def test_run_spark_job(self):
# given ske
ske = "opends spark it-test"
Staroid().cluster().create(ske, TEST_REGION)
# init with ske name
ods.init(ske=ske)
spark = ods.spark("test", delta=True).session()
df = spark.createDataFrame([{"hello": "world"} for x in range(100)])
self.assertEqual(100, df.count())
# delta
s3_delta_table = "{}/delta-table".format(os.environ["S3_LOCATION"])
spark.range(
0, 5).write.format("delta").mode("overwrite").save(s3_delta_table)
delta = spark.read.format("delta").load(s3_delta_table)
self.assertEqual(5, delta.count())
# delete cluster instance
spark.stop()
ods.spark("test").delete()
# clean up
Staroid().cluster().delete(ske)
def test_read_config(self):
# unset env
at = None
ac = None
if "STAROID_ACCESS_TOKEN" in os.environ:
at = os.environ["STAROID_ACCESS_TOKEN"]
del os.environ["STAROID_ACCESS_TOKEN"]
if "STAROID_ACCOUNT" in os.environ:
ac = os.environ["STAROID_ACCOUNT"]
del os.environ["STAROID_ACCOUNT"]
# given
fp = tempfile.NamedTemporaryFile()
fp.write(b"access_token: abc\naccount: GITHUB/user1")
fp.flush()
# when
s = Staroid(config_path=fp.name)
# then
self.assertEqual("abc", s.get_access_token())
self.assertEqual("GITHUB/user1", s.get_account())
# restore env
if at != None:
os.environ["STAROID_ACCESS_TOKEN"] = at
if ac != None:
os.environ["STAROID_ACCOUNT"] = ac
def test_crud(self):
# given
s = Staroid(access_token=os.environ["STAROID_ACCESS_TOKEN"],
account=os.environ["STAROID_ACCOUNT"])
all_accounts = s.get_all_accounts()
clusters_before = s.cluster().get_all()
# when create
c = s.cluster().create("staroid-python it-test")
# then
self.assertEqual("staroid-python it-test", c.name())
# when already exists on create
c = s.cluster().create("staroid-python it-test")
# then
self.assertEqual("staroid-python it-test", c.name())
# when list
clusters_after = s.cluster().get_all()
# then
self.assertEqual(len(clusters_after), len(clusters_before) + 1)
# when delete
s.cluster().delete("staroid-python it-test")
# then
clusters_after_deleted = s.cluster().get_all()
self.assertEqual(len(clusters_after_deleted), len(clusters_before))
def __init__(self, provider_config, cluster_name):
NodeProvider.__init__(self, provider_config, cluster_name)
self.__cached = {}
self.__star = Staroid(access_token=provider_config["access_token"],
account=provider_config["account"])
self.__ske = self._get_config_or_env(provider_config, "ske",
"STAROID_SKE")
self.__ske_region = self._get_config_or_env(provider_config,
"ske_region",
"STAROID_SKE_REGION")
def test_download_chisel(self):
# given
tmp_dir = tempfile.mkdtemp()
s = Staroid(cache_dir=tmp_dir)
# when
chisel_path = s.get_chisel_path()
# then
self.assertIsNotNone(chisel_path)
self.assertTrue(os.path.isfile(chisel_path))
# clean up
shutil.rmtree(pathlib.Path(tmp_dir))
def test_read_default_account(self):
# given access_token is set but account is not set
ac = None
if "STAROID_ACCOUNT" in os.environ:
ac = os.environ["STAROID_ACCOUNT"]
del os.environ["STAROID_ACCOUNT"]
# when
s = Staroid()
# then
self.assertNotEqual(None, s.get_account())
# restore env
if ac != None:
os.environ["STAROID_ACCOUNT"] = ac
def __init__(self, staroid=None, ske=None, cache_dir=None):
self.__ske = None
if staroid == None:
self._staroid = Staroid()
else:
self._staroid = staroid
if cache_dir == None:
self.__cache_dir = "{}/.ods".format(str(Path.home()))
else:
self.__cache_dir = cache_dir
# configure from env var
if "STAROID_SKE" in os.environ:
self.__ske = os.environ["STAROID_SKE"]
# configure from args
if ske != None:
self.__ske = ske
def test_crud_namespace(self):
# given
s = Staroid(access_token=os.environ["STAROID_ACCESS_TOKEN"], account=os.environ["STAROID_ACCOUNT"])
c = s.cluster().create("staroid-python it-test-namespace")
# when create a namespace
ns_api = s.namespace(c)
ns = ns_api.create("instance1", "GITHUB/staroids/namespace:master")
# then namespace becomes RUNNING
wait_for_phase(ns_api, ns, "RUNNING")
self.assertEqual("RUNNING", ns_api.get_by_id(ns.id()).phase())
# start shell
ns_api.shell_start("instance1")
resources = ns_api.get_all_resources("instance1")
self.assertTrue(len(resources["services"]) > 0)
# start tunnel
ns_api.start_tunnel("instance1", ["57683:localhost:57683"])
