def _renew_message_visibility(self, receipt_handle: str):
interval = self.renewal_period
new_timeout = self.visibility_timeout
cur_time = time.time()
while True:
time.sleep((cur_time + interval) - time.time())
cur_time += interval
new_timeout += interval
with self.receipt_handle_mutex:
if receipt_handle in self.stop_renewal:
self.stop_renewal.remove(receipt_handle)
break
try:
self.sqs_client.change_message_visibility(
QueueUrl=self.queue_url,
ReceiptHandle=receipt_handle,
VisibilityTimeout=new_timeout,
)
except botocore.exceptions.ClientError as err:
if err.response["Error"][
"Code"] == "InvalidParameterValue":
# unexpected; this error is thrown when attempting to renew a message that has been deleted
continue
elif err.response["Error"][
"Code"] == "AWS.SimpleQueueService.NonExistentQueue":
# there may be a delay between the cron may deleting the queue and this worker stopping
log.info(
"failed to renew message visibility because the queue was not found"
)
else:
self.stop_renewal.remove(receipt_handle)
raise err
def garbage_collect(
self,
exclude_disk_model_ids: List[str] = [],
dry_run: bool = False) -> Tuple[bool, List[str], List[str]]:
"""
Removes stale in-memory and on-disk models based on LRU policy.
Also calls the "remove" callback before removing the models from this object. The callback must not raise any exceptions.
Must be called with a write lock unless dry_run is set to true.
Args:
exclude_disk_model_ids: Model IDs to exclude from removing from disk. Necessary for locally-provided models.
dry_run: Just test if there are any models to remove. If set to true, this method can then be called with a read lock.
Returns:
A 3-element tuple. First element tells whether models had to be collected. The 2nd and 3rd elements contain the model IDs that were removed from memory and disk respectively.
"""
collected = False
if self._mem_cache_size <= 0 or self._disk_cache_size <= 0:
return collected
stale_mem_model_ids = self._lru_model_ids(self._mem_cache_size,
filter_in_mem=True)
stale_disk_model_ids = self._lru_model_ids(self._disk_cache_size -
len(exclude_disk_model_ids),
filter_in_mem=False)
if self._remove_callback and not dry_run:
self._remove_callback(stale_mem_model_ids)
# don't delete excluded model IDs from disk
stale_disk_model_ids = list(
set(stale_disk_model_ids) - set(exclude_disk_model_ids))
stale_disk_model_ids = stale_disk_model_ids[len(stale_disk_model_ids) -
self._disk_cache_size:]
if not dry_run:
logger.info(
f"unloading models {stale_mem_model_ids} from memory using the garbage collector"
)
logger.info(
f"unloading models {stale_disk_model_ids} from disk using the garbage collector"
)
for model_id in stale_mem_model_ids:
self.remove_model_by_id(model_id, mem=True, disk=False)
for model_id in stale_disk_model_ids:
self.remove_model_by_id(model_id, mem=False, disk=True)
if len(stale_mem_model_ids) > 0 or len(stale_disk_model_ids) > 0:
collected = True
return collected, stale_mem_model_ids, stale_disk_model_ids
def start(
self,
message_fn: Callable[[Dict[str, Any]], None],
message_failure_fn: Callable[[Dict[str, Any]], None],
on_job_complete_fn: Optional[Callable[[Dict[str, Any]], None]] = None,
):
no_messages_found_in_previous_iteration = False
signal_handler = SignalHandler()
while not signal_handler.received_signal():
response = self.sqs_client.receive_message(
QueueUrl=self.queue_url,
MaxNumberOfMessages=1,
WaitTimeSeconds=self.message_wait_time,
VisibilityTimeout=self.visibility_timeout,
MessageAttributeNames=["All"],
)
if response.get("Messages") is None or len(
response["Messages"]) == 0:
visible_messages, invisible_messages = self._get_total_messages_in_queue(
)
if visible_messages + invisible_messages == 0:
if no_messages_found_in_previous_iteration and self.stop_if_no_messages:
log.info("no messages left in queue, exiting...")
