group_ids, - instance_type=master_type, + security_group_ids=[main_group.id] + additional_group_ids, + instance_type=main_type, placement=opts.zone, min_count=1, max_count=1, @@ -726,8 +726,8 @@ def launch_cluster(conn, opts, cluster_name): instance_initiated_shutdown_behavior=opts.instance_initiated_shutdown_behavior, instance_profile_name=opts.instance_profile_name) - master_nodes = master_res.instances - print("Launched master in %s, regid = %s" % (zone, master_res.id)) + main_nodes = main_res.instances + print("Launched main in %s, regid = %s" % (zone, main_res.id)) # This wait time corresponds to SPARK-4983 print("Waiting for AWS to propagate instance metadata...") @@ -740,24 +740,24 @@ def launch_cluster(conn, opts, cluster_name): map(str.strip, tag.split(':', 1)) for tag in opts.additional_tags.split(',') ) - for master in master_nodes: - master.add_tags( - dict(additional_tags, Name='{cn}-master-{iid}'.format(cn=cluster_name, iid=master.id)) + for main in main_nodes: + main.add_tags( + dict(additional_tags, Name='{cn}-main-{iid}'.format(cn=cluster_name, iid=main.id)) ) - for slave in slave_nodes: - slave.add_tags( - dict(additional_tags, Name='{cn}-slave-{iid}'.format(cn=cluster_name, iid=slave.id)) + for subordinate in subordinate_nodes: + subordinate.add_tags( + dict(additional_tags, Name='{cn}-subordinate-{iid}'.format(cn=cluster_name, iid=subordinate.id)) ) # Return all the instances - return (master_nodes, slave_nodes) + return (main_nodes, subordinate_nodes) def get_existing_cluster(conn, opts, cluster_name, die_on_error=True): """ Get the EC2 instances in an existing cluster if available. - Returns a tuple of lists of EC2 instance objects for the masters and slaves. + Returns a tuple of lists of EC2 instance objects for the mains and subordinates. """ print("Searching for existing cluster {c} in region {r}...".format( c=cluster_name, r=opts.region)) @@ -774,51 +774,51 @@ def get_instances(group_names): instances = itertools.chain.from_iterable(r.instances for r in reservations) return [i for i in instances if i.state not in ["shutting-down", "terminated"]] - master_instances = get_instances([cluster_name + "-master"]) - slave_instances = get_instances([cluster_name + "-slaves"]) + main_instances = get_instances([cluster_name + "-main"]) + subordinate_instances = get_instances([cluster_name + "-subordinates"]) - if any((master_instances, slave_instances)): - print("Found {m} master{plural_m}, {s} slave{plural_s}.".format( - m=len(master_instances), - plural_m=('' if len(master_instances) == 1 else 's'), - s=len(slave_instances), - plural_s=('' if len(slave_instances) == 1 else 's'))) + if any((main_instances, subordinate_instances)): + print("Found {m} main{plural_m}, {s} subordinate{plural_s}.".format( + m=len(main_instances), + plural_m=('' if len(main_instances) == 1 else 's'), + s=len(subordinate_instances), + plural_s=('' if len(subordinate_instances) == 1 else 's'))) - if not master_instances and die_on_error: - print("ERROR: Could not find a master for cluster {c} in region {r}.".format( + if not main_instances and die_on_error: + print("ERROR: Could not find a main for cluster {c} in region {r}.".format( c=cluster_name, r=opts.region), file=sys.stderr) sys.exit(1) - return (master_instances, slave_instances) + return (main_instances, subordinate_instances) -# Execute a cmd on master and slave nodes -def ssh_cluster(master_nodes, slave_nodes, opts, cmd): - master = get_dns_name(master_nodes[0], opts.private_ips) - ssh(master, opts, cmd) - for slave in slave_nodes: - slave_address = get_dns_name(slave, opts.private_ips) - ssh(slave_address, opts, cmd) +# Execute a cmd on main and subordinate nodes +def ssh_cluster(main_nodes, subordinate_nodes, opts, cmd): + main = get_dns_name(main_nodes[0], opts.private_ips) + ssh(main, opts, cmd) + for subordinate in subordinate_nodes: + subordinate_address = get_dns_name(subordinate, opts.private_ips) + ssh(subordinate_address, opts, cmd) # Deploy configuration files and run setup scripts on a newly launched # or started EC2 cluster. -def setup_cluster(conn, master_nodes, slave_nodes, opts, deploy_ssh_key): - master = get_dns_name(master_nodes[0], opts.private_ips) +def setup_cluster(conn, main_nodes, subordinate_nodes, opts, deploy_ssh_key): + main = get_dns_name(main_nodes[0], opts.private_ips) if deploy_ssh_key: - print("Generating cluster's SSH key on master...") + print("Generating cluster's SSH key on main...") key_setup = """ [ -f ~/.ssh/id_rsa ] || (ssh-keygen -q -t rsa -N '' -f ~/.ssh/id_rsa && cat ~/.ssh/id_rsa.pub >> ~/.ssh/authorized_keys) """ - ssh(master, opts, key_setup) - dot_ssh_tar = ssh_read(master, opts, ['tar', 'c', '.ssh']) - print("Transferring cluster's SSH key to slaves...") - for slave in slave_nodes: - slave_address = get_dns_name(slave, opts.private_ips) - print(slave_address) - ssh_write(slave_address, opts, ['tar', 'x'], dot_ssh_tar) + ssh(main, opts, key_setup) + dot_ssh_tar = ssh_read(main, opts, ['tar', 'c', '.ssh']) + print("Transferring cluster's SSH key to subordinates...") + for subordinate in subordinate_nodes: + subordinate_address = get_dns_name(subordinate, opts.private_ips) + print(subordinate_address) + ssh_write(subordinate_address, opts, ['tar', 'x'], dot_ssh_tar) modules = ['spark', 'ephemeral-hdfs', 'persistent-hdfs', 'mapreduce', 'spark-standalone'] @@ -835,10 +835,10 @@ def setup_cluster(conn, master_nodes, slave_nodes, opts, deploy_ssh_key): # NOTE: We should clone the repository before running deploy_files to # prevent ec2-variables.sh from being overwritten - print("Cloning spark-ec2 scripts from {r}/tree/{b} on master...".format( + print("Cloning spark-ec2 scripts from {r}/tree/{b} on main...".format( r=opts.spark_ec2_git_repo, b=opts.spark_ec2_git_branch)) ssh( - host=master, + host=main, opts=opts, command="rm -rf spark-ec2" + " && " @@ -846,37 +846,37 @@ def setup_cluster(conn, master_nodes, slave_nodes, opts, deploy_ssh_key): b=opts.spark_ec2_git_branch) ) - print("Deploying files to master...") + print("Deploying files to main...") deploy_files( conn=conn, root_dir=SPARK_EC2_DIR + "/" + "deploy.generic", opts=opts, - master_nodes=master_nodes, - slave_nodes=slave_nodes, + main_nodes=main_nodes, + subordinate_nodes=subordinate_nodes, modules=modules ) if opts.deploy_root_dir is not None: - print("Deploying {s} to master...".format(s=opts.deploy_root_dir)) + print("Deploying {s} to main...".format(s=opts.deploy_root_dir)) deploy_user_files( root_dir=opts.deploy_root_dir, opts=opts, - master_nodes=master_nodes + main_nodes=main_nodes ) - print("Running setup on master...") - setup_spark_cluster(master, opts) + print("Running setup on main...") + setup_spark_cluster(main, opts) print("Done!") -def setup_spark_cluster(master, opts): - ssh(master, opts, "chmod u+x spark-ec2/setup.sh") - ssh(master, opts, "spark-ec2/setup.sh") - print("Spark standalone cluster started at http://%s:8080" % master) +def setup_spark_cluster(main, opts): + ssh(main, opts, "chmod u+x spark-ec2/setup.sh") + ssh(main, opts, "spark-ec2/setup.sh") + print("Spark standalone cluster started at http://%s:8080" % main) if opts.ganglia: - print("Ganglia started at http://%s:5080/ganglia" % master) + print("Ganglia started at http://%s:5080/ganglia" % main) def is_ssh_available(host, opts, print_ssh_output=True): @@ -1047,13 +1047,13 @@ def get_num_disks(instance_type): # Deploy the configuration file templates in a given local directory to # a cluster, filling in any template parameters with information about the -# cluster (e.g. lists of masters and slaves). Files are only deployed to -# the first master instance in the cluster, and we expect the setup +# cluster (e.g. lists of mains and subordinates). Files are only deployed to +# the first main instance in the cluster, and we expect