Hydrosphere has an internal Model Registry as centralized storage for Model Versions. When you build a Dockerized model and upload it to Hydrosphere or create new model versions, they get uploaded/stored to the configured model registry in the form of images. This organizes and simplifies model management across the platform and production lifecycle.

Gateway

Gateway is a service responsible for routing requests to/from or between Servables and Applications and validating these requests for matching a Model's/Application signature.

The Gateway maps a model’s name to a corresponding container. Whenever it receives a request via HTTP API, GRPC, or Kafka Streams, it communicates with that container via the gRPC protocol.

Manager

Manager is responsible for:

• Building a Docker Image from your ML model for future deployment

• Storing these images inside a Docker Registry deployed alongside with

  manager service

• Versioning these images as Model Versions

• Creating running instances of these Model Versions called Servables

  inside Kubernetes cluster

• Combining multiple Model Versions into a linear graph with a single

  endpoint called Application

Automatic Outlier Detection

Sonar

Sonar service is responsible for managing metrics, training and production data storage, calculating profiles, and shadowing data to the Model Versions which are used as an outlier detection metrics.

Drift Report

What is Hydrosphere?

Hydrosphere is an open-source MLOps platform for deploying, managing, and monitoring ML models in production with Kubernetes.

Hydrosphere supports all major machine learning frameworks, including Tensorflow, Keras, PyTorch, XGBoost, scikit-learn, fastai, etc. The platform is designed to effectively measure performance and health metrics of your production models, making it possible to spot early signs of performance drops and data drifts, get insights into why they happen.

Hydrosphere offers immediate value to ML-based products:

• Сovers all aspects of the production ML lifecycle - model versioning & deployment, traffic & contract management, data monitoring, gaining insights.

• Easy & fast management of production models that brings models to production in minutes by reducing time to upload, update, and roll your models into production.

• Allows to create reproducible, observable, and explainable machine learning pipelines.

• Provides understanding and control of models’ performance in production via data and target metrics analysis.

• Adds in-depth observability for your production models and data flowing through them.

• Improves business metrics of ML-based products as a result of a reduction in MTTR and MTTD incidents related to ML models due to early alerts once data drifts happen.

Why Hydrosphere?

Production ML is a dangerous place where numerous things can and usually do go wrong, making issues harder to discover and fix. Hydrosphere automates MLOps in the production part of the ML lifecycle combining best practices of CI/CD and DevOps and putting special emphasis on monitoring performance of ML models after their deployment.

MLOps problems Hydrosphere addresses:

• Non-interpretable, biased models

• Integration between the tools of each step of production ML lifecycle

• Long time to find & debug issues with production ML Models

• Monitoring for Model Degradation and Performance Loss

• Understanding the reasons behind wrong predictions

What Hydrosphere is not

Hydrosphere is not an ML model training framework. Before using Hydrosphere, you need to train your models with one of many existing frameworks for ML model training.

We suggest you use one of the orchestrators, such as Kubeflow or Airflow, to deliver your model to the Hydrosphere.

Hydrosphere Components

Hydrosphere platform includes all steps of a production ML model cycle - Versioning, Deployment, Monitoring, and Maintenance. This combination allows us to use a single tool to build an observable, reproducible, and scalable workflow, and start getting early warnings once anything goes wrong. These steps of an ML lifecycle are divided between three components that make up the Hydrosphere platform - Serving, Monitoring, and Interpretability.

Serving

Hydro Serving is responsible for framework-agnostic model deployment and management. It allows Data Scientists to upload, version, combine into linear pipelines and deploy ML models trained in any ML framework to a Docker/Kubernetes cluster and expose HTTP/gRPC/Kafka API to other parties.

Monitoring

Hydro Monitoring tracks model performance over time, raising alerts in case of detected issues. It provides a real-time updated UI, where you can monitor your models to see service health and usage. This constant monitoring of model health is crucial for any ML-based business as it’s tied to business and financial metrics.

Hydrosphere is capable of monitoring model quality with or without getting additional labeled data. Labeled data is often used in production drawing conclusions about the quality of a model’s predictions. Sometimes it is hard to get labeled data in production in a timely and cost-effective manner, especially when you deal with large volumes of complex data. Hydrosphere circumvents this issue by analyzing data that flows through a model as a proxy evaluating model quality to detect if ML models start to degrade and make unreliable predictions due to production data drifts from training data.