# stop tunnel
ns_api.stop_tunnel("instance1")
# stop shell
ns_api.shell_stop("instance1")
# pause
ns = ns_api.stop("instance1")
wait_for_phase(ns_api, ns, "PAUSED")
self.assertEqual("PAUSED", ns_api.get_by_id(ns.id()).phase())
# resume
ns = ns_api.start("instance1")
wait_for_phase(ns_api, ns, "RUNNING")
self.assertEqual("RUNNING", ns_api.get_by_id(ns.id()).phase())
# when delete a namespace
ns = ns_api.delete("instance1")
# then namespace becomes REMOVED
wait_for_phase(ns_api, ns, "REMOVED")
self.assertEqual("REMOVED", ns_api.get_by_id(ns.id()).phase())
# when delete
s.cluster().delete("staroid-python it-test-namespace")
class StaroidNodeProvider(NodeProvider):
def __init__(self, provider_config, cluster_name):
NodeProvider.__init__(self, provider_config, cluster_name)
self.__cached = {}
self.__star = Staroid(
access_token=provider_config["access_token"],
account=provider_config["account"],
)
self.__ske = self._get_config_or_env(provider_config, "ske", "STAROID_SKE")
self.__ske_region = self._get_config_or_env(
provider_config, "ske_region", "STAROID_SKE_REGION"
)
self._requests_lib = _try_import_requests()
def _get_config_or_env(self, config, config_key, env_name):
value = None
# check env first, so config can override env later
if env_name in os.environ:
value = os.environ[env_name]
if config_key in config and config[config_key] is not None:
value = config[config_key]
return value
def _connect_kubeapi_incluster(self, instance_name):
if not os.path.isdir("/var/run/secrets/kubernetes.io/serviceaccount"):
return None
kube_conf = config.load_incluster_config()
kube_client = client.ApiClient(kube_conf)
with open(
"/var/run/secrets/kubernetes.io/serviceaccount/namespace", "r"
) as file:
namespace = file.read().replace("\n", "")
self.__cached[instance_name] = {"kube_client": kube_client, "api_server": None}
self.namespace = namespace
return kube_client
def _connect_kubeapi(self, instance_name):
if instance_name in self.__cached:
return self.__cached[instance_name]["kube_client"]
# try incluster configuration first
kube_client = self._connect_kubeapi_incluster(instance_name)
if kube_client is not None:
return kube_client
# check if ske exists
cluster_api = self.__star.cluster()
ske = cluster_api.get(self.__ske)
if ske is None: # ske not exists
return None
# check if ray cluster instance exists
ns_api = self.__star.namespace(ske)
ns = ns_api.get(instance_name)
if ns is None: # instance not exists
return None
# check if staroid namespace is not PAUSED (stopped)
# or INACTIVE (terminated)
if ns.status() != "ACTIVE":
return None
# wait for the staroid namespace to be started
start_time = time.time()
timeout = 300
started = False
while time.time() - start_time < timeout:
if ns.phase() == "RUNNING":
started = True
break
time.sleep(3)
ns = ns_api.get(instance_name)
if started is False:
logger.info(log_prefix + "fail to start namespace")
return None
# start a shell service to create secure tunnel
ns_api.shell_start(instance_name)
local_port = find_free_port()
# fixed port number for kube api access through
# shell service in staroid
remote_port = 57683
# start a secure tunnel
ns_api.start_tunnel(
instance_name, ["{}:localhost:{}".format(local_port, remote_port)]
)
# wait for tunnel to be established by checking /version
local_kube_api_addr = "http://localhost:{}".format(local_port)
start_time = time.time()
established = False
while time.time() - start_time < timeout:
try:
r = self._requests_lib.get(
"{}/version".format(local_kube_api_addr), timeout=(3, 5)
)
if r.status_code == 200:
established = True
break
except self._requests_lib.exceptions.ConnectionError:
pass
time.sleep(3)
if established:
kube_conf = client.Configuration()
kube_conf.host = local_kube_api_addr
kube_client = client.ApiClient(kube_conf)
self.__cached[instance_name] = {
"kube_client": kube_client,
"api_server": local_kube_api_addr,
}
self.namespace = ns.namespace()
return kube_client
else:
self.__cached[instance_name] = None
return None
def non_terminated_nodes(self, tag_filters):
instance_name = self.cluster_name
kube_client = self._connect_kubeapi(instance_name)
if kube_client is None:
return []
core_api = client.CoreV1Api(kube_client)
