def start_fn():
provider = os.environ["CORTEX_PROVIDER"]
project_dir = os.environ["CORTEX_PROJECT_DIR"]
spec_path = os.environ["CORTEX_API_SPEC"]
model_dir = os.getenv("CORTEX_MODEL_DIR")
cache_dir = os.getenv("CORTEX_CACHE_DIR")
region = os.getenv("AWS_REGION")
tf_serving_port = os.getenv("CORTEX_TF_BASE_SERVING_PORT", "9000")
tf_serving_host = os.getenv("CORTEX_TF_SERVING_HOST", "localhost")
has_multiple_servers = os.getenv("CORTEX_MULTIPLE_TF_SERVERS")
if has_multiple_servers:
with LockedFile("/run/used_ports.json", "r+") as f:
used_ports = json.load(f)
for port in used_ports.keys():
if not used_ports[port]:
tf_serving_port = port
used_ports[port] = True
break
f.seek(0)
json.dump(used_ports, f)
f.truncate()
try:
api = get_api(provider, spec_path, model_dir, cache_dir, region)
client = api.predictor.initialize_client(
tf_serving_host=tf_serving_host, tf_serving_port=tf_serving_port
)
with FileLock("/run/init_stagger.lock"):
logger.info("loading the predictor from {}".format(api.predictor.path))
predictor_impl = api.predictor.initialize_impl(project_dir, client)
local_cache["api"] = api
local_cache["provider"] = provider
local_cache["client"] = client
local_cache["predictor_impl"] = predictor_impl
local_cache["predict_fn_args"] = inspect.getfullargspec(predictor_impl.predict).args
if api.server_side_batching_enabled:
dynamic_batching_config = api.api_spec["predictor"]["server_side_batching"]
local_cache["dynamic_batcher"] = DynamicBatcher(
predictor_impl,
max_batch_size=dynamic_batching_config["max_batch_size"],
batch_interval=dynamic_batching_config["batch_interval"]
/ NANOSECONDS_IN_SECOND, # convert nanoseconds to seconds
)
if util.has_method(predictor_impl, "post_predict"):
local_cache["post_predict_fn_args"] = inspect.getfullargspec(
predictor_impl.post_predict
).args
predict_route = "/predict"
local_cache["predict_route"] = predict_route
except:
logger.exception("failed to start api")
sys.exit(1)
app.add_api_route(local_cache["predict_route"], predict, methods=["POST"])
app.add_api_route(local_cache["predict_route"], get_summary, methods=["GET"])
return app
def start():
while not pathlib.Path("/mnt/workspace/init_script_run.txt").is_file():
time.sleep(0.2)
cache_dir = os.environ["CORTEX_CACHE_DIR"]
api_spec_path = os.environ["CORTEX_API_SPEC"]
job_spec_path = os.environ["CORTEX_JOB_SPEC"]
project_dir = os.environ["CORTEX_PROJECT_DIR"]
model_dir = os.getenv("CORTEX_MODEL_DIR")
tf_serving_port = os.getenv("CORTEX_TF_BASE_SERVING_PORT", "9000")
tf_serving_host = os.getenv("CORTEX_TF_SERVING_HOST", "localhost")
region = os.getenv("AWS_REGION")
has_multiple_servers = os.getenv("CORTEX_MULTIPLE_TF_SERVERS")
if has_multiple_servers:
with LockedFile("/run/used_ports.json", "r+") as f:
used_ports = json.load(f)
for port in used_ports.keys():
if not used_ports[port]:
tf_serving_port = port
used_ports[port] = True
break
f.seek(0)
json.dump(used_ports, f)
f.truncate()
api = get_api(api_spec_path, model_dir, cache_dir, region)
storage, _ = get_spec(api_spec_path, cache_dir, region)
job_spec = get_job_spec(storage, cache_dir, job_spec_path)
sqs_client = boto3.client("sqs", region_name=region)
client = api.predictor.initialize_client(tf_serving_host=tf_serving_host,
tf_serving_port=tf_serving_port)
try:
log.info("loading the predictor from {}".format(api.predictor.path))
metrics_client = MetricsClient(api.statsd)
predictor_impl = api.predictor.initialize_impl(
project_dir=project_dir,
client=client,
metrics_client=metrics_client,
job_spec=job_spec,
)
except UserRuntimeException as err:
err.wrap(f"failed to start job {job_spec['job_id']}")
log.error(str(err), exc_info=True)
sys.exit(1)
except Exception as err:
capture_exception(err)
log.error(f"failed to start job {job_spec['job_id']}", exc_info=True)
sys.exit(1)
# crons only stop if an unhandled exception occurs
def check_if_crons_have_failed():
while True:
for cron in api.predictor.crons:
if not cron.is_alive():
os.kill(os.getpid(), signal.SIGQUIT)
time.sleep(1)
threading.Thread(target=check_if_crons_have_failed, daemon=True).start()
local_cache["api"] = api
local_cache["job_spec"] = job_spec
local_cache["predictor_impl"] = predictor_impl
local_cache["predict_fn_args"] = inspect.getfullargspec(
predictor_impl.predict).args
local_cache["sqs_client"] = sqs_client
open("/mnt/workspace/api_readiness.txt", "a").close()
log.info("polling for batches...")