Interpretability

Hydrosphere Interpretability provides human-readable explanations of the predictions made by your ML models, as well as the explanations of monitoring analytics made by Hydrosphere Monitoring. It helps to evaluate and analyze models and understand what features influence their decisions. The Interpretability component demystifies your ML process, provides a new level of confidence about the reasons behind your models’ decisions and a certain level of trust business can rely on.

Resource definitions

Resource definitions describe Models, Applications, and Deployment Configurations in the YAML format. You can learn more about them in the section.

Serving

Models & Model Versions

A Model is a machine learning model or a processing function that consumes provided inputs and produces predictions or transformations.

Within the Hydrosphere platform, we break down a model into its versions. Each Model version represents a single Docker image containing all the artifacts that you have uploaded to the platform. Consequently, Model is a group of Model versions with the same name.

Runtimes

A Runtime is a Docker image with the predefined gRPC interface which loads and serves your model.

Servable

Servable is a deployed instance of a Model version combined with a Runtime. It exposes a gRPC endpoint that can be used to send requests.

Users should not use Servables as-is, since they are designed to be building blocks, rather than inference endpoints. Hydrosphere provides a better alternative to deploy a Model version — Application.

Applications

An Application is a pipeline of one or more stages, each consisting of one or multiple Model Versions. Data sent to an application stage is shadowed to all of its model versions. The output of a stage is picked randomly with respect to weights.

When a user creates an Application, the Manager service automatically deploys appropriate Servables. The Application handles monitoring of your models and can perform A/B traffic splits.

Each Application has publicly available HTTP and gRPC endpoints that you can send requests to.

Deployment Configurations

A Deployment Configuration is a collection of Kubernetes settings that you can set for your Servables and Model Versions used inside of Application stages.

Deployment Configuration covers:

• Horizontal Pod Autoscaler specs

• Container Specs

  • Resource requirements: limits and requests

• Pod Specs

  • Node Selectors

  • Affinity

  • Tolerations

• Deployment Specs

  • Replicas count

Model's Signature

A Model's Signature is a specification of your model computation which identifies the name of a function with its inputs and outputs, including their names, shapes, and data types.

Example of a signature defined in a YAML file:

Field

A Field is a basic element of a Model's signature. It has a name, shape, data type, and profile.

Example of a model's signature field defined in a YAML file:

Field`s Profile

A Profile is a special tag that tells how Hydrosphere should interpret the field's data.

There are multiple available tags: Numerical, Categorical, Image, Text, Audio, Video, etc.

Monitoring

Metrics

Data coming through deployed Model Versions can be monitored with metrics.

Metric is a Model Version that takes a combination of inputs & outputs from another monitored Model Version, receives every request and response from the monitored model, produces a single value, and compares it with a threshold to determine whether this request was healthy or not.

Every request is evaluated against all metrics assigned to the model.

Checks

A check is a boolean condition associated with a field of a Model Version signature which shows for every request whether the field value is acceptable or not.

For example, Min/Max checks ensure that a field value is in an acceptable range which is inferred from training data values.

Hydrosphere allows you to A/B test your ML models in production.

A/B testing is a great way of measuring how well your models perform or which of your model versions is more effective and taking data-driven decisions upon this knowledge.

Production ML applications always have specific goals, for example driving as many users as possible to perform some action. To achieve these goals, it’s necessary to run online experiments and compare model versions using metrics in order to measure your progress against them. This approach allows to track whether your development efforts lead to desired outcomes.

To perform a basic A/B experiment on an application consisting of 2 variants of a model, you need to train and upload both versions to Hydrosphere, create an application with a single execution stage from them, invoke it by simulating production data flow, then analyze production data using metrics of your choice.

Learn how to set up an A/B application:

• A/B Analysis for a Recommendation Model

Serving

• Model Registry

• Inference Pipelines

• A/B Model Version Deployment

• Traffic Shadowing

• Language-Agnostic Deployment

Monitoring

• Automatic Outlier Detection

• Data Drift Report

• Monitoring Dashboard and Data Health Metrics

• Alerts

Interpretability

• Prediction Explanations

• Data Projection

Third-Party Integrations

• Kubeflow Components

• AWS Sagemaker

Interpretability provides EDA (Exploratory Data Analysis) and explanations for predictions made by your models to make predictions understandable and actionable. It also produces explanations for monitoring metrics to let you know why a particular request was marked as an outlier. The component consists of 2 services:

• Explanations

• Data Projections

Both services are built with Celery to run asynchronous tasks from apps and consists of a client, a worker, and a broker that mediates in between. A client generates a task and initiates it by adding a message to a queue, а broker delivers it to a worker, then the worker executes the task.