# Match pods that are in the 'Pending' or 'Running' phase.
# Unfortunately there is no OR operator in field selectors, so we
# have to match on NOT any of the other phases.
field_selector = ",".join(
[
"status.phase!=Failed",
"status.phase!=Unknown",
"status.phase!=Succeeded",
"status.phase!=Terminating",
]
)
tag_filters[TAG_RAY_CLUSTER_NAME] = self.cluster_name
label_selector = to_label_selector(tag_filters)
pod_list = core_api.list_namespaced_pod(
self.namespace, field_selector=field_selector, label_selector=label_selector
)
return [pod.metadata.name for pod in pod_list.items]
def is_running(self, node_id):
kube_client = self.__cached[self.cluster_name]["kube_client"]
core_api = client.CoreV1Api(kube_client)
pod = core_api.read_namespaced_pod(node_id, self.namespace)
return pod.status.phase == "Running"
def is_terminated(self, node_id):
kube_client = self.__cached[self.cluster_name]["kube_client"]
core_api = client.CoreV1Api(kube_client)
pod = core_api.read_namespaced_pod(node_id, self.namespace)
return pod.status.phase not in ["Running", "Pending"]
def node_tags(self, node_id):
kube_client = self.__cached[self.cluster_name]["kube_client"]
core_api = client.CoreV1Api(kube_client)
pod = core_api.read_namespaced_pod(node_id, self.namespace)
return pod.metadata.labels
def external_ip(self, node_id):
raise NotImplementedError("Must use internal IPs with Kubernetes.")
def internal_ip(self, node_id):
kube_client = self.__cached[self.cluster_name]["kube_client"]
core_api = client.CoreV1Api(kube_client)
pod = core_api.read_namespaced_pod(node_id, self.namespace)
return pod.status.pod_ip
def get_node_id(self, ip_address, use_internal_ip=True) -> str:
if not use_internal_ip:
raise ValueError("Must use internal IPs with Staroid.")
return super().get_node_id(ip_address, use_internal_ip=use_internal_ip)
def set_node_tags(self, node_id, tags):
kube_client = self.__cached[self.cluster_name]["kube_client"]
core_api = client.CoreV1Api(kube_client)
max_retry = 10
for i in range(max_retry):
try:
pod = core_api.read_namespaced_pod(node_id, self.namespace)
pod.metadata.labels.update(tags)
core_api.patch_namespaced_pod(node_id, self.namespace, pod)
except ApiException as e:
if e.status == 409 and max_retry - 1 > i:
# conflict. pod modified before apply patch. retry
time.sleep(0.2)
continue
raise e
def create_node(self, node_config, tags, count):
instance_name = self.cluster_name
incluster = self._connect_kubeapi(instance_name)
if incluster is None:
# get or create ske
cluster_api = self.__star.cluster()
ske = cluster_api.create(self.__ske, self.__ske_region)
if ske is None:
raise Exception(
"Failed to create an SKE '{}' in '{}' region".format(
self.__ske, self.__ske_region
)
)
# create a namespace
ns_api = self.__star.namespace(ske)
