Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

feat: support autoscaling metrics when deploying models #1197

Merged
merged 14 commits into from
May 23, 2022
+234 −9
Merged

feat: support autoscaling metrics when deploying models #1197

Changes from 2 commits
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
03d7885
feat: support autoscaling metrics when deploying models
munagekar May 5, 2022
4bb666b
feat: support model deploy to endpoint with autoscaling metrics
munagekar May 5, 2022
9501eb0
fix autoscaling_target_accelerator_duty_cycle check
munagekar May 11, 2022
7a601c7
fix docstring: specify that autoscaling_params are optional
munagekar May 11, 2022
b278298
bug fix: add autoscaling_target_cpu_utilization to custom_resource_spec
munagekar May 11, 2022
c82ba65
add tests
munagekar May 11, 2022
e14d1ae
add _TEST_METRIC_NAME_CPU_UTILIZATION and _TEST_METRIC_NAME_GPU_UTILI…
munagekar May 17, 2022
fbceb92
remove not required arguments in tests
munagekar May 17, 2022
0f2353e
Merge branch 'main' into feat/autoscaling-endpoint
munagekar May 17, 2022
8a28d7a
fix tests: wait for LRO to complete even if not sync
munagekar May 17, 2022
779b805
Merge branch 'main' into feat/autoscaling-endpoint
munagekar May 19, 2022
12ad857
fix lint: run black
munagekar May 20, 2022
4acd36d
Merge branch 'main' into feat/autoscaling-endpoint
sararob May 20, 2022
1f19010
Merge branch 'main' into feat/autoscaling-endpoint
sararob May 23, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
91 changes: 83 additions & 8 deletions google/cloud/aiplatform/models.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -644,6 +644,8 @@ def deploy(
metadata: Optional[Sequence[Tuple[str, str]]] = (),
sync=True,
deploy_request_timeout: Optional[float] = None,
autoscaling_target_cpu_utilization: Optional[int] = None,
autoscaling_target_accelerator_duty_cycle: Optional[int] = None,
) -> None:
"""Deploys a Model to the Endpoint.

Expand Down Expand Up @@ -717,6 +719,13 @@ def deploy(
be immediately returned and synced when the Future has completed.
deploy_request_timeout (float):
Optional. The timeout for the deploy request in seconds.
autoscaling_target_cpu_utilization (int):
Target CPU Utilization to use for Autoscaling Replicas.
A default value of 60 will be used if not specified.
autoscaling_target_accelerator_duty_cycle (int):
Target Accelerator Duty Cycle.
Must also set accelerator_type and accelerator_count if specified.
A default value of 60 will be used if not specified.
"""
self._sync_gca_resource_if_skipped()

Expand Down Expand Up @@ -747,6 +756,8 @@ def deploy(
metadata=metadata,
sync=sync,
deploy_request_timeout=deploy_request_timeout,
autoscaling_target_cpu_utilization=autoscaling_target_cpu_utilization,
autoscaling_target_accelerator_duty_cycle=autoscaling_target_accelerator_duty_cycle
)

@base.optional_sync()
Expand All @@ -767,6 +778,8 @@ def _deploy(
metadata: Optional[Sequence[Tuple[str, str]]] = (),
sync=True,
deploy_request_timeout: Optional[float] = None,
autoscaling_target_cpu_utilization: Optional[int] = None,
autoscaling_target_accelerator_duty_cycle: Optional[int] = None,
) -> None:
"""Deploys a Model to the Endpoint.

