Skip to content

Latest commit

 

History

History
926 lines (658 loc) · 36.1 KB

concepts.rst

File metadata and controls

926 lines (658 loc) · 36.1 KB
Raw

Core Concepts

The main idea of BentoML is that the Data Science team should be able to ship their models in a way that is easy to test, easy to deploy, and easy to integrate with. And to do so, Data Scientists need tools that help them build and ship prediction services, instead of uploading pickled model files or Protobuf files to a server and hoping things work out.

bentoml.BentoService <bentoml-bentoservice-label> is the base class for building such prediction services using BentoML. And here's the minimal BentoService example from the Getting Started Guide <quickstart>:

import bentoml
from bentoml.adapters import DataframeInput
from bentoml.artifact import SklearnModelArtifact

@bentoml.env(auto_pip_dependencies=True)
@bentoml.artifacts([SklearnModelArtifact('model')])
class IrisClassifier(bentoml.BentoService):

    @bentoml.api(input=DataframeInput())
    def predict(self, df):
        return self.artifacts.model.predict(df)

Each BentoService class can contain multiple models declared through the @bentoml.artifact API, and multiple APIs for accessing this service. Each API definition requires a InputAdapter type, which defines the expected input data format of this API. BentoML provides API input adapters that covers most model serving use cases including DataframeInput, TfTensorInput, ImageInput and JsonInput.

Once an ML model is trained, a BentoService instance can bundle with the trained model with the BentoService#pack <bentoml-bentoservice-pack-label> method. This trained model is then accessible within the API function code via self.artifacts.ARTIFACT_NAME. In the example above, the artifact is initialized with the name "model", so the user code can get access to the model via self.artifacts.model.

The BentoService instance is now ready to be used for inference. But more importantly, BentoML solves the problem of saving the entire BentoService to disk, distribute the saved file, and reproduce the exact same prediction service in testing and production environment.

To save the BentoService instance, simply call the BentoService#save <bentoml-bentoservice-save-label> method. In this process, BentoML will:

  1. Saves the model based on the ML training framework and artifact type used
  2. Automatically extracts all the pip dependencies required by your BentoService class and put into a requirements.txt file
  3. Saves all the local python code dependencies
  4. Put all the generated files into one file directory, which, by default, is a location managed by BentoML
from sklearn import svm
from sklearn import datasets

clf = svm.SVC(gamma='scale')
iris = datasets.load_iris()
X, y = iris.data, iris.target
clf.fit(X, y)

# Create a iris classifier service with the newly trained model
iris_classifier_service = IrisClassifier()
iris_classifier_service.pack("model", clf)

# Test invoking BentoService instance
iris_classifier_service.predict([[5.1, 3.5, 1.4, 0.2]])

# Save the entire prediction service to file bundle
saved_path = iris_classifier_service.save()

The saved BentoService file directory is called a Bento. It is a versioned file directory that contains all the information required to run this prediction service.

Think of Bento as a docker container image or a software binary, but for an ML model. A Bento can be generated at the end of every training job, reflecting your newest code changes and training data changes. You can then easily store and distribute the Bento file, test the prediction service, and then update it to production model serving endpoint.

BentoML keeps track of all the services saved and provides web UI and CLI commands for model management. By default, BentoML saves all the model files and metadata in the local file system. For team settings, it is recommended to run a shared BentoML server for the entire team, which stores all of their Bento files and metadata in the cloud (e.g. RDS, S3). This allows your ML team to easily share, find and use each others' models and model serving endpoints. Read more about it here <guides/yatai_service>.

BentoML CLI Listing recent Bento:

> bentoml list
BENTO_SERVICE                         CREATED_AT        APIS                       ARTIFACTS
IrisClassifier:20200121114004_360ECB  2020-01-21 19:40  predict<DataframeInput>  model<SklearnModelArtifact>
IrisClassifier:20200120082658_4169CF  2020-01-20 16:27  predict<DataframeInput>  clf<PickleArtifact>
...

BentoML model management web UI:

BentoML YataiService Bento Repository Page

BentoML YataiService Bento Details Page

Creating BentoService

Users build their prediction service by subclassing bentoml.BentoService <bentoml-bentoservice-label>. It is recommended to always put the source code of your BentoService class into an individual Python file and check it into source control(e.g. git) along with your model training code.