Interpretability services use MongoDB as both a Celery broker and backend storage to save task results. To save and retrieve model training and production data, the Interpretability component uses S3 storage.

When Explanation or Data Projection receives a task they create a new temporary Servable specifically for the model they need to make an explanation for. They use this Servable to run data through it in order to make new predictions and delete it after.

Prediction Explanations

Prediction Explanations generate explanations of model predictions to help you understand them. Depending on the type of data your model uses, it provides an explanation as either a set of logical predicates if your data is in a tabular format or a saliency map if your data is in the image format. Saliency Map is a heat map that highlights parts of a picture that a prediction was based on.

Data Projections

Data Projection visualizes high-dimensional data in a 2D scatter plot with an automatically trained UMAP transformer to let you evaluate data structure and spot clusters, outliers, novel data, or any other patterns. It is especially helpful if your model works with high-dimensional data, such as images or text embeddings.

Hydrosphere is composed of several microservices, united to efficiently serve and monitor machine learning models in production. Hydrosphere features are divided between multiple services. You can learn more about each of them in this section.

UI / nginx

​​​​

Hydrosphere is a platform for deploying, versioning, and monitoring your machine learning models in production. It is language-agnostic and framework-agnostic, with support for all major programming languages and frameworks - Python, Java, Tensorflow, Pytorch, etc.

What to do next?

⭐️ Star Hydrosphere repo on Github

💦 Explore our Getting Started tutorial

🥳 Join Hydrosphere Slack Community

A Hydrosphere user can create a linear inference pipeline from multiple model versions. Such pipelines are called Applications.

A/B Deployment

Hydrosphere users can use multiple inside of the same stage. Hydrosphere shadows traffic to all model versions inside of an application stage.

Users can specify the likelihood that a model output will be selected as an application stage output by using the weight argument.

Traffic Shadowing

Hydrosphere shadows traffic to all model versions inside of an application stage.

If you want to shadow your traffic between model versions without producing output from them simply set weight parameter to 0. This way your model version will receive all incoming traffic, but its output will never be chosen as an output of an application stage.

Monitoring Dashboard lets you track your performance and get a high-level view of your data health.

Monitoring Dashboard plots all requests streaming through a model version which are colored in respect with how "healthy" they are. On the horizontal axis we group our data by batches and on the vertical axis we group data by signature fields. In this plot cells are determined by their batch and field. Cells are colored from green to red, depending on the average request health inside the batch.

Hydrosphere is a language-agnostic platform. You can use it with models written in any language and trained in any framework. Your ML models can come from any background, without restrictions of your choices regarding ML model development tools.

In Hydrosphere you operate ML models as Runtimes, which are Docker containers packed with predefined dependencies and gRPC interfaces for loading and serving them on the platform with a model inside. All models that you upload to Hydrosphere must have the corresponding runtimes.

Runtimes are created by building a Docker container with dependencies required for the language that matches your model. You can either use our pre-made runtimes or create your own runtime.

The Hydrosphere component responsible for building Docker images from models for deployment, storing them in the registry, versioning, and more is Manager.

For each model with uploaded training data, Hydrosphere creates an outlier detection (Auto OD) metric, which assigns an outlier score to each request. A request is labeled as an outlier if the outlier score is greater than the 97th percentile of training data outlier scores distribution.

You can observe those models deployed as metrics in your monitoring dashboard. These metrics provide you with information about how novel/anomalous your data is.

If these values of the metric deviate significantly from the average, you can tell that you experience a data drift and need to re-evaluate your ML pipeline to check for errors.

Supported Models

Right now Auto OD feature works only for Models with numerical scalar fields and uploaded training data.

Data Projection is a service that visualizes high-dimensional data in a 2D scatter plot with an automatically trained transformer to let you evaluate the data structure and spot clusters, outliers, novel data, or any other patterns. This is especially helpful if your model works with high-dimensional data, such as images or text embeddings.