return
no_messages_found_in_previous_iteration = True
time.sleep(self.not_found_sleep_time)
continue
no_messages_found_in_previous_iteration = False
message = response["Messages"][0]
receipt_handle = message["ReceiptHandle"]
renewer = threading.Thread(
target=self._renew_message_visibility,
args=(receipt_handle, ),
daemon=True,
)
renewer.start()
if is_on_job_complete(message):
self._handle_on_job_complete(message, on_job_complete_fn)
else:
self._handle_message(message, message_fn, message_failure_fn)
def __init__(self, config):
num_success = 0
num_fail = 0
for i in range(config["num_requests"]):
if i > 0:
time.sleep(config["sleep"])
try:
# response = requests.get(config["endpoint"])
response = requests.post(config["endpoint"],
json=config["data"])
except Exception as e:
num_fail += 1
cortex_logger.error(
e,
extra={
"error": True,
"request_number": i,
},
)
continue
if response.status_code == 200:
num_success += 1
cortex_logger.info("successful request",
extra={
"request_success": True,
"request_number": i
})
else:
num_fail += 1
cortex_logger.error(
response.text,
extra={
"error": True,
"code": response.status_code,
"request_number": i,
},
)
cortex_logger.warn(
"FINISHED",
extra={
"finished": True,
"num_success": num_success,
"num_fail": num_fail
},
)
def _remove_models(self, model_ids: List[str]) -> None:
"""
Remove models from TFS.
Must only be used when caching enabled.
"""
logger.info(f"unloading models with model IDs {model_ids} from TFS")
models = {}
for model_id in model_ids:
model_name, model_version = model_id.rsplit("-", maxsplit=1)
if model_name not in models:
models[model_name] = [model_version]
else:
models[model_name].append(model_version)
model_names = []
model_versions = []
for model_name, versions in models.items():
model_names.append(model_name)
model_versions.append(versions)
self._client.remove_models(model_names, model_versions)
def _run_inference(self, model_input: Any, model_name: str,
model_version: str) -> dict:
"""
When processes_per_replica = 1 and caching enabled, check/load model and make prediction.
When processes_per_replica > 0 and caching disabled, attempt to make prediction regardless.
Args:
model_input: Input to the model.
model_name: Name of the model, as it's specified in predictor:models:paths or in the other case as they are named on disk.
model_version: Version of the model, as it's found on disk. Can also infer the version number from the "latest" version tag.
Returns:
The prediction.
"""
model = None
tag = ""
if model_version == "latest":
tag = model_version
if not self._caching_enabled:
# determine model version
if tag == "latest":
versions = self._client.poll_available_model_versions(
model_name)
if len(versions) == 0:
raise UserException(
f"model '{model_name}' accessed with tag {tag} couldn't be found"
)
model_version = str(max(map(lambda x: int(x), versions)))
model_id = model_name + "-" + model_version
return self._client.predict(model_input, model_name, model_version)
if not self._multiple_processes and self._caching_enabled:
# determine model version
try:
if tag == "latest":
model_version = self._get_latest_model_version_from_tree(
model_name, self._models_tree.model_info(model_name))
except ValueError:
# if model_name hasn't been found
raise UserRuntimeException(
f"'{model_name}' model of tag {tag} wasn't found in the list of available models"
)
models_stats = []
for model_id in self._models.get_model_ids():
models_stats = self._models.has_model_id(model_id)
# grab shared access to model tree
available_model = True
logger.info(
f"grabbing access to model {model_name} of version {model_version}"
)
with LockedModelsTree(self._models_tree, "r", model_name,
model_version):
# check if the versioned model exists
model_id = model_name + "-" + model_version
if model_id not in self._models_tree:
available_model = False
logger.info(
f"model {model_name} of version {model_version} is not available"
)
raise WithBreak
# retrieve model tree's metadata
upstream_model = self._models_tree[model_id]
current_upstream_ts = int(
upstream_model["timestamp"].timestamp())
logger.info(
f"model {model_name} of version {model_version} is available"
)
if not available_model:
if tag == "":
raise UserException(
f"model '{model_name}' of version '{model_version}' couldn't be found"
)
raise UserException(
f"model '{model_name}' accessed with tag '{tag}' couldn't be found"
)
# grab shared access to models holder and retrieve model
update_model = False
prediction = None
tfs_was_unresponsive = False
with LockedModel(self._models, "r", model_name, model_version):
logger.info(
f"checking the {model_name} {model_version} status")
status, local_ts = self._models.has_model(
model_name, model_version)
if status in ["not-available", "on-disk"
] or (status != "not-available"
and local_ts != current_upstream_ts):
logger.info(