BentoML is designed to be easily inserted to the end of your model training workflow, where you can import your BentoService class and create a Bento(a saved BentoService bundled with trained model artifacts). This makes it easy to manage, test and deploy all the models you and your team have created overtime.

Note

The BentoService class can not be defined in the __main__ module, meaning the class itself should not be defined in a Jupyter notebook cell or a python interactive shell. You can however use the %writefile magic command in jupyter notebook to write the BentoService class definition to a separate file, see example in BentoML quickstart notebook.

BentoService can only be created using Python as the programming language. But it is possible to use models trained with other languages/frameworks with BentoML and benefit from BentoML's model management, API server, dockerization and performance optimizations. To do so, you will need to create custom artifact <guides/custom_artifact>. Support for R and Spark MLlib models are on our roadmap.

Defining Service Environment

The bentoml.env <bentoml-env-label> decorator is the API for defining the environment settings and dependencies of your prediction service. And here are the types of dependencies supported by BentoML:

PyPI Packages

Python PyPI package is the most common type of dependencies. BentoML provides a mechanism that automatically figures out the PyPI packages required by your BentoService python class, simply use the auto_pip_dependencies=True option.

@bentoml.env(auto_pip_dependencies=True)
class ExamplePredictionService(bentoml.BentoService):

    @bentoml.api(input=DataframeInput())
    def predict(self, df):
        return self.artifacts.model.predict(df)

If you had specific versions of PyPI packages required for model serving that are different from your training environment, or if the auto_pip_dependencies=True option did not work for your case(bug report highly appreciated), you can also specify the list of PyPI packages manually, e.g.:

@bentoml.env(
  pip_dependencies=['scikit-learn']
)
class ExamplePredictionService(bentoml.BentoService):

    @bentoml.api(input=DataframeInput())
    def predict(self, df):
        return self.artifacts.model.predict(df)

Similarly, if you already have a list of PyPI packages required for model serving in a requirements.txt file, then simply pass in the file path via @bentoml.env(requirements_txt_file='./requirements.txt').

Conda Packages

Conda packages can be specified similarly, here's an example prediction service relying on an H2O model that requires the h2o conda packages:

@bentoml.artifacts([H2oModelArtifact('model')])
@bentoml.env(
  pip_dependencies=['pandas', 'h2o==3.24.0.2'],
  conda_channels=['h2oai'],
  conda_dependencies=['h2o==3.24.0.2']
)
class ExamplePredictionService(bentoml.BentoService):

  @bentoml.api(input=DataframeInput())
  def predict(self, df):
      return self.artifacts.model.predict(df)

Note

One caveat with Conda Packages here, is that it does not work with AWS Lambda deployment due to the limitation of the AWS Lambda platform.

Using other Docker base images ^^^^^^^^^^^^^^^^

By default, BentoML uses a default Docker base image and installs your model and its dependencies on top of it. This base image contains all of BentoML's dependencies and an installation of conda which helps BentoML to package and use the right Python version for your bundle.

However, there may be times when you need to use other Docker images (e.g. have some pre-build dependencies layers, company base image, using an Alpine-based image, etc.). BentoML makes it really easy to switch between base images by doing allowing you to specify a docker_base_image.

# e.g. using BentoML slim image
@env(docker_base_image="bentoml/model-server:0.8.12-slim-py37")
@artifacts([SklearnModelArtifact('model')])
class ExamplePredictionService(BentoService):
  ...

In fact, one such base image that many may find useful are the BentoML slim base images. The original base image weighs in at roughly ~853MB whereas the slim version weighs in at ~360MB.

> docker image ls

REPOSITORY                             TAG                   IMAGE ID            CREATED              SIZE
bentoml/model-server                   0.8.12-slim-py37      109b451ed537        6 minutes ago        360MB
bentoml/model-server                   0.8.12                f034fa23264c        33 minutes ago       853MB

This means that each image built on top of these slim images will be significantly smaller.