try:
sqs_handler = SQSHandler(
sqs_client=sqs_client,
queue_url=job_spec["sqs_url"],
renewal_period=MESSAGE_RENEWAL_PERIOD,
visibility_timeout=INITIAL_MESSAGE_VISIBILITY,
not_found_sleep_time=MESSAGE_NOT_FOUND_SLEEP,
message_wait_time=SQS_POLL_WAIT_TIME,
dead_letter_queue_url=job_spec.get("sqs_dead_letter_queue"),
stop_if_no_messages=True,
)
sqs_handler.start(
message_fn=handle_batch_message,
message_failure_fn=handle_batch_failure,
on_job_complete_fn=handle_on_job_complete,
)
except UserRuntimeException as err:
err.wrap(f"failed to run job {job_spec['job_id']}")
log.error(str(err), exc_info=True)
sys.exit(1)
except Exception as err:
capture_exception(err)
log.error(f"failed to run job {job_spec['job_id']}", exc_info=True)
sys.exit(1)
def start():
while not pathlib.Path("/mnt/workspace/init_script_run.txt").is_file():
time.sleep(0.2)
cache_dir = os.environ["CORTEX_CACHE_DIR"]
provider = os.environ["CORTEX_PROVIDER"]
api_spec_path = os.environ["CORTEX_API_SPEC"]
job_spec_path = os.environ["CORTEX_JOB_SPEC"]
project_dir = os.environ["CORTEX_PROJECT_DIR"]
model_dir = os.getenv("CORTEX_MODEL_DIR")
tf_serving_port = os.getenv("CORTEX_TF_BASE_SERVING_PORT", "9000")
tf_serving_host = os.getenv("CORTEX_TF_SERVING_HOST", "localhost")
region = os.getenv("AWS_REGION")
has_multiple_servers = os.getenv("CORTEX_MULTIPLE_TF_SERVERS")
if has_multiple_servers:
with LockedFile("/run/used_ports.json", "r+") as f:
used_ports = json.load(f)
for port in used_ports.keys():
if not used_ports[port]:
tf_serving_port = port
used_ports[port] = True
break
f.seek(0)
json.dump(used_ports, f)
f.truncate()
api = get_api(provider, api_spec_path, model_dir, cache_dir, region)
storage, api_spec = get_spec(provider, api_spec_path, cache_dir, region)
job_spec = get_job_spec(storage, cache_dir, job_spec_path)
try:
client = api.predictor.initialize_client(
tf_serving_host=tf_serving_host, tf_serving_port=tf_serving_port)
logger.info("loading the predictor from {}".format(api.predictor.path))
predictor_impl = api.predictor.initialize_impl(project_dir, client,
job_spec)
except CortexException as err:
err.wrap(f"failed to start job {job_spec['job_id']}")
logger.error(str(err), exc_info=True)
sys.exit(1)
except:
logger.error(f"failed to start job {job_spec['job_id']}",
exc_info=True)
sys.exit(1)
local_cache["api_spec"] = api
local_cache["provider"] = provider
local_cache["job_spec"] = job_spec
local_cache["predictor_impl"] = predictor_impl
local_cache["predict_fn_args"] = inspect.getfullargspec(
predictor_impl.predict).args
local_cache["sqs_client"] = boto3.client("sqs", region_name=region)
open("/mnt/workspace/api_readiness.txt", "a").close()
logger.info("polling for batches...")