Expand Down Expand Up @@ -840,6 +853,13 @@ def _deploy(
be immediately returned and synced when the Future has completed.
deploy_request_timeout (float):
Optional. The timeout for the deploy request in seconds.
autoscaling_target_cpu_utilization (int):
Target CPU Utilization to use for Autoscaling Replicas.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you add "Optional" to the beginning of this docstring? Same for autoscaling_target_accelerator_duty_cycle.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed in 7a601c7

A default value of 60 will be used if not specified.
autoscaling_target_accelerator_duty_cycle (int):
Target Accelerator Duty Cycle.
Must also set accelerator_type and accelerator_count if specified.
A default value of 60 will be used if not specified.
Raises:
ValueError: If there is not current traffic split and traffic percentage
is not 0 or 100.
Expand All @@ -866,6 +886,8 @@ def _deploy(
explanation_parameters=explanation_parameters,
metadata=metadata,
deploy_request_timeout=deploy_request_timeout,
autoscaling_target_cpu_utilization=autoscaling_target_cpu_utilization,
autoscaling_target_accelerator_duty_cycle=autoscaling_target_accelerator_duty_cycle
)

_LOGGER.log_action_completed_against_resource("model", "deployed", self)
Expand All @@ -892,6 +914,8 @@ def _deploy_call(
explanation_parameters: Optional[explain.ExplanationParameters] = None,
metadata: Optional[Sequence[Tuple[str, str]]] = (),
deploy_request_timeout: Optional[float] = None,
autoscaling_target_cpu_utilization: Optional[int] = None,
autoscaling_target_accelerator_duty_cycle: Optional[int] = None,
):
"""Helper method to deploy model to endpoint.

Expand Down Expand Up @@ -965,6 +989,13 @@ def _deploy_call(
be immediately returned and synced when the Future has completed.
deploy_request_timeout (float):
Optional. The timeout for the deploy request in seconds.
autoscaling_target_cpu_utilization (int):
Target CPU Utilization to use for Autoscaling Replicas.
A default value of 60 will be used if not specified.
autoscaling_target_accelerator_duty_cycle (int):
Target Accelerator Duty Cycle.
Must also set accelerator_type and accelerator_count if specified.
A default value of 60 will be used if not specified.
Raises:
ValueError: If there is not current traffic split and traffic percentage
is not 0 or 100.
Expand All @@ -980,6 +1011,12 @@ def _deploy_call(
"Both `accelerator_type` and `accelerator_count` should be specified or None."
)

if not accelerator_type or not accelerator_count and autoscaling_target_accelerator_duty_cycle:
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't think this check is working as expected (I tried deploying a model with only autoscaling_target_cpu_utilization set and it raised this error). You can change it to something like:

if autoscaling_target_accelerator_duty_cycle is not None and (not accelerator_type or not accelerator_count):

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed in 9501eb0

raise ValueError(
"Both `accelerator_type` and `accelerator_count` should be set "
"when specifying autoscaling_target_accelerator_duty_cycle`"
)

deployed_model = gca_endpoint_compat.DeployedModel(
model=model.resource_name,
display_name=deployed_model_display_name,
Expand All @@ -995,7 +1032,8 @@ def _deploy_call(
in model.supported_deployment_resources_types
)
provided_custom_machine_spec = (
machine_type or accelerator_type or accelerator_count
machine_type or accelerator_type or accelerator_count or
autoscaling_target_accelerator_duty_cycle or autoscaling_target_accelerator_duty_cycle
)

# If the model supports both automatic and dedicated deployment resources,
Expand All @@ -1015,22 +1053,37 @@ def _deploy_call(
_LOGGER.info(f"Using default machine_type: {machine_type}")

if use_dedicated_resources:

dedicated_resources = gca_machine_resources_compat.DedicatedResources(
min_replica_count=min_replica_count,
max_replica_count=max_replica_count,
)

machine_spec = gca_machine_resources_compat.MachineSpec(
machine_type=machine_type
)

if autoscaling_target_cpu_utilization:
autoscaling_metric_spec = gca_machine_resources_compat.AutoscalingMetricSpec(
metric_name="aiplatform.googleapis.com/prediction/online/cpu/utilization",
target=autoscaling_target_cpu_utilization
)
dedicated_resources.autoscaling_metric_specs.extend([autoscaling_metric_spec])

if accelerator_type and accelerator_count:
utils.validate_accelerator_type(accelerator_type)
machine_spec.accelerator_type = accelerator_type
machine_spec.accelerator_count = accelerator_count