> docker image ls

REPOSITORY                               TAG                 IMAGE ID            CREATED              SIZE
jzhao2k19/iris                           latest              bfc9b81c7535        About a minute ago   1.54GB
jzhao2k19/iris-slim                      latest              4e8d87a0c18a        4 minutes ago        577MB

However, as with using any alternative Docker base image, there are a few things to keep in mind. The regular base image uses conda, whereas the slim image does not. This has a few consequences. BentoML uses conda to ensure the Python version used matches the one you used to save your bundle. This means that you should manually select the right slim image for your bundle. For example, if you used Python 3.8 to train your model, you would use bentoml/model-server:0.8.12-slim-py38. Currently, BentoML support Python 3.6, 3.7, and 3.8. Additionally, this means that BentoML will ignore the environment.yml, meaning that user-defined conda packages and dependencies will be ignored. In the following example, only pandas will be installed, as the conda_channels and conda_dependencies will be ignored.

@bentoml.env(
  pip_dependencies=['pandas'],
  conda_channels=['h2oai'],
  conda_dependencies=['h2o==3.24.0.2'],
  docker_base_image="bentoml/model-server:0.8.12-slim-py37"
)
class ExamplePredictionService(bentoml.BentoService):
  ...

Init Bash Script

Init setup script is used for customizing the API serving docker container. Users can insert arbitrary bash script into the docker build process via init setup script, to install extra system dependencies or do other setups required by the prediction service.

@bentoml.env(
    auto_pip_dependencies=True,
    setup_sh="./my_init_script.sh"
)
class ExamplePredictionService(bentoml.BentoService):
    ...

@bentoml.env(
    auto_pip_dependencies=True,
    setup_sh="""\
#!/bin/bash
set -e

apt-get install --no-install-recommends nvidia-driver-430
...
  """
)
class ExamplePredictionService(bentoml.BentoService):
    ...

If you have a specific docker base image that you would like to use for your API server, we are working on custom docker base image support. contact us if you are interested in helping with beta testing this feature.

Packaging Model Artifacts

BentoML's model artifact API allow users to specify the trained models required by a BentoService. BentoML automatically handles model serialization and deserialization when saving and loading a BentoService.

Thus BentoML asks the user to choose the right Artifact class for the machine learning framework they are using. BentoML has built-in artifact class for most popular ML frameworks and you can find the list of supported frameworks here <api/artifacts>. If the ML framework you're using is not in the list, let us know and we will consider adding its support.

To specify the model artifacts required by your BentoService, use the bentoml.artifacts decorator and gives it a list of artifact types. And give each model artifact a unique name within the prediction service. Here's an example prediction service that packs two trained models:

import bentoml
from bentoml.adapters import DataframeInput
from bentoml.artifact import SklearnModelArtifact, XgboostModelArtifact

@bentoml.env(auto_pip_dependencies=True)
@artifacts([
    SklearnModelArtifact("model_a"),
    XgboostModelArtifact("model_b")
])
class MyPredictionService(bentoml.BentoService):

    @bentoml.api(input=DataframeInput())
    def predict(self, df):
        # assume the output of model_a will be the input of model_b in this example:
        df = self.artifacts.model_a.predict(df)

        return self.artifacts.model_b.predict(df)
svc = MyPredictionService()
svc.pack('model_a', my_sklearn_model_object)
svc.pack('model_b', my_xgboost_model_object)
svc.save()

For most model serving scenarios, we recommend one model per prediction service, and decouple non-related models into separate services. The only exception is when multiple models are depending on each other, such as the example above.

API Function and Adapters

BentoService API is the entry point for clients to access a prediction service. It is defined by writing the API handling function(a class method within the BentoService class) which gets called when client sent an inference request. User will need to annotate this method with @bentoml.api decorator and pass in an InputAdapter instance, which defines the desired input format for the API function. For example, if your model is expecting tabular data as input, you can use DataframeInput for your API, e.g.:

class ExamplePredictionService(bentoml.BentoService):

    @bentoml.api(input=DataframeInput())
    def predict(self, df):
        assert type(df) == pandas.core.frame.DataFrame
        return postprocessing(model_output)

When using DataframeInput, BentoML will convert the inference requests sent from the client, either in the form of a JSON HTTP request or a CSV file, into a pandas.DataFrame object and pass it down to the user-defined API function.