try:
sqs_loop()
except CortexException as err:
err.wrap(f"failed to run job {job_spec['job_id']}")
logger.error(str(err), exc_info=True)
sys.exit(1)
except:
logger.error(f"failed to run job {job_spec['job_id']}", exc_info=True)
sys.exit(1)
def start_fn():
provider = os.environ["CORTEX_PROVIDER"]
project_dir = os.environ["CORTEX_PROJECT_DIR"]
spec_path = os.environ["CORTEX_API_SPEC"]
model_dir = os.getenv("CORTEX_MODEL_DIR")
cache_dir = os.getenv("CORTEX_CACHE_DIR")
region = os.getenv("AWS_REGION")
tf_serving_port = os.getenv("CORTEX_TF_BASE_SERVING_PORT", "9000")
tf_serving_host = os.getenv("CORTEX_TF_SERVING_HOST", "localhost")
try:
has_multiple_servers = os.getenv("CORTEX_MULTIPLE_TF_SERVERS")
if has_multiple_servers:
with LockedFile("/run/used_ports.json", "r+") as f:
used_ports = json.load(f)
for port in used_ports.keys():
if not used_ports[port]:
tf_serving_port = port
used_ports[port] = True
break
f.seek(0)
json.dump(used_ports, f)
f.truncate()
api = get_api(provider, spec_path, model_dir, cache_dir, region)
client = api.predictor.initialize_client(
tf_serving_host=tf_serving_host, tf_serving_port=tf_serving_port)
with FileLock("/run/init_stagger.lock"):
logger.info("loading the predictor from {}".format(
api.predictor.path))
predictor_impl = api.predictor.initialize_impl(project_dir, client)
# crons only stop if an unhandled exception occurs
def check_if_crons_have_failed():
while True:
for cron in api.predictor.crons:
if not cron.is_alive():
os.kill(os.getpid(), signal.SIGQUIT)
time.sleep(1)
threading.Thread(target=check_if_crons_have_failed,
daemon=True).start()
local_cache["api"] = api
local_cache["provider"] = provider
local_cache["client"] = client
local_cache["predictor_impl"] = predictor_impl
local_cache["predict_fn_args"] = inspect.getfullargspec(
predictor_impl.predict).args
if api.python_server_side_batching_enabled:
dynamic_batching_config = api.api_spec["predictor"][
"server_side_batching"]
local_cache["dynamic_batcher"] = DynamicBatcher(
predictor_impl,
max_batch_size=dynamic_batching_config["max_batch_size"],
batch_interval=dynamic_batching_config["batch_interval"] /
NANOSECONDS_IN_SECOND, # convert nanoseconds to seconds
)
if util.has_method(predictor_impl, "post_predict"):
local_cache["post_predict_fn_args"] = inspect.getfullargspec(
predictor_impl.post_predict).args
predict_route = "/predict"
local_cache["predict_route"] = predict_route
except (UserRuntimeException, Exception) as err:
if not isinstance(err, UserRuntimeException):
capture_exception(err)
logger.exception("failed to start api")
sys.exit(1)
app.add_api_route(local_cache["predict_route"], predict, methods=["POST"])
app.add_api_route(local_cache["predict_route"],
get_summary,
methods=["GET"])
return app
def init():
project_dir = os.environ["CORTEX_PROJECT_DIR"]
spec_path = os.environ["CORTEX_API_SPEC"]
model_dir = os.getenv("CORTEX_MODEL_DIR")
cache_dir = os.getenv("CORTEX_CACHE_DIR")
region = os.getenv("AWS_REGION")
tf_serving_port = os.getenv("CORTEX_TF_BASE_SERVING_PORT", "9000")
tf_serving_host = os.getenv("CORTEX_TF_SERVING_HOST", "localhost")
has_multiple_servers = os.getenv("CORTEX_MULTIPLE_TF_SERVERS")
if has_multiple_servers:
with LockedFile("/run/used_ports.json", "r+") as f:
used_ports = json.load(f)
for port in used_ports.keys():
if not used_ports[port]:
tf_serving_port = port
used_ports[port] = True
break
f.seek(0)
json.dump(used_ports, f)
f.truncate()
api = get_api(spec_path, model_dir, cache_dir, region)
config: Dict[str, Any] = {
"api": None,
"client": None,
"predictor_impl": None,
"module_proto_pb2_grpc": None,
}
proto_without_ext = pathlib.Path(api.predictor.protobuf_path).stem
module_proto_pb2 = importlib.import_module(proto_without_ext + "_pb2")
module_proto_pb2_grpc = importlib.import_module(proto_without_ext +
"_pb2_grpc")
client = api.predictor.initialize_client(tf_serving_host=tf_serving_host,
tf_serving_port=tf_serving_port)
with FileLock("/run/init_stagger.lock"):
logger.info("loading the predictor from {}".format(api.predictor.path))
metrics_client = MetricsClient(api.statsd)