Data Projection is an important tool, which helps to describe complex things in a simple way. One good visualization can show more than text or data. Monitoring and interpretation of machine learning models are hard tasks that require analyzing a lot of raw data: training data, production requests, as well as model outputs.

Essentially, this data is just numbers that in their original form of vectors and matrices do not have any meaning since it is hard to extract any meaning from thousands of vectors of numbers. In Hydrosphere we want to make monitoring easier and clearer that is why we created a data projection service that can visualize your data in a single plot.

Usage

To start working with Data Projection you need to create a model that has an output field with an embedding of your data. Embeddings are real-valued vectors that represent the input features in a lower dimensionality.

1. Create a model with an embedding field

   Data Projection service delegates the creation of embeddings to the user. It expects that model will create embedding from input features and pass it as output vector. Thus embedding field is required, models without this field are not supported. Data Projection also expects that output labels field is called class and model confidence is called respectively confidence. Other outputs are ignored.

2. Send data through your model

3. Check Data Projection service inside the Model Details menu

Each point in the plot presents a request. Requests with similar features are close to each other. You can select a specific request point and inspect what it consists of.

Above plot, there are several scores: global score, stability score, MSID score, etc. These scores reflect the quality of projection of multidimensional requests to 2D. To interpret scores you refer to technical documentation on Data Projection service.

In the Colorize menu, you can choose how to colorize model requests: by class, by monitoring metric or by confidence. Data Projection searchers specifically for output scalars class and confidence.

In the Accent Points menu, you can highlight the nearest in original space points to the selected one by picking the nearest variant. Counterfactuals will show you nearest points to selected but with a different predicted label.

Drift Report service creates a statistical report based on a comparison of training and production data distributions. It compares these two sets of data by a set of statistical tests and finds deviations.

Drift report uses multiple different tests with p=.95 for different features:

Numerical features:

• Levene's test with a trimmed mean

• Welch's t-test

• Mood's test

• Kolmogorov–Smirnov test

Categorical features:

• Chi-Square test

• Unseen categories

Supported Models

Right now Drift Report feature works only for Models with numerical scalar fields.

Hydrosphere Serving Components for Kubeflow Pipelines provide integration between Hydrosphere model serving benefits and Kubeflow orchestration capabilities. This allows launching training jobs as well as serving the same models in Kubernetes in a single pipeline.

You can find examples of sample pipelines here.

Serving components

Deploy

The Deploy component allows you to upload a model, trained in a Kubeflow pipelines workflow to a Hydrosphere platform.

For more information, check Hydrosphere Deploy Kubeflow Component

Release

The Release component allows you to create an Application from a model previously uploaded to Hydrosphere platform. This application will be capable of serving prediction requests by HTTP or gRPC.

For more information, check Hydrosphere Release Kubeflow Component

Prediction Explanation service is designed to help Hydrosphere users understand the underlying causes of changes in predictions coming from their models.

Prediction Explanation generates explanations of predictions produced by your models and tells you why a model made a particular prediction. Depending on the type of data your model uses, Prediction Explanation provides an explanation as either a set of logical predicates (if your data is in a tabular format) or a saliency map (if your data is in the image format). A saliency map is a heat map that highlights parts of a picture that a prediction was based on.

Contents

Overview

This section contains tutorials to help you get started with the Hydrosphere platform. A tutorial shows how to accomplish a goal rather than a single basic task.

Typically, a tutorial has several sections. When a tutorial section has several pieces of code to illustrate it, they can be shown as a group of tabs that you can switch between.

For guides on performing more basic technical steps, please look in the How-To section:

Overview

**** sends data about any failed health checks of live production models and applications to Prometheus AlertManager. Once a user deploys a model to production, adds training data and starts sending production requests, these requests start getting checked by Sonar. If Sonar detects an anomaly (for example, a data check failed, or a metric value exceeded the threshold), AlertManager sends an appropriate alert.

Users can manage alerts by setting up AlertManager for Prometheus on Kubernetes. This can be helpful when you have models that you get too many alerts from and need to filter, group, or partly silence them. AlertManager can take care of grouping, inhibition, silencing of alerts, and routing them to the receiver integration of your choice. To configure alerts, modify the prometheus-am-configmap-<release_name> ConfigMap.