User can write arbitrary python code within the API function that process the data. Besides passing the prediction input data to the model for inference, user can also write Python code for data fetching, data pre-processing and post-processing within the API function. For example:

from my_lib import preprocessing, postprocessing, fetch_user_profile_from_database

class ExamplePredictionService(bentoml.BentoService):

    @bentoml.api(input=DataframeInput())
    def predict(self, df):
        user_profile_column = fetch_user_profile_from_database(df['user_id'])
        df['user_profile'] = user_profile_column
        model_input = preprocessing(df)
        model_output = self.artifacts.model.predict(model_input)
        return postprocessing(model_output)

Note

Check out the list of API InputAdapters <api/adapters> that BentoML provides.

It is important to notice that in BentoML, the input variable passed into the user-defined API function is always a list of inference inputs. BentoML users must make sure their API function code is processing a batch of input data.

This design made it possible for BentoML to do Micro-Batching in online API serving, which is one of the most effective optimization technique for model serving systems.

API Function Return Value

The output of an API function can be any of the follow types:

pandas.DataFrame
pandas.Series
numpy.ndarray
tensorflow.Tensor

# List of JSON Serializable
# JSON = t.Union[str, int, float, bool, None, t.Mapping[str, 'JSON'], t.List['JSON']]
List[JSON]

It is user API function's responsibility to make sure the list of prediction results matches the order of input sequence and have the exact same length.

Note

It is possible for API function to handle and return a single inference request at one time before BentoML 0.7.0, but it is no longer recommended after introducing the adaptive micro batching feature.

Service with Multiple APIs

A BentoService can contain multiple APIs, which makes it easy to build prediction service that supports different access patterns for different clients, e.g.:

from my_lib import process_custom_json_format

class ExamplePredictionService(bentoml.BentoService):

    @bentoml.api(input=DataframeInput())
    def predict(self, df: pandas.Dataframe):
        return self.artifacts.model.predict(df)

    @bentoml.api(input=JsonInput())
    def predict_json(self, json_arr):
        df = process_custom_json_format(json-arr)
        return self.artifacts.model.predict(df)

Make sure to give each API a different name. BentoML uses the method name as the API's name, which will become part the serving endpoint it generates.

Operational API

User can also create APIs that, instead of handling an inference request, handles request for updating prediction service configs or retraining models with new arrived data. Operational API is still a beta feature, contact us if you're interested in learning more.

Saving BentoService

After writing your model training/evaluation code and BentoService definition, here are the steps required to create a BentoService instance and save it for serving:

  1. Model Training
  2. Create BentoService instance
  3. Pack trained model artifacts with BentoService#pack <bentoml-bentoservice-pack-label>
  4. Save to a Bento with BentoService#save <bentoml-bentoservice-save-label>

As illustrated in the previous example:

from sklearn import svm
from sklearn import datasets

# 1. Model training
clf = svm.SVC(gamma='scale')
iris = datasets.load_iris()
X, y = iris.data, iris.target
clf.fit(X, y)

# 2. Create BentoService instance
iris_classifier_service = IrisClassifier()

# 3. Pack trained model artifacts
iris_classifier_service.pack("model", clf)

# 4. Save
saved_path = iris_classifier_service.save()

How Save Works

BentoService#save_to_dir(path) <bentoml-bentoservice-save-label> is the primitive operation for saving the BentoService to a target directory. save_to_dir serializes the model artifacts and saves all the related code, dependencies and configs into a the given path.

Users can then use bentoml.load(path) <bentoml-load-label> to load the exact same BentoService instance back from the saved file path. This made it possible to easily distribute your prediction service to test and production environment in a consistent manner.

BentoService#save <bentoml-bentoservice-save-label> essentially calls BentoService#save_to_dir(path) <bentoml-bentoservice-save-label> under the hood, while keeping track of all the prediction services you've created and maintaining the file structures and metadata information of those saved bundle.

Model Serving

Once a BentoService is saved as a Bento, it is ready to be deployed for many different types of serving workloads.

There are 3 main types of model serving -

  • Online Serving - clients access predictions via API endpoints in near real-time
  • Offline Batch Serving - pre-compute predictions and save results in a storage system
  • Edge Serving - distribute model and run it on mobile or IoT devices

BentoML has great support for online serving and offline batch serving. It has a high-performance API server that can load a saved Bento and expose a REST API for client access. It also provide tools to load the Bento and feed it with a batch of inputs for offline inference. Edge serving is only supported when the client has the Python runtime, e.g. model serving in a router or a Raspberry Pi.

