Home

Awesome

<div align="center">

<a href="https://app.clear.ml"><img src="https://github.com/allegroai/clearml/blob/master/docs/clearml-logo.svg?raw=true" width="250px"></a>

ClearML Serving - Model deployment made easy

clearml-serving v1.3.1 </br> :sparkles: Model Serving (ML/DL) Made Easy :tada: <br> :fire: NEW version 1.3 :rocket: 20% faster !

GitHub license PyPI pyversions PyPI version shields.io Artifact Hub Slack Channel

🌟 ClearML is open-source - Leave a star to support the project! 🌟

</div>

clearml-serving is a command line utility for model deployment and orchestration.
It enables model deployment including serving and preprocessing code to a Kubernetes cluster or custom container based solution.

:fire: NEW :confetti_ball: Take it for a spin with a simple docker-compose command :magic_wand: :sparkles:

<a><img src="https://github.com/allegroai/clearml-serving/blob/main/docs/design_diagram.png?raw=true" width="100%"></a>

Features:

ClearML Serving Design

ClearML Serving Design Principles

Modular , Scalable , Flexible , Customizable , Open Source

Installation

Prerequisites

:nail_care: Initial Setup

  1. Setup your ClearML Server or use the Free tier Hosting
  2. Setup local access (if you haven't already), see instructions here
  3. Install clearml-serving CLI:
pip3 install clearml-serving
  1. Create the Serving Service Controller
  1. Write down the Serving Service UID
  2. Clone clearml-serving repository
git clone https://github.com/allegroai/clearml-serving.git
  1. Edit the environment variables file (docker/example.env) with your clearml-server credentials and Serving Service UID. For example, you should have something like
cat docker/example.env

  CLEARML_WEB_HOST="https://app.clear.ml"
  CLEARML_API_HOST="https://api.clear.ml"
  CLEARML_FILES_HOST="https://files.clear.ml"
  CLEARML_API_ACCESS_KEY="<access_key_here>"
  CLEARML_API_SECRET_KEY="<secret_key_here>"
  CLEARML_SERVING_TASK_ID="<serving_service_id_here>"
  1. Spin the clearml-serving containers with docker-compose (or if running on Kubernetes use the helm chart)
cd docker && docker-compose --env-file example.env -f docker-compose.yml up

If you need Triton support (keras/pytorch/onnx etc.), use the triton docker-compose file

cd docker && docker-compose --env-file example.env -f docker-compose-triton.yml up

:muscle: If running on a GPU instance w/ Triton support (keras/pytorch/onnx etc.), use the triton gpu docker-compose file

cd docker && docker-compose --env-file example.env -f docker-compose-triton-gpu.yml up

Notice: Any model that registers with "Triton" engine, will run the pre/post processing code on the Inference service container, and the model inference itself will be executed on the Triton Engine container.

:ocean: Optional: advanced setup - S3/GS/Azure access

To add access credentials and allow the inference containers to download models from your S3/GS/Azure object-storage, add the respective environment variables to your env files (example.env) See further details on configuring the storage access here

AWS_ACCESS_KEY_ID
AWS_SECRET_ACCESS_KEY
AWS_DEFAULT_REGION

GOOGLE_APPLICATION_CREDENTIALS

AZURE_STORAGE_ACCOUNT
AZURE_STORAGE_KEY

:information_desk_person: Concepts

CLI - Secure configuration interface for on-line model upgrade/deployment on running Serving Services

Serving Service Task - Control plane object storing configuration on all the endpoints. Support multiple separated instance, deployed on multiple clusters.

Inference Services - Inference containers, performing model serving pre/post processing. Also support CPU model inferencing.

Serving Engine Services - Inference engine containers (e.g. Nvidia Triton, TorchServe etc.) used by the Inference Services for heavier model inference.

Statistics Service - Single instance per Serving Service collecting and broadcasting model serving & performance statistics

Time-series DB - Statistics collection service used by the Statistics Service, e.g. Prometheus

Dashboards - Customizable dashboard-ing solution on top of the collected statistics, e.g. Grafana

:point_right: Toy model (scikit learn) deployment example

  1. Train toy scikit-learn model
  1. Register the new Model on the Serving Service
  1. Spin the Inference Container
  1. Test new model inference endpoint

Notice, now that we have an inference container running, we can add new model inference endpoints directly with the CLI. The inference container will automatically sync once every 5 minutes.

Notice On the first few requests the inference container needs to download the model file and preprocessing python code, this means the request might take a little longer, once everything is cached, it will return almost immediately.

Notes:

Review the model repository in the ClearML web UI, under the "serving examples" Project on your ClearML account/server (free hosted or self-deployed).

Inference services status, console outputs and machine metrics are available in the ClearML UI in the Serving Service project (default: "DevOps" project)

To learn more on training models and the ClearML model repository, see the ClearML documentation

:turtle: Registering & Deploying new models manually

Uploading an existing model file into the model repository can be done via the clearml RestAPI, the python interface, or with the clearml-serving CLI.