For more information about Prometheus AlertManager please refer to its .

Overview

In this tutorial, you will learn how to configure deployed Applications.

By the end of this tutorial you will know how to:

• Train and upload an example model version

• Create a Deployment Configuration

• Create an Application from the uploaded model version with previously created deployment configuration

• Examine settings of a Kubernetes cluster

Prerequisites

• Hydrosphere platform installed in Kubernetes cluster

• Python SDK / CLI

Upload a Model

In this section, we describe the resources required to create and upload an example model used in further sections. If you have no prior experience with uploading models to the Hydrosphere platform we suggest that you visit the Getting Started Tutorial.

Here are the resources used to train sklearn.ensemble.GradientBoostingClassifier and upload it to the Hydrosphere cluster.

numpy~=1.18
scipy==1.4.1
scikit-learn~=0.23

kind: Model
name: my-model
runtime: hydrosphere/serving-runtime-python-3.7:2.3.2
install-command: pip install -r requirements.txt
payload:
  - src/
  - requirements.txt
  - model.joblib
contract:
  name: infer
  inputs:
    x:
      shape: [30]
      type: double
  outputs:
    y:
      shape: scalar
      type: int64

import joblib
import pandas as pd
from sklearn.datasets import make_blobs
from sklearn.ensemble import GradientBoostingClassifier

# initialize data
X, y = make_blobs(n_samples=3000, n_features=30)

# create a model
model = GradientBoostingClassifier(n_estimators=200)
model.fit(X, y)

# Save training data and model
pd.DataFrame(X).to_csv("training_data.csv", index=False)
joblib.dump(model, "model.joblib")

import joblib
import numpy as np

# Load model once
model = joblib.load("/model/files/model.joblib")


def infer(x):
    # Make a prediction
    y = model.predict(x[np.newaxis])

    # Return the scalar representation of y
    return {"y": np.asscalar(y)}

Our folder structure should look like this:

dep_config_tutorial
├── model.joblib
├── train.py
├── requirements.txt
├── serving.yaml
└── src
    └── func_main.py

After we have made sure that all files are placed correctly, we can upload the model to the Hydrosphere platform by running hs upload from the command line.

hs upload

Create a Deployment Configuration

Next, we are going to create and upload an instance of Deployment Configuration to the Hydrosphere platform.

Deployment Configurations describe with which Kubernetes settings Hydrosphere should deploy servables. You can specify Pod Affinity and Tolerations, the number of desired pods in deployment, ResourceRequirements, and Environment Variables for the model container, and HorizontalPodAutoScaler settings.

Created Deployment Configurations can be attached to Servables and Model Variants inside of Application.

Deployment Configurations are immutable and cannot be changed after they've been uploaded to the Hydrosphere platform.

You can create and upload Deployment Configuration to Hydrosphere via YAML Resource definition or via Python SDK.

For this tutorial, we'll create a deployment configuration with 2 initial pods per deployment, HPA, and FOO environment variable with value bar.

kind: DeploymentConfiguration
name: my-dep-config
deployment:
  replicaCount: 2
hpa:
  minReplicas: 2
  maxReplicas: 4
  cpuUtilization: 70
container:
  env:
    FOO: bar

hs apply -f deployment_configuration.yaml

from hydrosdk import Cluster, DeploymentConfigurationBuilder

cluster = Cluster("http://localhost")

dep_config_builder = DeploymentConfigurationBuilder("my-dep-config", cluster)
dep_config = dep_config_builder. \
    with_replicas(replica_count=2). \
    with_env({"FOO":"bar"}). \
    with_hpa(max_replicas=4,
             min_replicas=2,
             target_cpu_utilization_percentage=70).build()

Create an Application

kind: Application
name: my-app-with-config
pipeline:
  - - model: my-model:1
      weight: 100
      deploymentConfiguartion: my-config

hs apply -f application.yaml

from application import ApplicationBuilder, ExecutionStageBuilder
from hydrosdk import ModelVersion, Cluster, DeploymentConfiguration

cluster = Cluster('http:\\localhost')
my_model = ModelVersion.find(cluster, "my-model", 1)
my_config = DeploymentConfiguration.find(cluster, "my-config")

stage = ExecutionStageBuilder().with_model_variant(model_version=my_model,
                                                   weight=100,
                                                   deployment_configuration=my_config).build()

app = ApplicationBuilder(cluster, "my-app-with-config").with_stage(stage).build()

Examine Kubernetes Settings

Replicas

You can check whether with_replicas was successful by calling kubectl get deployment -A -o wide and checking the READYcolumn.