Online API Serving

Once a BentoService is saved, you can easily start the REST API server to test out sending request and interacting with the server. For example, after saving the BentoService in the Getting Started Guide <quickstart>, you can start a API server right away with:

bentoml serve IrisClassifier:latest

If you are using save_to_dir <bentoml-bentoservice-save-label> , or you have directly copied the saved Bento file directory from other machine, the BentoService IrisClassifier is not registered with your local BentoML repository. In that case, you can still start the server by providing the path to the saved BentoService:

bentoml serve $saved_path

The REST API request format is determined by each API's input type and input config. More details can be found in the BentoML API InputAdapters References <bentoml-api-adapters-label>.

For running production API server, make sure to run bentoml serve-gunicorn command instead, or use Docker container for deployment.

bentoml serve-gunicorn $saved_path --enable-microbatch --workers=2 --port=3000

API Server Dockerization

When you are ready to deploy the service to production, a docker image containing your model API server can be easily created with BentoML. When saving a Bento, a Dockerfile is also generated by BentoML in the same directory. Dockerfile is a text document that contains all the commands required for creating a docker image, and docker build command builds an image from a Dockerfile.

# Find the saved path of the latest version of IrisClassifier Bento
saved_path=$(bentoml get IrisClassifier:latest -q | jq -r ".uri.uri")

# Build docker image using saved_path directory as the build context, replace the
# {username} below to your docker hub account name
docker build -t {username}/iris_classifier_bento_service $saved_path

# Run a container with the docker image built and expose port 5000
docker run -p 5000:5000 {username}/iris_classifier_bento_service

# Push the docker image to docker hub for deployment
docker push {username}/iris_classifier_bento_service

Here's an example deployment you can create in a Kubernetes cluster using the docker image built above:

apiVersion: apps/v1 # for k8s versions before 1.9.0 use apps/v1beta2
kind: Deployment
metadata:
  name: iris_classifier
spec:
  selector:
    matchLabels:
      app: iris_classifier
  replicas: 3
  template:
    metadata:
      labels:
        app: iris_classifier
    spec:
      containers:
      - name: iris_classifier_bento_service
        image: {username}/iris_classifier_bento_service:latest
        ports:
        - containerPort: 5000

Adaptive Micro-Batching

Micro batching is a technique where incoming prediction requests are grouped into small batches to achieve the performance advantage of batch processing in model inference tasks. BentoML implemented such a micro batching layer that is inspired by the paper Clipper: A Low-Latency Online Prediction Serving System.

Given the mass performance improvement a model serving system get from micro-batching, BentoML APIs were designed to work with micro-batching without any code changes on the user side. It is why all the API InputAdapters are designed to accept a list of input data, as described in the concepts-api-func-and-adapters section.

Currently, micro-batching is still a beta feature, users can enable micro-batching by passing a flag when running BentoML API server:

# Launch micro batching API server from CLI
bentoml serve-gunicorn $saved_path --enable-microbatch

# Launch model server docker image with micro batching enabled
docker run -p 5000:5000 -e BENTOML_ENABLE_MICROBATCH=True {username}/iris-classifier:latest

Programmatic Access

A saved BentoService can also be loaded from saved Bento and access directly from Python. There are two main ways this can be done:

  1. Load from a saved Bento directory with bentoml.load(path) <bentoml-load-label> API
import bentoml

bento_service = bentoml.load(saved_path)
result = bento_service.predict(input_data)

The benefit of this approach is its flexibility. Users can easily invoke saved BentoService in their backend applications, and programmatically choose which model to load and how they are used for inference.

  1. Install BentoService as a PyPI package

A Bento directory is also pip-installable as demonstrated in the Getting Started Guide <quickstart>:

pip install $saved_path
# Your bentoML model class name will become the package name
import IrisClassifier

installed_svc = IrisClassifier.load()
installed_svc.predict([[5.1, 3.5, 1.4, 0.2]])

This approach made sure that all the required pip dependencies are installed for the BentoService when being installed. It is convenient when your Data Science team is shipping the prediction service as a standalone python package that can be shared by a variety of different developers to integrate with.