ns = ns_api.create(
instance_name,
self.provider_config["project"],
# Configure 'start-head' param to 'false'.
# head node will be created using Kubernetes api.
params=[{"group": "Misc", "name": "start-head", "value": "false"}],
)
if ns is None:
raise Exception(
"Failed to create a cluster '{}' in SKE '{}'".format(
instance_name, self.__ske
)
)
# 'ray down' will change staroid namespace status to "PAUSE"
# in this case we need to start namespace again.
if ns.status() == "PAUSE":
ns = ns_api.start(instance_name)
# kube client
kube_client = self._connect_kubeapi(instance_name)
core_api = client.CoreV1Api(kube_client)
apps_api = client.AppsV1Api(kube_client)
# retrieve container image
image = None
if self.provider_config["image_from_project"]:
ray_images = apps_api.read_namespaced_deployment(
name="ray-images", namespace=self.namespace
)
py_ver = self.provider_config["python_version"].replace(".", "-")
containers = ray_images.spec.template.spec.containers
for c in containers:
if py_ver in c.image:
image = c.image
break
logger.info(log_prefix + "use image {}".format(image))
# create head node
conf = node_config.copy()
pod_spec = conf.get("pod", conf)
service_spec = conf.get("service")
node_uuid = str(uuid4())
tags[TAG_RAY_CLUSTER_NAME] = self.cluster_name
tags["ray-node-uuid"] = node_uuid
pod_spec["metadata"]["namespace"] = self.namespace
if "labels" in pod_spec["metadata"]:
pod_spec["metadata"]["labels"].update(tags)
else:
pod_spec["metadata"]["labels"] = tags
if "generateName" not in pod_spec["metadata"]:
pod_spec["metadata"]["generateName"] = (
"ray-" + pod_spec["metadata"]["labels"]["ray-node-type"] + "-"
)
if "component" not in pod_spec["metadata"]["labels"]:
pod_spec["metadata"]["labels"]["component"] = (
"ray-" + pod_spec["metadata"]["labels"]["ray-node-type"]
)
if image is not None:
containers = pod_spec["spec"]["containers"]
for c in containers:
if c["name"] == "ray-node":
c["image"] = image
node_type = pod_spec["metadata"]["labels"]["ray-node-type"]
if node_type == "head":
if "STAROID_ACCESS_TOKEN" in os.environ:
c["env"].append(
{
"name": "STAROID_ACCESS_TOKEN",
"value": os.environ["STAROID_ACCESS_TOKEN"],
}
)
if "STAROID_ACCOUNT" in os.environ:
c["env"].append(
{
"name": "STAROID_ACCOUNT",
"value": os.environ["STAROID_ACCOUNT"],
}
)
if "STAROID_SKE" in os.environ:
c["env"].append(
{
"name": "STAROID_SKE",
"value": os.environ["STAROID_SKE"],
}
)
logger.info(
log_prefix + "calling create_namespaced_pod (count={}).".format(count)
)
new_nodes = []
for _ in range(count):
pod = core_api.create_namespaced_pod(self.namespace, pod_spec)
new_nodes.append(pod)
new_svcs = []
if service_spec is not None:
logger.info(
log_prefix + "calling create_namespaced_service "
"(count={}).".format(count)
)
for new_node in new_nodes:
metadata = service_spec.get("metadata", {})
metadata["name"] = new_node.metadata.name
service_spec["metadata"] = metadata
service_spec["spec"]["selector"] = {"ray-node-uuid": node_uuid}
svc = core_api.create_namespaced_service(self.namespace, service_spec)
new_svcs.append(svc)
def terminate_node(self, node_id):
logger.info(log_prefix + "calling delete_namespaced_pod")
kube_client = self.__cached[self.cluster_name]["kube_client"]
core_api = client.CoreV1Api(kube_client)
core_api.delete_namespaced_pod(node_id, self.namespace)
try:
core_api.delete_namespaced_service(node_id, self.namespace)
except ApiException:
pass
if node_id.startswith("ray-head"):