To learn more on training models and the ClearML model repository, see the ClearML documentation

Notice we can also provide a differnt storage destination for the model, such as S3/GS/Azure, by passing --destination="s3://bucket/folder", gs://bucket/folder, azure://bucket/folder. Yhere is no need to provide a unique path tp the destination argument, the location of the model will be a unique path based on the serving service ID and the model name

:rabbit: Automatic model deployment

The clearml Serving Service support automatic model deployment and upgrades, directly connected with the model repository and API. When the model auto-deploy is configured, a new model versions will be automatically deployed when you "publish" or "tag" a new model in the clearml model repository. This automation interface allows for simpler CI/CD model deployment process, as a single API automatically deploy (or remove) a model from the Serving Service.

:bulb: Automatic model deployment example

  1. Configure the model auto-update on the Serving Service
  1. Deploy the Inference container (if not already deployed)
  2. Publish a new model the model repository
from clearml import Model
Model(model_id="unique_model_id_here").publish()
  1. The new model is available on a new endpoint version (1), test with: curl -X POST "http://127.0.0.1:8080/serve/test_model_sklearn_auto/1" -H "accept: application/json" -H "Content-Type: application/json" -d '{"x0": 1, "x1": 2}'

:bird: Canary endpoint setup

Canary endpoint deployment add a new endpoint where the actual request is sent to a preconfigured set of endpoints with pre-provided distribution. For example, let's create a new endpoint "test_model_sklearn_canary", we can provide a list of endpoints and probabilities (weights).

clearml-serving --id <service_id> model canary --endpoint "test_model_sklearn_canary" --weights 0.1 0.9 --input-endpoints test_model_sklearn/2 test_model_sklearn/1

This means that any request coming to /test_model_sklearn_canary/ will be routed with probability of 90% to /test_model_sklearn/1/ and with probability of 10% to /test_model_sklearn/2/.

Note:

As with any other Serving Service configuration, we can configure the Canary endpoint while the Inference containers are already running and deployed, they will get updated in their next update cycle (default: once every 5 minutes)

We can also prepare a "fixed" canary endpoint, always splitting the load between the last two deployed models:

clearml-serving --id <service_id> model canary --endpoint "test_model_sklearn_canary" --weights 0.1 0.9 --input-endpoints-prefix test_model_sklearn/

This means that is we have two model inference endpoints: /test_model_sklearn/1/ and /test_model_sklearn/2/. The 10% probability (weight 0.1) will match the last (order by version number) endpoint, i.e. /test_model_sklearn/2/ and the 90% will match /test_model_sklearn/2/. When we add a new model endpoint version, e.g. /test_model_sklearn/3/, the canary distribution will automatically match the 90% probability to /test_model_sklearn/2/ and the 10% to the new endpoint /test_model_sklearn/3/.

Example:

  1. Add two endpoints:
  1. Add Canary endpoint:
  1. Test Canary endpoint:

:bar_chart: Model monitoring and performance metrics :bell:

Grafana Screenshot

ClearML serving instances send serving statistics (count/latency) automatically to Prometheus and Grafana can be used to visualize and create live dashboards.

The default docker-compose installation is preconfigured with Prometheus and Grafana, do notice that by default data/ate of both containers is not persistent. To add persistence we do recommend adding a volume mount.

You can also add many custom metrics on the input/predictions of your models. Once a model endpoint is registered, adding custom metric can be done using the CLI. For example, assume we have our mock scikit-learn model deployed on endpoint test_model_sklearn, we can log the requests inputs and outputs (see examples/sklearn/preprocess.py example):

clearml-serving --id <serving_service_id_here> metrics add --endpoint test_model_sklearn --variable-scalar
x0=0,0.1,0.5,1,10 x1=0,0.1,0.5,1,10 y=0,0.1,0.5,0.75,1

This will create a distribution histogram (buckets specified via a list of less-equal values after = sign), that we will be able to visualize on Grafana. Notice we can also log time-series values with --variable-value x2 or discrete results (e.g. classifications strings) with --variable-enum animal=cat,dog,sheep. Additional custom variables can be in the preprocess and postprocess with a call to collect_custom_statistics_fn({'new_var': 1.337}) see clearml_serving/preprocess/preprocess_template.py

With the new metrics logged we can create a visualization dashboard over the latency of the calls, and the output distribution.

Grafana model performance example:

Notice: If not specified all serving requests will be logged, to change the default configure "CLEARML_DEFAULT_METRIC_LOG_FREQ", for example CLEARML_DEFAULT_METRIC_LOG_FREQ=0.2 means only 20% of all requests will be logged. You can also specify per endpoint log frequency with the clearml-serving CLI. Check the CLI documentation with clearml-serving metrics --help

:fire: Model Serving Examples

:pray: Status

Contributing

PRs are always welcomed :heart: See more details in the ClearML Guidelines for Contributing.