  1. Command-Line Access

Similarly, a Bento can be loaded for inference using the BentoML CLI tool. The CLI command bentoml is available once you've installed BentoML via pip. And to load a saved Bento file, simply use the bentoml run command and give it either the name and version pair, or the Bento's path:

# With BentoService name and version pair
bentoml run IrisClassifier:latest predict --input='[[5.1, 3.5, 1.4, 0.2]]'
bentoml run IrisClassifier:latest predict --input='./iris_test_data.csv'

# With BentoService's saved path
bentoml run $saved_path predict --input='[[5.1, 3.5, 1.4, 0.2]]'
bentoml run $saved_path predict --input='./iris_test_data.csv'

Or if you have already pip-install'd the BentoService, it provides a CLI command specifically for this BentoService. The CLI command is the same as the BentoService class name:

IrisClassifier run predict --input='[[5.1, 3.5, 1.4, 0.2]]'
IrisClassifier run predict --input='./iris_test_data.csv'

Offline Batch Serving

All three methods in the Programmatic Access section above, can be used for doing single-machine batch offline model serving. Depends on the format of input data. An inference computation job can be started either with BentoService's Python API or Bash CLI command. This made it very easy to integrate with Job scheduling tools such as Apache Airflow and Celery.

For batch serving on large dataset running on a cluster, BentoML team is building a Apache Spark UDF loader for BentoService. This feature is still in Beta testing phase. Contact us if you are interested in helping to test or improve it.

Model Management

By default, BentoService#save <bentoml-bentoservice-save-label> will save all the BentoService saved bundle files under ~/bentoml/repository/ directory, following by the service name and service version as sub-directory name. And all the metadata of saved BentoService are stored in a local SQLite database file at ~/bentoml/storage.db.

Users can easily query and use all the BentoService they have created, for example, to list all the BentoService created:

> bentoml list
BENTO_SERVICE                                   AGE                  APIS                        ARTIFACTS
IrisClassifier:20200323212422_A1D30D            1 day and 22 hours   predict<DataframeInput>   model<SklearnModelArtifact>
IrisClassifier:20200304143410_CD5F13            3 weeks and 4 hours  predict<DataframeInput>   model<SklearnModelArtifact>
SentimentAnalysisService:20191219090607_189CFE  13 weeks and 6 days  predict<DataframeInput>   model<SklearnModelArtifact>
TfModelService:20191216125343_06BCA3            14 weeks and 2 days  predict<JsonInput>        model<TensorflowSavedModelArtifact>

> bentoml get IrisClassifier
BENTO_SERVICE                         CREATED_AT        APIS                       ARTIFACTS
IrisClassifier:20200121114004_360ECB  2020-01-21 19:45  predict<DataframeInput>  model<SklearnModelArtifact>
IrisClassifier:20200121114004_360ECB  2020-01-21 19:40  predict<DataframeInput>  model<SklearnModelArtifact>

> bentoml get IrisClassifier:20200323212422_A1D30D
{
  "name": "IrisClassifier",
  "version": "20200323212422_A1D30D",
  "uri": {
    "type": "LOCAL",
    "uri": "/Users/chaoyu/bentoml/repository/IrisClassifier/20200323212422_A1D30D"
  },
  "bentoServiceMetadata": {
    "name": "IrisClassifier",
    "version": "20200323212422_A1D30D",
    "createdAt": "2020-03-24T04:24:39.517239Z",
    "env": {
      "condaEnv": "name: bentoml-IrisClassifier\nchannels:\n- defaults\ndependencies:\n- python=3.7.5\n- pip\n",
      "pipDependencies": "bentoml==0.6.3\nscikit-learn",
      "pythonVersion": "3.7.5"
    },
    "artifacts": [
      {
        "name": "model",
        "artifactType": "SklearnModelArtifact"
      }
    ],
    "apis": [
      {
        "name": "predict",
        "InputType": "DataframeInput",
        "docs": "BentoService API",
        "inputConfig": {
          "orient": "records",
          "typ": "frame",
          "is_batch_input": true,
          "input_dtypes": null
        }
      }
    ]
  }
}