HPA

To check whether with_hpa was successful you should get a list of all created Horizontal Pod Autoscaler Resources. You can do so by calling kubectl get hpa -A

The output is similar to:

NAME                        REFERENCE                                            TARGETS    MINPODS   MAXPODS   REPLICAS   AGE
my-model-1-tumbling-star    CrossVersionObjectReference/my-model-1-tumbling-star 20%/70%    2         4         2          1d

Environment Variables

To list all environment variables run kubectl exec my-model-1-tumbling-star -it /bin/bash and then execute the printenv command which prints ann system variables.

The output is similar to:

MY_MODEL_1_TUMBLING_STAR_SERVICE_PORT_GRPC=9091
...
FOO=bar

Python SDK offers a simple and convenient way of integrating a user's workflow scripts with Hydrosphere API.

Source code: PyPI:

You can learn more about it in its documentation .

Installation

You can use pip to install hydrosdk

Usage

You can access the locally deployed Hydrosphere platform from previous by running the following code:

Overview

In this tutorial, you will learn how to retrospectively compare the behavior of two different models.

By the end of this tutorial you will know how to:

• Set up an A/B application

• Analyze production data

Prerequisites

Set Up an A/B Application

Prepare a model for uploading

Upload Model A

We train and upload our model with 5 components as movie_rec:v1

Upload Model B

Next, we train and upload a new version of our original model with 20 components as movie_rec:v2

We can check that we have multiple versions of our model by running:

Create an Application

The following code will create such an application:

Invoking movie-ab-app

We'll simulate production data flow by repeatedly asking our model for recommendations.

Analyze production data

Read Data from parquet

Each request-response pair is stored in S3 (or in minio if deployed locally) in parquet files. We'll use fastparquet package to read these files and use s3fs package to connect to S3.

The only file in the feature-lake folder is ['feature-lake/movie_rec']. Data stored in S3 is stored under the following path: feature-lake/MODEL_NAME/MODEL_VERSION/YEAR/MONTH/DAY/*.parquet

Now that we have loaded the data, we can start analyzing it.

Compare production data with new labeled data

To compare differences between model versions we'll use two metrics:

1. Latency - we compare the time delay between the request received and the response produced.

2. Mean Top-3 Hit Rate - we compare recommendations to those the user has rated. If they match then increase the hit rate by 1. Do this for the complete test set to get the hit rate.

Latencies

Let's calculate the 95th percentile of our latency distributions per model version and plot them. Latencies are stored in the _hs_latency column in our dataframes.

In our case, the output was 13.0ms against 12.0ms. Results may differ.

Furthermore, we can visualize our data. To plot latency distribution we'll use the Matplotlib library.

Mean Top-3 Hit Rate

Next, we'll calculate hit rates. To do so, we need new labeled data. For recommender systems, this data is usually available after a user has clicked\watched\liked\rated the item we've recommended to him. We'll use the test part of movielens as labeled data.

To measure how well our models were recommending movies we'll use a hit rate metric. It calculates how many movies users have watched and rated with 4 or 5 out of 3 movies recommended to him.

In our case the mean_hit_rate variable is {'v1': 0.137, 'v2': 0.141} . Which means that the second model version is better in terms of hit rate.

You have successfully completed the tutorial! 🚀

Now you know how to read and analyze automatically stored data.

Overview

In this tutorial, you will learn how to train and deploy a model for a classification task based on the . The main steps of this process are data preparation, training a model, uploading a model to the cluster, and making a prediction on test samples.

By the end of this tutorial you will know how to:

• Prepare data

• Train a model

• Deploy a model with SDK

• Explore models via UI

• Deploy a model with CLI and resource definition

Prerequisites

For this tutorial, you need to have Hydrosphere Platform deployed and Hydrosphere CLI (hs) along with Python SDK (hydrosdk) _**_installed on your local machine. If you don't have them yet, please follow these guides first:

For this tutorial, you can use a local cluster. To ensure that, run hs cluster in your terminal. This command will show the name and server address of a cluster you’re currently using. If it shows that you're not using a local cluster, you can configure one with the following commands:

Data preparation

Model training always requires some amount of initial preparation, most of which is data preparation. The Adult Dataset consists of 14 descriptors, 5 of which are numerical and 9 categorical, including the class column.