the setup # script to be run on that instance to copy them to other nodes. # # root_dir should be an absolute path to the directory with the files we want to deploy. -def deploy_files(conn, root_dir, opts, master_nodes, slave_nodes, modules): - active_master = get_dns_name(master_nodes[0], opts.private_ips) +def deploy_files(conn, root_dir, opts, main_nodes, subordinate_nodes, modules): + active_main = get_dns_name(main_nodes[0], opts.private_ips) num_disks = get_num_disks(opts.instance_type) hdfs_data_dirs = "/mnt/ephemeral-hdfs/data" @@ -1065,7 +1065,7 @@ def deploy_files(conn, root_dir, opts, master_nodes, slave_nodes, modules): mapred_local_dirs += ",/mnt%d/hadoop/mrlocal" % i spark_local_dirs += ",/mnt%d/spark" % i - cluster_url = "%s:7077" % active_master + cluster_url = "%s:7077" % active_main if "." in opts.spark_version: # Pre-built Spark deploy @@ -1078,13 +1078,13 @@ def deploy_files(conn, root_dir, opts, master_nodes, slave_nodes, modules): print("Deploying Spark via git hash; Tachyon won't be set up") modules = filter(lambda x: x != "tachyon", modules) - master_addresses = [get_dns_name(i, opts.private_ips) for i in master_nodes] - slave_addresses = [get_dns_name(i, opts.private_ips) for i in slave_nodes] + main_addresses = [get_dns_name(i, opts.private_ips) for i in main_nodes] + subordinate_addresses = [get_dns_name(i, opts.private_ips) for i in subordinate_nodes] worker_instances_str = "%d" % opts.worker_instances if opts.worker_instances else "" template_vars = { - "master_list": '\n'.join(master_addresses), - "active_master": active_master, - "slave_list": '\n'.join(slave_addresses), + "main_list": '\n'.join(main_addresses), + "active_main": active_main, + "subordinate_list": '\n'.join(subordinate_addresses), "cluster_url": cluster_url, "hdfs_data_dirs": hdfs_data_dirs, "mapred_local_dirs": mapred_local_dirs, @@ -1095,7 +1095,7 @@ def deploy_files(conn, root_dir, opts, master_nodes, slave_nodes, modules): "tachyon_version": tachyon_v, "hadoop_major_version": opts.hadoop_major_version, "spark_worker_instances": worker_instances_str, - "spark_master_opts": opts.master_opts + "spark_main_opts": opts.main_opts } if opts.copy_aws_credentials: @@ -1125,12 +1125,12 @@ def deploy_files(conn, root_dir, opts, master_nodes, slave_nodes, modules): text = text.replace("{{" + key + "}}", template_vars[key]) dest.write(text) dest.close() - # rsync the whole directory over to the master machine + # rsync the whole directory over to the main machine command = [ 'rsync', '-rv', '-e', stringify_command(ssh_command(opts)), "%s/" % tmp_dir, - "%s@%s:/" % (opts.user, active_master) + "%s@%s:/" % (opts.user, active_main) ] subprocess.check_call(command) # Remove the temp directory we created above @@ -1140,16 +1140,16 @@ def deploy_files(conn, root_dir, opts, master_nodes, slave_nodes, modules): # Deploy a given local directory to a cluster, WITHOUT parameter substitution. # Note that unlike deploy_files, this works for binary files. # Also, it is up to the user to add (or not) the trailing slash in root_dir. -# Files are only deployed to the first master instance in the cluster. +# Files are only deployed to the first main instance in the cluster. # # root_dir should be an absolute path. -def deploy_user_files(root_dir, opts, master_nodes): - active_master = get_dns_name(master_nodes[0], opts.private_ips) +def deploy_user_files(root_dir, opts, main_nodes): + active_main = get_dns_name(main_nodes[0], opts.private_ips) command = [ 'rsync', '-rv', '-e', stringify_command(ssh_command(opts)), "%s" % root_dir, - "%s@%s:/" % (opts.user, active_master) + "%s@%s:/" % (opts.user, active_main) ] subprocess.check_call(command) @@ -1249,10 +1249,10 @@ def get_zones(conn, opts): # Gets the number of items in a partition def get_partition(total, num_partitions, current_partitions): - num_slaves_this_zone = total // num_partitions + num_subordinates_this_zone = total // num_partitions if (total % num_partitions) - current_partitions > 0: - num_slaves_this_zone += 1 - return num_slaves_this_zone + num_subordinates_this_zone += 1 + return num_subordinates_this_zone # Gets the IP address, taking into account the --private-ips flag @@ -1302,21 +1302,21 @@ def real_main(): print("Warning: Unrecognized EC2 instance type for instance-type: {t}".format( t=opts.instance_type), file=stderr) - if opts.master_instance_type != "": - if opts.master_instance_type not in EC2_INSTANCE_TYPES: - print("Warning: Unrecognized EC2 instance type for master-instance-type: {t}".format( - t=opts.master_instance_type), file=stderr) + if opts.main_instance_type != "": + if opts.main_instance_type not in EC2_INSTANCE_TYPES: + print("Warning: Unrecognized EC2 instance type for main-instance-type: {t}".format( + t=opts.main_instance_type), file=stderr) # Since we try instance types even if we can't resolve them, we check if they resolve first # and, if they do, see if they resolve to the same virtualization type. if opts.instance_type in EC2_INSTANCE_TYPES and \ - opts.master_instance_type in EC2_INSTANCE_TYPES: + opts.main_instance_type in EC2_INSTANCE_TYPES: if EC2_INSTANCE_TYPES[opts.instance_type] != \ - EC2_INSTANCE_TYPES[opts.master_instance_type]: - print("Error: spark-ec2 currently does not support having a master and slaves " + EC2_INSTANCE_TYPES[opts.main_instance_type]: + print("Error: spark-ec2 currently does not support having a main and subordinates " "with different AMI virtualization types.", file=stderr) - print("master instance virtualization type: {t}".format( - t=EC2_INSTANCE_TYPES[opts.master_instance_type]), file=stderr) - print("slave instance virtualization type: {t}".format( + print("main instance virtualization type: {t}".format( + t=EC2_INSTANCE_TYPES[opts.main_instance_type]), file=stderr) + print("subordinate instance virtualization type: {t}".format( t=EC2_INSTANCE_TYPES[opts.instance_type]), file=stderr) sys.exit(1) @@ -1356,48 +1356,48 @@ def real_main(): opts.zone = random.choice(conn.get_all_zones()).name if action == "launch": - if opts.slaves <= 0: - print("ERROR: You have to start at least 1 slave", file=sys.stderr) + if opts.subordinates <= 0: + print("ERROR: You have to start at least 1 subordinate", file=sys.stderr) sys.exit(1) if opts.resume: - (master_nodes, slave_nodes) = get_existing_cluster(conn, opts, cluster_name) + (main_nodes, subordinate_nodes) = get_existing_cluster(conn, opts, cluster_name) else: - (master_nodes, slave_nodes) = launch_cluster(conn, opts, cluster_name) + (main_nodes, subordinate_nodes) = launch_cluster(conn, opts, cluster_name) wait_for_cluster_state( conn=conn, opts=opts, - cluster_instances=(master_nodes + slave_nodes), + cluster_instances=(main_nodes + subordinate_nodes), cluster_state='ssh-ready' ) - setup_cluster(conn, master_nodes, slave_nodes, opts, True) + setup_cluster(conn, main_nodes, subordinate_nodes, opts, True) elif action == "destroy": - (master_nodes, slave_nodes) = get_existing_cluster( + (main_nodes, subordinate_nodes) = get_existing_cluster( conn, opts, cluster_name, die_on_error=False) - if any(master_nodes + slave_nodes): + if any(main_nodes + subordinate_nodes): print("The following instances will be terminated:") - for inst in master_nodes + slave_nodes: + for inst in main_nodes + subordinate_nodes: print("> %s" % get_dns_name(inst, opts.private_ips)) print("ALL DATA ON ALL NODES WILL BE LOST!!") msg = "Are you sure you want to destroy the cluster {c}? (y/N) ".format(c=cluster_name) response = raw_input(msg) if response == "y": - print("Terminating master...") - for inst in master_nodes: + print("Terminating main...") + for inst in main_nodes: inst.terminate() - print("Terminating slaves...") - for inst in slave_nodes: + print("Terminating subordinates...") + for inst in subordinate_nodes: inst.terminate() # Delete