f"model {model_name} of version {model_version} is not loaded (with status {status} or different timestamp)"
)
update_model = True
raise WithBreak
# run prediction
logger.info(
f"run the prediction on model {model_name} of version {model_version}"
)
self._models.get_model(model_name, model_version, tag)
try:
prediction = self._client.predict(model_input, model_name,
model_version)
except grpc.RpcError as e:
# effectively when it got restarted
if len(
self._client.poll_available_model_versions(
model_name)) > 0:
raise
tfs_was_unresponsive = True
# remove model from disk and memory references if TFS gets unresponsive
if tfs_was_unresponsive:
with LockedModel(self._models, "w", model_name, model_version):
available_versions = self._client.poll_available_model_versions(
model_name)
status, _ = self._models.has_model(model_name,
model_version)
if not (status == "in-memory"
and model_version not in available_versions):
raise WithBreak
logger.info(
f"removing model {model_name} of version {model_version} because TFS got unresponsive"
)
self._models.remove_model(model_name, model_version)
# download, load into memory the model and retrieve it
if update_model:
# grab exclusive access to models holder
with LockedModel(self._models, "w", model_name, model_version):
# check model status
status, local_ts = self._models.has_model(
model_name, model_version)
# refresh disk model
if status == "not-available" or (
status in ["on-disk", "in-memory"]
and local_ts != current_upstream_ts):
# unload model from TFS
if status == "in-memory":
try:
logger.info(
f"unloading model {model_name} of version {model_version} from TFS"
)
self._models.unload_model(
model_name, model_version)
except Exception:
logger.info(
f"failed unloading model {model_name} of version {model_version} from TFS"
)
raise
# remove model from disk and references
if status in ["on-disk", "in-memory"]:
logger.info(
f"removing model references from memory and from disk for model {model_name} of version {model_version}"
)
self._models.remove_model(model_name,
model_version)
# download model
if model_name not in self._spec_models.get_local_model_names(
):
logger.info(
f"downloading model {model_name} of version {model_version} from the {upstream_model['provider']} upstream"
)
date = self._models.download_model(
upstream_model["provider"],
upstream_model["bucket"],
model_name,
model_version,
upstream_model["path"],
)
if not date:
raise WithBreak
current_upstream_ts = int(date.timestamp())
# load model
try:
logger.info(
f"loading model {model_name} of version {model_version} into memory"
)
self._models.load_model(
model_name,
model_version,
current_upstream_ts,
[tag],
kwargs={
"model_name":
model_name,
"model_version":
model_version,
"signature_key":
self._determine_model_signature_key(
model_name),
},
)
except Exception as e:
raise UserRuntimeException(
f"failed (re-)loading model {model_name} of version {model_version} (thread {td.get_ident()})",
str(e),
)
# run prediction
self._models.get_model(model_name, model_version, tag)
prediction = self._client.predict(model_input, model_name,
model_version)
return prediction
def _get_model(self, model_name: str, model_version: str) -> Any:
"""
Checks if versioned model is on disk, then checks if model is in memory,
and if not, it loads it into memory, and returns the model.
Args:
model_name: Name of the model, as it's specified in predictor:models:paths or in the other case as they are named on disk.
model_version: Version of the model, as it's found on disk. Can also infer the version number from the "latest" tag.
Exceptions:
RuntimeError: if another thread tried to load the model at the very same time.
Returns:
The model as returned by self._load_model method.
None if the model wasn't found or if it didn't pass the validation.
"""
model = None
tag = ""
if model_version == "latest":
tag = model_version
if not self._caching_enabled:
# determine model version
if tag == "latest":
model_version = self._get_latest_model_version_from_disk(
model_name)
model_id = model_name + "-" + model_version
# grab shared access to versioned model
resource = os.path.join(self._lock_dir, model_id + ".txt")
with LockedFile(resource, "r", reader_lock=True) as f:
# check model status
file_status = f.read()
if file_status == "" or file_status == "not-available":
raise WithBreak
current_upstream_ts = int(file_status.split(" ")[1])
update_model = False
# grab shared access to models holder and retrieve model
with LockedModel(self._models, "r", model_name, model_version):
status, local_ts = self._models.has_model(
model_name, model_version)
if status == "not-available" or (
status == "in-memory"
and local_ts != current_upstream_ts):
update_model = True
raise WithBreak
model, _ = self._models.get_model(model_name,
model_version, tag)
# load model into memory and retrieve it
if update_model:
with LockedModel(self._models, "w", model_name,