Categorical features are usually presented as strings. This is not an appropriate data type for sending it into a model, so we need to transform it first. We can remove rows that contain question marks in some samples. Once the preprocessing is complete, you can delete the DataFrame (df):

Training a model

There are many classifiers that you can potentially use for this step. In this example, we’ll apply the Random Forest classifier. After preprocessing, the dataset will be separated into train and test subsets. The test set will be used to check whether our deployed model can process requests on the cluster. After the training step, we can save a model with joblib.dump() in a model/ model folder.

Deploy a model with SDK

The code in the func_main.py should be as follows:

It’s important to make sure that variables will be in the right order after we transform our dictionary for a prediction. So in cols we preserve column names as a list sorted by order of their appearance in the DataFrame.

To start working with the model in a cluster, we need to install the necessary libraries used in func_main.py. Create a requirements.txt in the folder with your model and add the following libraries to it:

After this, your model directory with all necessary dependencies should look as follows:

Now we are ready to upload our model to the cluster.

Use X.dtypes to check what types of data you have for each column. You can use int64 fields for all variables including income, which is our dependent variable and we can name it as 'y' in a signature for further prediction.

Besides, you can specify the type of profiling for each variable using ProfilingType so Hydrosphere could know what this variable is about and analyze it accordingly. For this purpose, we can create a dictionary, which could contain keys as our variables and values as our profiling types. Otherwise, you can describe them one by one as a parameter in the input.

Finally, we can complete our signature with the .build() method.

Next, we need to specify which files will be uploaded to the cluster. We use path to define the root model folder and payload to point out paths to all files that we need to upload.

At this point, we can combine all our efforts into the LocalModel object. LocalModels are models before they get uploaded to the cluster. They contain all the information required to instantiate a ModelVersion in a Hydrosphere cluster. We’ll name this model adult_model.

Now we are ready to upload our model to the cluster. This process consists of several steps:

1. Once LocalModel is prepared we can apply the upload method to upload it.

2. Then we can lock any interaction with the model until it will be successfully uploaded.

3. ModelVersion helps to check whether our model was successfully uploaded to the platform by looking for it.

Predictors provide a predict method which we can use to send our data to the model. We can try to make predictions for our test set that has preliminarily been converted to a list of dictionaries. You can check the results using the name you have used for an output of Signature and preserve it in any format you would prefer. Before making a prediction don't forget to make a small pause to finish all necessary loadings.

Explore the UI

If you want to interact with your model via Hydrosphere UI, you can go to http://localhost. Here you can find all your models. Click on a model to view information about it: versions, building logs, created applications, model's environments, and other services associated with deployed models.

🎉 You have successfully finished this tutorial! 🎉

Next Steps

Next, you can:

1. Go to the next tutorial and learn how to create a custom Monitoring Metric and attach it to your deployed model:

1. Explore the extended part of this tutorial to learn how to use YAML resource definitions to upload a ModelVersion and create an Application.

Deploy a model with CLI and Resource Definitions

Model deployment with a resource definition repeats all the steps of that with SDK, but in one file. A considerable advantage of using a resource definition is that besides describing your model it allows creating an application by simply adding an object to the contract after the separation line at the bottom. Just name your application and provide the name and version of a model you want to tie to it.

To start uploading, run hs apply -f serving.yaml. To monitor your model you can use Hydrosphere UI as was previously shown.

Hydrosphere CLI, orhs, is a command-line interface designed to work with the Hydrosphere platform.

Source code: PyPI:

Installation

Use pip to install hs:

Check the installation:

Usage

hs cluster

This command lets you operate cluster instances. A cluster points to your Hydrosphere instance. You can use this command to work with different Hydrosphere instances.

See hs cluster --help for more information.

hs upload

This command lets you upload models to the Hydrosphere platform. During the upload, hs looks for a serving.yaml file in the current directory. This file must contain a definition of the model ().

See hs upload --help for more information.

hs apply

This command is an extended version of the hs upload command, which also allows you to operate applications and host selector resources.