security groups as well if opts.delete_groups: - group_names = [cluster_name + "-master", cluster_name + "-slaves"] + group_names = [cluster_name + "-main", cluster_name + "-subordinates"] wait_for_cluster_state( conn=conn, opts=opts, - cluster_instances=(master_nodes + slave_nodes), + cluster_instances=(main_nodes + subordinate_nodes), cluster_state='terminated' ) print("Deleting security groups (this will take some time)...") @@ -1441,38 +1441,38 @@ def real_main(): print("Try re-running in a few minutes.") elif action == "login": - (master_nodes, slave_nodes) = get_existing_cluster(conn, opts, cluster_name) - if not master_nodes[0].public_dns_name and not opts.private_ips: - print("Master has no public DNS name. Maybe you meant to specify --private-ips?") + (main_nodes, subordinate_nodes) = get_existing_cluster(conn, opts, cluster_name) + if not main_nodes[0].public_dns_name and not opts.private_ips: + print("Main has no public DNS name. Maybe you meant to specify --private-ips?") else: - master = get_dns_name(master_nodes[0], opts.private_ips) - print("Logging into master " + master + "...") + main = get_dns_name(main_nodes[0], opts.private_ips) + print("Logging into main " + main + "...") proxy_opt = [] if opts.proxy_port is not None: proxy_opt = ['-D', opts.proxy_port] subprocess.check_call( - ssh_command(opts) + proxy_opt + ['-t', '-t', "%s@%s" % (opts.user, master)]) + ssh_command(opts) + proxy_opt + ['-t', '-t', "%s@%s" % (opts.user, main)]) - elif action == "reboot-slaves": + elif action == "reboot-subordinates": response = raw_input( "Are you sure you want to reboot the cluster " + - cluster_name + " slaves?\n" + - "Reboot cluster slaves " + cluster_name + " (y/N): ") + cluster_name + " subordinates?\n" + + "Reboot cluster subordinates " + cluster_name + " (y/N): ") if response == "y": - (master_nodes, slave_nodes) = get_existing_cluster( + (main_nodes, subordinate_nodes) = get_existing_cluster( conn, opts, cluster_name, die_on_error=False) - print("Rebooting slaves...") - for inst in slave_nodes: + print("Rebooting subordinates...") + for inst in subordinate_nodes: if inst.state not in ["shutting-down", "terminated"]: print("Rebooting " + inst.id) inst.reboot() - elif action == "get-master": - (master_nodes, slave_nodes) = get_existing_cluster(conn, opts, cluster_name) - if not master_nodes[0].public_dns_name and not opts.private_ips: - print("Master has no public DNS name. Maybe you meant to specify --private-ips?") + elif action == "get-main": + (main_nodes, subordinate_nodes) = get_existing_cluster(conn, opts, cluster_name) + if not main_nodes[0].public_dns_name and not opts.private_ips: + print("Main has no public DNS name. Maybe you meant to specify --private-ips?") else: - print(get_dns_name(master_nodes[0], opts.private_ips)) + print(get_dns_name(main_nodes[0], opts.private_ips)) elif action == "stop": response = raw_input( @@ -1480,17 +1480,17 @@ def real_main(): cluster_name + "?\nDATA ON EPHEMERAL DISKS WILL BE LOST, " + "BUT THE CLUSTER WILL KEEP USING SPACE ON\n" + "AMAZON EBS IF IT IS EBS-BACKED!!\n" + - "All data on spot-instance slaves will be lost.\n" + + "All data on spot-instance subordinates will be lost.\n" + "Stop cluster " + cluster_name + " (y/N): ") if response == "y": - (master_nodes, slave_nodes) = get_existing_cluster( + (main_nodes, subordinate_nodes) = get_existing_cluster( conn, opts, cluster_name, die_on_error=False) - print("Stopping master...") - for inst in master_nodes: + print("Stopping main...") + for inst in main_nodes: if inst.state not in ["shutting-down", "terminated"]: inst.stop() - print("Stopping slaves...") - for inst in slave_nodes: + print("Stopping subordinates...") + for inst in subordinate_nodes: if inst.state not in ["shutting-down", "terminated"]: if inst.spot_instance_request_id: inst.terminate() @@ -1498,33 +1498,33 @@ def real_main(): inst.stop() elif action == "start": - (master_nodes, slave_nodes) = get_existing_cluster(conn, opts, cluster_name) - print("Starting slaves...") - for inst in slave_nodes: + (main_nodes, subordinate_nodes) = get_existing_cluster(conn, opts, cluster_name) + print("Starting subordinates...") + for inst in subordinate_nodes: if inst.state not in ["shutting-down", "terminated"]: inst.start() - print("Starting master...") - for inst in master_nodes: + print("Starting main...") + for inst in main_nodes: if inst.state not in ["shutting-down", "terminated"]: inst.start() wait_for_cluster_state( conn=conn, opts=opts, - cluster_instances=(master_nodes + slave_nodes), + cluster_instances=(main_nodes + subordinate_nodes), cluster_state='ssh-ready' ) # Determine types of running instances - existing_master_type = master_nodes[0].instance_type - existing_slave_type = slave_nodes[0].instance_type - # Setting opts.master_instance_type to the empty string indicates we - # have the same instance type for the master and the slaves - if existing_master_type == existing_slave_type: - existing_master_type = "" - opts.master_instance_type = existing_master_type - opts.instance_type = existing_slave_type - - setup_cluster(conn, master_nodes, slave_nodes, opts, False) + existing_main_type = main_nodes[0].instance_type + existing_subordinate_type = subordinate_nodes[0].instance_type + # Setting opts.main_instance_type to the empty string indicates we + # have the same instance type for the main and the subordinates + if existing_main_type == existing_subordinate_type: + existing_main_type = "" + opts.main_instance_type = existing_main_type + opts.instance_type = existing_subordinate_type + + setup_cluster(conn, main_nodes, subordinate_nodes, opts, False) else: print("Invalid action: %s" % action, file=stderr) diff --git a/tensorflowonspark/TFCluster.py b/tensorflowonspark/TFCluster.py index b19f56d..d7508d8 100644 --- a/tensorflowonspark/TFCluster.py +++ b/tensorflowonspark/TFCluster.py @@ -208,7 +208,7 @@ def tensorboard_url(self): def run(sc, map_fun, tf_args, num_executors, num_ps, tensorboard=False, input_mode=InputMode.TENSORFLOW, - log_dir=None, driver_ps_nodes=False, master_node=None, reservation_timeout=600, queues=['input', 'output', 'error'], + log_dir=None, driver_ps_nodes=False, main_node=None, reservation_timeout=600, queues=['input', 'output', 'error'], eval_node=False): """Starts the TensorFlowOnSpark cluster and Runs the TensorFlow "main" function on the Spark executors @@ -222,7 +222,7 @@ def run(sc, map_fun, tf_args, num_executors, num_ps, tensorboard=False, input_mo :input_mode: TFCluster.InputMode :log_dir: directory to save tensorboard event logs. If None, defaults to a fixed path on local filesystem. :driver_ps_nodes: run the PS nodes on the driver locally instead of on the spark executors; this help maximizing computing resources (esp. GPU). You will need to set cluster_size = num_executors + num_ps - :master_node: name of the "master" or "chief" node in the cluster_template, used for `tf.estimator` applications. + :main_node: name of the "main" or "chief" node in the cluster_template, used for `tf.estimator` applications. :reservation_timeout: number of seconds after which cluster reservation times out (600 sec default) :queues: *INTERNAL_USE* :eval_node: run evaluator node for distributed Tensorflow @@ -239,13 +239,13 @@ def run(sc, map_fun, tf_args, num_executors, num_ps, tensorboard=False, input_mo raise Exception('running evaluator nodes is only supported in InputMode.TENSORFLOW') # compute size of TF cluster and validate against number of Spark executors - num_master = 1 if master_node else 0 + num_main = 1 if main_node else 0 num_eval = 1 if eval_node else 0 - num_workers = max(num_executors - num_ps - num_eval - num_master, 0) - total_nodes = num_ps + num_master + num_eval + num_workers + num_workers = max(num_executors - num_ps - num_eval - num_main, 0) + total_nodes = num_ps + num_main + num_eval + num_workers assert total_nodes == num_executors, "TensorFlow cluster requires {} nodes, but only {} executors available".format(total_nodes, num_executors) - assert num_master + num_workers > 0, "TensorFlow cluster requires at least one worker or master/chief node" + assert num_main + num_workers > 0, "TensorFlow cluster requires at least one worker or main/chief node" # create a cluster template for scheduling TF nodes onto executors executors = list(range(num_executors)) @@ -254,8 +254,8 @@ def run(sc, map_fun, tf_args, num_executors, num_ps, tensorboard=False, input_mo if num_ps > 0: cluster_template['ps'] = executors[:num_ps] del executors[:num_ps] - if master_node: - cluster_template[master_node] = executors[:1] + if main_node: + cluster_template[main_node] = executors[:1] del executors[:1] if eval_node: cluster_template['evaluator'] = executors[:1] diff --git a/tensorflowonspark/TFSparkNode.py b/tensorflowonspark/TFSparkNode.py index 8ae22ea..b1dfbcd 100644 --- a/tensorflowonspark/TFSparkNode.py +++ b/tensorflowonspark/TFSparkNode.py @@ -274,8 +274,8 @@ def _mapfn(iter): hosts.append("{0}:{1}".format(nhost, nport)) cluster_spec[njob] = hosts - # update TF_CONFIG if cluster spec has a 'master' node (i.e. tf.estimator) - if 'master' in cluster_spec or 'chief' in cluster_spec: + # update TF_CONFIG if cluster spec has a 'main' node (i.e. tf.estimator) + if 'main' in cluster_spec or 'chief' in cluster_spec: tf_config = json.dumps({ 'cluster': cluster_spec, 'task': {'type': job_name, 'index': task_index}, diff --git a/test/test.py b/test/test.py index d02336d..cbc1308 100644 --- a/test/test.py +++ b/test/test.py @@ -10,8 +10,8 @@ class SparkTest(unittest.TestCase): @classmethod def setUpClass(cls): - master = os.getenv('MASTER') - assert master is not None, "Please start a Spark standalone cluster and export MASTER to your env." + main = os.getenv('MASTER') + assert main is not None, "Please start a Spark standalone cluster and export MASTER to your env." num_workers = os.getenv('SPARK_WORKER_INSTANCES') assert num_workers is not None, "Please export SPARK_WORKER_INSTANCES to your env." @@ -21,7 +21,7 @@ def setUpClass(cls): assert spark_jars, "Please add path to tensorflow/ecosystem/hadoop jar to SPARK_CLASSPATH." cls.conf = SparkConf().set('spark.jars', spark_jars) - cls.sc = SparkContext(master, cls.__name__, conf=cls.conf) + cls.sc = SparkContext(main, cls.__name__, conf=cls.conf) cls.spark = SparkSession.builder.getOrCreate() @classmethod diff --git a/test/test_pipeline.py b/test/test_pipeline.py index 0950260..69c94ed 100644 --- a/test/test_pipeline.py +++ b/test/test_pipeline.py @@ -191,7 +191,7 @@ def sparse_train(args, ctx): cost = tf.reduce_mean(tf.square(y_ - y), name='cost') optimizer = tf.train.GradientDescentOptimizer(0.1).minimize(cost, global_step) - with tf.train.MonitoredTrainingSession(master=server.target, + with tf.train.MonitoredTrainingSession(main=server.target, is_chief=(ctx.task_index == 0), checkpoint_dir=args.model_dir, save_checkpoint_steps=20) as sess: @@ -344,7 +344,7 @@ def end(self, session): optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(cost, global_step) chief_hooks = [ExportHook(ctx.absolute_path(args.export_dir), x, y)] if args.export_dir else [] - with tf.train.MonitoredTrainingSession(master=server.target, + with tf.train.MonitoredTrainingSession(main=server.target, is_chief=(ctx.task_index == 0), checkpoint_dir=args.model_dir, chief_only_hooks=chief_hooks) as sess: @@ -385,7 +385,7 @@ def _get_examples(batch_size): cost = tf.reduce_mean(tf.square(y_ - y), name='cost') optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(cost, global_step) - with tf.train.MonitoredTrainingSession(master=server.target, + with tf.train.MonitoredTrainingSession(main=server.target, is_chief=(ctx.task_index == 0), checkpoint_dir=args.model_dir) as sess: step = 0