model_version):
status, _ = self._models.has_model(
model_name, model_version)
if status == "not-available" or (
status == "in-memory"
and local_ts != current_upstream_ts):
if status == "not-available":
logger.info(
f"loading model {model_name} of version {model_version} (thread {td.get_ident()})"
)
else:
logger.info(
f"reloading model {model_name} of version {model_version} (thread {td.get_ident()})"
)
try:
self._models.load_model(
model_name,
model_version,
current_upstream_ts,
[tag],
)
except Exception as e:
raise UserRuntimeException(
f"failed (re-)loading model {model_name} of version {model_version} (thread {td.get_ident()})",
str(e),
)
model, _ = self._models.get_model(
model_name, model_version, tag)
if not self._multiple_processes and self._caching_enabled:
# determine model version
try:
if tag == "latest":
model_version = self._get_latest_model_version_from_tree(
model_name, self._models_tree.model_info(model_name))
except ValueError:
# if model_name hasn't been found
raise UserRuntimeException(
f"'{model_name}' model of tag latest wasn't found in the list of available models"
)
# grab shared access to model tree
available_model = True
with LockedModelsTree(self._models_tree, "r", model_name,
model_version):
# check if the versioned model exists
model_id = model_name + "-" + model_version
if model_id not in self._models_tree:
available_model = False
raise WithBreak
# retrieve model tree's metadata
upstream_model = self._models_tree[model_id]
current_upstream_ts = int(
upstream_model["timestamp"].timestamp())
if not available_model:
return None
# grab shared access to models holder and retrieve model
update_model = False
with LockedModel(self._models, "r", model_name, model_version):
status, local_ts = self._models.has_model(
model_name, model_version)
if status in ["not-available", "on-disk"
] or (status != "not-available"
and local_ts != current_upstream_ts):
update_model = True
raise WithBreak
model, _ = self._models.get_model(model_name, model_version,
tag)
# download, load into memory the model and retrieve it
if update_model:
# grab exclusive access to models holder
with LockedModel(self._models, "w", model_name, model_version):
# check model status
status, local_ts = self._models.has_model(
model_name, model_version)
# refresh disk model
if status == "not-available" or (
status in ["on-disk", "in-memory"]
and local_ts != current_upstream_ts):
if status == "not-available":
logger.info(
f"model {model_name} of version {model_version} not found locally; continuing with the download..."
)
elif status == "on-disk":
logger.info(
f"found newer model {model_name} of vesion {model_version} on the {upstream_model['provider']} upstream than the one on the disk"
)
else:
logger.info(
f"found newer model {model_name} of vesion {model_version} on the {upstream_model['provider']} upstream than the one loaded into memory"
)
# remove model from disk and memory
if status == "on-disk":
logger.info(
f"removing model from disk for model {model_name} of version {model_version}"
)
self._models.remove_model(model_name,
model_version)
if status == "in-memory":
logger.info(
f"removing model from disk and memory for model {model_name} of version {model_version}"
)
self._models.remove_model(model_name,
model_version)
# download model
logger.info(
f"downloading model {model_name} of version {model_version} from the {upstream_model['provider']} upstream"
)
date = self._models.download_model(
upstream_model["provider"],
upstream_model["bucket"],
model_name,
model_version,
upstream_model["path"],
)
if not date:
raise WithBreak
current_upstream_ts = int(date.timestamp())
# load model
try:
logger.info(
f"loading model {model_name} of version {model_version} into memory"
)
self._models.load_model(
model_name,
model_version,
current_upstream_ts,
[tag],
)
except Exception as e:
raise UserRuntimeException(
f"failed (re-)loading model {model_name} of version {model_version} (thread {td.get_ident()})",
str(e),
)
# retrieve model
model, _ = self._models.get_model(model_name,
model_version, tag)
return model
def _signal_handler(self, sys_signal, _):
log.info(f"handling signal {sys_signal}, exiting gracefully")
self.__received_signal = True
def _extract_signatures(
self, signature_def, signature_key, model_name: str, model_version: str
):
logger.info(
"signature defs found in model '{}' for version '{}': {}".format(
model_name, model_version, signature_def
)
)
available_keys = list(signature_def.keys())
if len(available_keys) == 0:
raise UserException(
"unable to find signature defs in model '{}' of version '{}'".format(
model_name, model_version
)
)
if signature_key is None:
if len(available_keys) == 1:
logger.info(
"signature_key was not configured by user, using signature key '{}' for model '{}' of version '{}' (found in the signature def map)".format(
available_keys[0],
model_name,
model_version,