Similarly, the Bento name and version pair can be used to load and run those BentoService directly. For example:

> bentoml serve IrisClassifier:latest
* Serving Flask app "IrisClassifier" (lazy loading)
* Environment: production
  WARNING: This is a development server. Do not use it in a production deployment.
  Use a production WSGI server instead.
* Debug mode: off
* Running on http://127.0.0.1:5000/ (Press CTRL+C to quit)

> bentoml run IrisClassifier:latest predict --input='[[5.1, 3.5, 1.4, 0.2]]'
[0]

Customizing Model Repository

BentoML has a standalone component YataiService that handles model storage and deployment. BentoML uses a local YataiService instance by default, which saves BentoService files to ~/bentoml/repository/ directory and other metadata to ~/bentoml/storage.db.

Users can also customize this to make it work for team settings, making it possible for a team of data scientists to easily share, use and deploy models and prediction services created by each other. To do so, the user will need to setup a host server that runs YataiService, from BentoML cli command `yatai-service-start`:

> bentoml yatai-service-start --help
Usage: bentoml yatai-service-start [OPTIONS]

  Start BentoML YataiService for model management and deployment

Options:
  --db-url TEXT         Database URL following RFC-1738, and usually can
                        include username, password, hostname, database name as
                        well as optional keyword arguments for additional
                        configuration
  --repo-base-url TEXT  Base URL for storing saved BentoService bundle files,
                        this can be a filesystem path(POSIX/Windows), or an S3
                        URL, usually starts with "s3://"
  --grpc-port INTEGER   Port for Yatai server
  --ui-port INTEGER     Port for Yatai web UI
  --ui / --no-ui        Start BentoML YataiService without Web UI
  -q, --quiet           Hide all warnings and info logs
  --verbose, --debug    Show debug logs when running the command
  --help                Show this message and exit.

BentoML provides a pre-built docker image for running YataiService. For each BentoML release, a new image will be pushed to docker hub under bentoml/yatai-service with the same image tag as the PyPI package version. For example, use the following command to start a YataiService of BentoML version 0.7.2, loading data from your local BentoML repository under the local ~/bentoml directory:

> docker run -v ~/bentoml:/bentoml \
    -p 3000:3000 \
    -p 50051:50051 \
    bentoml/yatai-service:0.7.2 \
    --db-url=sqlite:///bentoml/storage.db \
    --repo-base-url=/bentoml/repository

The recommended way to deploy YataiService for teams, is to back it by a remote PostgreSQL database and an S3 bucket. For example, deploy the following docker container to run a YataiService configured with remote database and S3 storage, as well as AWS credentials for managing deployments created on AWS:

> docker run -p 3000:3000 -p 50051:50051 \
    -e AWS_SECRET_ACCESS_KEY=... -e AWS_ACCESS_KEY_ID=...  \
    bentoml/yatai-service:0.7.2 \
    --db-url postgresql://scott:tiger@localhost:5432/bentomldb \
    --repo-base-url s3://my-bentoml-repo/

* Starting BentoML YataiService gRPC Server
* Debug mode: off
* Web UI: running on http://127.0.0.1:3000
* Running on 127.0.0.1:50051 (Press CTRL+C to quit)
* Usage: `bentoml config set yatai_service.url=127.0.0.1:50051`
* Help and instructions: https://docs.bentoml.org/en/latest/guides/yatai_service.html
* Web server log can be found here: /Users/chaoyu/bentoml/logs/yatai_web_server.log

After deploying the YataiService server, get the server IP address and run the following command to configure BentoML client to use this remote YataiService for model management and deployments. You will need to replace 127.0.0.1 with an IP address or URL that is accessible for your team:

bentoml config set yatai_service.url=127.0.0.1:50051

Once you've run the command above, all the BentoML model management operations will be sent to the remote server, including saving BentoService, query saved BentoServices or creating model serving deployments.

Note

BentoML's YataiService does not provide any kind of authentication. To secure your deployment, we recommend only make the server accessible within your VPC for you data science team to have access.

BentoML team also provides hosted YataiService for enterprise teams, that has all the security best practices built-in, to bootstrap the end-to-end model management and model serving deployment workflow. Contact us to learn more about our offerings.