# Stop namespace on staroid after remove ray-head node.
instance_name = self.cluster_name
cluster_api = self.__star.cluster()
ske = cluster_api.get(self.__ske)
ns_api = self.__star.namespace(ske)
ns_api.get(instance_name)
del self.__cached[instance_name]
ns_api.stop_tunnel(instance_name)
ns_api.stop(instance_name)
def terminate_nodes(self, node_ids):
for node_id in node_ids:
self.terminate_node(node_id)
def get_command_runner(
self,
log_prefix,
node_id,
auth_config,
cluster_name,
process_runner,
use_internal_ip,
docker_config=None,
):
instance_name = self.cluster_name
# initialize connection
self._connect_kubeapi(instance_name)
command_runner = StaroidCommandRunner(
log_prefix,
self.namespace,
node_id,
auth_config,
process_runner,
self.__cached[cluster_name]["api_server"],
)
return command_runner
@staticmethod
def bootstrap_config(cluster_config):
"""Bootstraps the cluster config by adding env defaults if needed."""
return cluster_config
def test_initialize(self):
s = Staroid()
def test_initialize(self):
s = Staroid(access_token=os.environ["STAROID_ACCESS_TOKEN"],
account=os.environ["STAROID_ACCOUNT"])
s.get_all_accounts()
class Ods:
def __init__(self, staroid=None, ske=None, cache_dir=None):
self.__ske = None
if staroid == None:
self._staroid = Staroid()
else:
self._staroid = staroid
if cache_dir == None:
self.__cache_dir = "{}/.ods".format(str(Path.home()))
else:
self.__cache_dir = cache_dir
# configure from env var
if "STAROID_SKE" in os.environ:
self.__ske = os.environ["STAROID_SKE"]
# configure from args
if ske != None:
self.__ske = ske
def create_or_get_cache_dir(self, module=""):
"create (if not exists) or return cache dir path for module"
cache_dir = "{}/{}".format(self.__cache_dir, module)
if not os.path.exists(cache_dir):
os.makedirs(cache_dir)
return cache_dir
def download_chisel_if_not_exists(self):
self._staroid.get_chisel_path()
def _start_instance_on_staroid(self, instance_name, commit_url):
cluster = self._staroid.cluster().get(self.__ske)
if cluster == None:
raise Exception("Can't get ske cluster")
ns_api = self._staroid.namespace(cluster)
ns = ns_api.create(instance_name, commit_url)
if ns == None:
raise Exception("Can't create instance")
# if instnace is stopped, restart
if ns.status() == "PAUSE":
ns_api.start(instance_name)
# wait for phase to become RUNNING
return self.__wait_for_ns_phase(ns_api, ns, "RUNNING", 600)
def _start_tunnel(self, instance_name, tunnels):
cluster = self._staroid.cluster().get(self.__ske)
if cluster == None:
raise Exception("Can't get ske cluster")
ns_api = self._staroid.namespace(cluster)
ns = ns_api.get(instance_name)
ns_api.shell_start(instance_name)
ns_api.start_tunnel(instance_name, tunnels)
def _stop_tunnel(self, instance_name):
cluster = self._staroid.cluster().get(self.__ske)
if cluster == None:
raise Exception("Can't get ske cluster")
ns_api = self._staroid.namespace(cluster)
ns_api.stop_tunnel(instance_name)
ns_api.shell_stop(instance_name)
def _stop_instance_on_staroid(self, instance_name):
cluster = self._staroid.cluster().get(self.__ske)
if cluster == None:
raise Exception("Can't get ske cluster")
ns_api = self._staroid.namespace(cluster)
ns = ns_api.stop(instance_name)
ns = self.__wait_for_ns_phase(ns_api, ns, "PAUSED", 600)
return ns
def _delete_instance_on_staroid(self, instance_name):
cluster = self._staroid.cluster().get(self.__ske)
if cluster == None:
raise Exception("Can't get ske cluster")
ns_api = self._staroid.namespace(cluster)
ns = ns_api.delete(instance_name)
ns = self.__wait_for_ns_phase(ns_api, ns, "REMOVED", 600)
def __wait_for_ns_phase(self, ns_api, ns, phase, timeout):
start_time = time.time()
sleep_time = 1
max_sleep_time = 7
while ns.phase() != phase:
if time.time() - start_time > timeout:
raise Exception("Timeout")
# sleep
time.sleep(sleep_time)
if sleep_time < max_sleep_time:
sleep_time += 1
# check
ns = ns_api.get_by_id(ns.id())
return ns