)
)
signature_key = available_keys[0]
elif "predict" in signature_def:
logger.info(
"signature_key was not configured by user, using signature key 'predict' for model '{}' of version '{}' (found in the signature def map)".format(
model_name,
model_version,
)
)
signature_key = "predict"
else:
raise UserException(
"signature_key was not configured by user, please specify one the following keys '{}' for model '{}' of version '{}' (found in the signature def map)".format(
", ".join(available_keys), model_name, model_version
)
)
else:
if signature_def.get(signature_key) is None:
possibilities_str = "key: '{}'".format(available_keys[0])
if len(available_keys) > 1:
possibilities_str = "keys: '{}'".format("', '".join(available_keys))
raise UserException(
"signature_key '{}' was not found in signature def map for model '{}' of version '{}', but found the following {}".format(
signature_key, model_name, model_version, possibilities_str
)
)
signature_def_val = signature_def.get(signature_key)
if signature_def_val.get("inputs") is None:
raise UserException(
"unable to find 'inputs' in signature def '{}' for model '{}'".format(
signature_key, model_name
)
)
parsed_signatures = {}
for input_name, input_metadata in signature_def_val["inputs"].items():
if input_metadata["tensorShape"] == {}:
# a scalar with rank 0 and empty shape
shape = "scalar"
elif input_metadata["tensorShape"].get("unknownRank", False):
# unknown rank and shape
#
# unknownRank is set to True if the model input has no rank
# it may lead to an undefined behavior if unknownRank is only checked for its presence
# so it also gets to be tested against its value
shape = "unknown"
elif input_metadata["tensorShape"].get("dim", None):
# known rank and known/unknown shape
shape = [int(dim["size"]) for dim in input_metadata["tensorShape"]["dim"]]
else:
raise UserException(
"invalid 'tensorShape' specification for input '{}' in signature key '{}' for model '{}'",
input_name,
signature_key,
model_name,
)
parsed_signatures[input_name] = {
"shape": shape if type(shape) == list else [shape],
"type": DTYPE_TO_TF_TYPE[input_metadata["dtype"]].name,
}
return signature_key, parsed_signatures
def model_downloader(
predictor_type: PredictorType,
bucket_provider: str,
bucket_name: str,
model_name: str,
model_version: str,
model_path: str,
temp_dir: str,
model_dir: str,
) -> Optional[datetime.datetime]:
"""
Downloads model to disk. Validates the cloud model path and the downloaded model as well.
Args:
predictor_type: The predictor type as implemented by the API.
bucket_provider: Provider for the bucket. Can be "s3" or "gs".
bucket_name: Name of the bucket where the model is stored.
model_name: Name of the model. Is part of the model's local path.
model_version: Version of the model. Is part of the model's local path.
model_path: Model prefix of the versioned model.
temp_dir: Where to temporarily store the model for validation.
model_dir: The top directory of where all models are stored locally.
Returns:
The model's timestamp. None if the model didn't pass the validation, if it doesn't exist or if there are not enough permissions.
"""
logger.info(
f"downloading from bucket {bucket_name}/{model_path}, model {model_name} of version {model_version}, temporarily to {temp_dir} and then finally to {model_dir}"
)
if bucket_provider == "s3":
client = S3(bucket_name)
if bucket_provider == "gs":
client = GCS(bucket_name)
# validate upstream cloud model
sub_paths, ts = client.search(model_path)
try:
validate_model_paths(sub_paths, predictor_type, model_path)
except CortexException:
logger.info(
f"failed validating model {model_name} of version {model_version}")
return None
# download model to temp dir
temp_dest = os.path.join(temp_dir, model_name, model_version)
try:
client.download_dir_contents(model_path, temp_dest)
except CortexException:
logger.info(
f"failed downloading model {model_name} of version {model_version} to temp dir {temp_dest}"
)
shutil.rmtree(temp_dest)
return None
# validate model
model_contents = glob.glob(os.path.join(temp_dest, "**"), recursive=True)
model_contents = util.remove_non_empty_directory_paths(model_contents)
try:
validate_model_paths(model_contents, predictor_type, temp_dest)
except CortexException:
logger.info(
f"failed validating model {model_name} of version {model_version} from temp dir"
)
shutil.rmtree(temp_dest)
return None
# move model to dest dir
model_top_dir = os.path.join(model_dir, model_name)
ondisk_model_version = os.path.join(model_top_dir, model_version)
logger.info(
f"moving model {model_name} of version {model_version} to final dir {ondisk_model_version}"
)
if os.path.isdir(ondisk_model_version):
shutil.rmtree(ondisk_model_version)
shutil.move(temp_dest, ondisk_model_version)
return max(ts)
def predict(self, payload):
cortex_logger.info("received payload", extra={"payload": payload})
return payload