predictor_impl = api.predictor.initialize_impl(
project_dir=project_dir,
client=client,
metrics_client=metrics_client,
proto_module_pb2=module_proto_pb2,
)
# crons only stop if an unhandled exception occurs
def check_if_crons_have_failed():
while True:
for cron in api.predictor.crons:
if not cron.is_alive():
os.kill(os.getpid(), signal.SIGQUIT)
time.sleep(1)
threading.Thread(target=check_if_crons_have_failed, daemon=True).start()
ServicerClass = get_servicer_from_module(module_proto_pb2_grpc)
class PredictorServicer(ServicerClass):
def __init__(self, predict_fn_args, predictor_impl, api):
self.predict_fn_args = predict_fn_args
self.predictor_impl = predictor_impl
self.api = api
def Predict(self, payload, context):
try:
kwargs = build_predict_kwargs(self.predict_fn_args, payload,
context)
response = self.predictor_impl.predict(**kwargs)
self.api.post_status_code_request_metrics(200)
except Exception:
logger.error(traceback.format_exc())
self.api.post_status_code_request_metrics(500)
context.abort(grpc.StatusCode.INTERNAL,
"internal server error")
return response
config["api"] = api
config["client"] = client
config["predictor_impl"] = predictor_impl
config["predict_fn_args"] = inspect.getfullargspec(
predictor_impl.predict).args
config["module_proto_pb2"] = module_proto_pb2
config["module_proto_pb2_grpc"] = module_proto_pb2_grpc
config["predictor_servicer"] = PredictorServicer
return config
def start():
while not pathlib.Path("/mnt/workspace/init_script_run.txt").is_file():
time.sleep(0.2)
cache_dir = os.environ["CORTEX_CACHE_DIR"]
provider = os.environ["CORTEX_PROVIDER"]
api_spec_path = os.environ["CORTEX_API_SPEC"]
job_spec_path = os.environ["CORTEX_JOB_SPEC"]
project_dir = os.environ["CORTEX_PROJECT_DIR"]
model_dir = os.getenv("CORTEX_MODEL_DIR")
tf_serving_port = os.getenv("CORTEX_TF_BASE_SERVING_PORT", "9000")
tf_serving_host = os.getenv("CORTEX_TF_SERVING_HOST", "localhost")
region = os.getenv("AWS_REGION")
has_multiple_servers = os.getenv("CORTEX_MULTIPLE_TF_SERVERS")
if has_multiple_servers:
with LockedFile("/run/used_ports.json", "r+") as f:
used_ports = json.load(f)
for port in used_ports.keys():
if not used_ports[port]:
tf_serving_port = port
used_ports[port] = True
break
f.seek(0)
json.dump(used_ports, f)
f.truncate()
api = get_api(provider, api_spec_path, model_dir, cache_dir, region)
storage, _ = get_spec(provider, api_spec_path, cache_dir, region)
job_spec = get_job_spec(storage, cache_dir, job_spec_path)
client = api.predictor.initialize_client(tf_serving_host=tf_serving_host,
tf_serving_port=tf_serving_port)
try:
logger.info("loading the predictor from {}".format(api.predictor.path))
metrics_client = MetricsClient(api.statsd)
predictor_impl = api.predictor.initialize_impl(
project_dir=project_dir,
client=client,
metrics_client=metrics_client,
job_spec=job_spec,
)
except UserRuntimeException as err:
err.wrap(f"failed to start job {job_spec['job_id']}")
logger.error(str(err), exc_info=True)
sys.exit(1)
except Exception as err:
capture_exception(err)
logger.error(f"failed to start job {job_spec['job_id']}",
exc_info=True)
sys.exit(1)
# crons only stop if an unhandled exception occurs
def check_if_crons_have_failed():
while True:
for cron in api.predictor.crons:
if not cron.is_alive():
os.kill(os.getpid(), signal.SIGQUIT)
time.sleep(1)
threading.Thread(target=check_if_crons_have_failed, daemon=True).start()
local_cache["api"] = api
local_cache["provider"] = provider
local_cache["job_spec"] = job_spec
local_cache["predictor_impl"] = predictor_impl
local_cache["predict_fn_args"] = inspect.getfullargspec(
predictor_impl.predict).args
local_cache["sqs_client"] = boto3.client("sqs", region_name=region)
open("/mnt/workspace/api_readiness.txt", "a").close()
logger.info("polling for batches...")
try:
sqs_loop()
except UserRuntimeException as err:
err.wrap(f"failed to run job {job_spec['job_id']}")
logger.error(str(err), exc_info=True)
sys.exit(1)
except Exception as err:
capture_exception(err)
logger.error(f"failed to run job {job_spec['job_id']}", exc_info=True)
sys.exit(1)