deployed_model.dedicated_resources = (
gca_machine_resources_compat.DedicatedResources(
machine_spec=machine_spec,
min_replica_count=min_replica_count,
max_replica_count=max_replica_count,
)
)
if autoscaling_target_accelerator_duty_cycle:
autoscaling_metric_spec = gca_machine_resources_compat.AutoscalingMetricSpec(
metric_name="aiplatform.googleapis.com/prediction/online/accelerator/duty_cycle",
target=autoscaling_target_accelerator_duty_cycle
)
dedicated_resources.autoscaling_metric_specs.extend([autoscaling_metric_spec])

dedicated_resources.machine_spec = machine_spec
deployed_model.dedicated_resources = dedicated_resources

elif supports_automatic_resources:
deployed_model.automatic_resources = (
Expand Down Expand Up @@ -1995,6 +2048,8 @@ def deploy(
encryption_spec_key_name: Optional[str] = None,
sync=True,
deploy_request_timeout: Optional[float] = None,
autoscaling_target_cpu_utilization: Optional[int] = None,
autoscaling_target_accelerator_duty_cycle: Optional[int] = None,
) -> Endpoint:
"""Deploys model to endpoint. Endpoint will be created if unspecified.

Expand Down Expand Up @@ -2079,6 +2134,13 @@ def deploy(
be immediately returned and synced when the Future has completed.
deploy_request_timeout (float):
Optional. The timeout for the deploy request in seconds.
autoscaling_target_cpu_utilization (int):
Target CPU Utilization to use for Autoscaling Replicas.
A default value of 60 will be used if not specified.
autoscaling_target_accelerator_duty_cycle (int):
Target Accelerator Duty Cycle.
Must also set accelerator_type and accelerator_count if specified.
A default value of 60 will be used if not specified.
Returns:
endpoint ("Endpoint"):
Endpoint with the deployed model.
Expand Down Expand Up @@ -2113,6 +2175,8 @@ def deploy(
or initializer.global_config.encryption_spec_key_name,
sync=sync,
deploy_request_timeout=deploy_request_timeout,
autoscaling_target_cpu_utilization=autoscaling_target_cpu_utilization,
autoscaling_target_accelerator_duty_cycle=autoscaling_target_accelerator_duty_cycle,
)

@base.optional_sync(return_input_arg="endpoint", bind_future_to_self=False)
Expand All @@ -2134,6 +2198,8 @@ def _deploy(
encryption_spec_key_name: Optional[str] = None,
sync: bool = True,
deploy_request_timeout: Optional[float] = None,
autoscaling_target_cpu_utilization: Optional[int] = None,
autoscaling_target_accelerator_duty_cycle: Optional[int] = None,
) -> Endpoint:
"""Deploys model to endpoint. Endpoint will be created if unspecified.

Expand Down Expand Up @@ -2218,6 +2284,13 @@ def _deploy(
be immediately returned and synced when the Future has completed.
deploy_request_timeout (float):
Optional. The timeout for the deploy request in seconds.
autoscaling_target_cpu_utilization (int):
Target CPU Utilization to use for Autoscaling Replicas.
A default value of 60 will be used if not specified.
autoscaling_target_accelerator_duty_cycle (int):
Target Accelerator Duty Cycle.
Must also set accelerator_type and accelerator_count if specified.
A default value of 60 will be used if not specified.
Returns:
endpoint ("Endpoint"):
Endpoint with the deployed model.
Expand Down Expand Up @@ -2253,6 +2326,8 @@ def _deploy(
explanation_parameters=explanation_parameters,
metadata=metadata,
deploy_request_timeout=deploy_request_timeout,
autoscaling_target_cpu_utilization=autoscaling_target_cpu_utilization,
autoscaling_target_accelerator_duty_cycle=autoscaling_target_accelerator_duty_cycle,
)

_LOGGER.log_action_completed_against_resource("model", "deployed", endpoint)
Expand Down