See hs apply --help for more information.

hs profile

This command lets you upload your training data to build profiles.

• $ hs profile push - upload training data to compute its profiles.

• $ hs profile status - show profiling status for a given model.

See hs profile --help for more information.

hs app

This command provides information about available applications.

• $ hs app list - list all existing applications.

• $ hs app rm - remove a certain application.

See hs app --help - for more information.

hs model

This command provides information about available models.

• $ hs model list - list all existing models.

• $ hs model rm - remove a certain model.

See hs model --help for more information.

Overview

Monitoring can be used to track the behavior of external models running outside of the Hydrosphere platform. This tutorial describes how to register an external model, trigger analysis over your requests, and retrieve results.

By the end of this tutorial you will know how to:

• Register a model

• Upload training data

• Assign custom metrics

• Invoke analysis

• Retrieve metrics

Prerequisites

For this tutorial, you need to have Hydrosphere Platform deployed on your local machine with Sonar component enabled. If you don't have it yet, please follow this guide first:

You also need a running external model, capable of producing predictions. Inputs and outputs of that model will be fed into Hydrosphere for monitoring purposes.

Model registration

First, you have to register an external model. To do that, submit a JSON document, defining your model.

Request document structure

This section describes the structure of the JSON document used to register external models within the platform.

Top-level members

The document must contain the following top-level members, describing the interface of your model:

• name: the name of the registered model. This name uniquely identifies a collection of model versions, registered within the Hydrosphere platform.

• contract: the interface of the registered model. This member describes inputs and outputs of the model, as well as other complementary metadata, such as model signature, and data profile for each field.

A document may contain additional top-level members, describing other details of your model.

• metadata: the metadata of the registered model. The structure of the object is not strictly defined. The only constraint is that the object must have a key-value structure, where a value can only be of a simple data type (string, number, boolean).

• monitoringConfiguration: monitoring configuration to be used for this model.

This example shows, how a model can be defined at the top level:

MonitoringConfiguration object

monitoringConfiguration object defines a monitoring configuration to be used for the model version. The object must contain the following members:

• batchSize: size of the batch to be used for aggregations.

The example below shows how a monitoringConfiguration object can be defined.

Contract object

Thecontract object appears in the document to define the interface of the model. The contract object must contain the following members:

• modelName: the original name of the model. It should be the same as the name of the registered model, defined on the level above;

• predict: the signature of the model. It defines the inputs and the outputs of the model.

The example below shows how a contract object can be defined.

Predict object

predict object describes the signature of the model. The signature object must contain the following members:

• signatureName: The signature of the model, used to process the request;

• inputs: A collection of fields, defining the inputs of the model. Each item in the collection describes a single data entry, its type, shape, and profile. A collection must contain at least one item;

• outputs: A collection of fields, defining the outputs of the model. Each item in the collection describes a single data entry, its type, shape, and profile. A collection must contain at least one item.

The example below shows how a predict object can be defined.

Field object

Items in the inputs / outputs collections are collectively called "fields". The field object must contain the following members:

• name: Name of the field;

• dtype: Data type of the field.

• profile: Data profile of the field.

• shape: Shape of the field.

The only valid options for dtype are:

• DT_STRING;

• DT_BOOL;

• DT_VARIANT;

• DT_HALF;

• DT_FLOAT;

• DT_DOUBLE;

• DT_INT8;

• DT_INT16;

• DT_INT32;

• DT_INT64;

• DT_UINT8;

• DT_UINT16;

• DT_UINT32;

• DT_UINT64;

• DT_QINT8;

• DT_QINT16;

• DT_QINT32;

• DT_QUINT8;

• DT_QUINT16;

• DT_COMPLEX64;

• DT_COMPLEX128;

The only valid options for profile are:

• NONE

• NUMERICAL

• TEXT

• IMAGE

The example below shows how a single field object can be defined.

Shape object

shape object defines the shape of the data that the model is processing. The shape object must contain the following members:

• dim: A collection of items, describing each dimension. A collection may be empty — in that case, the tensor will be interpreted as a scalar value.

• unknownRank: Boolean value. Identifies whether the defined shape is of unknown rank.

The example below shows how a shape object can be defined.

Dim object

dim object defines a dimension of the field. The dim object must contain the following members:

• size: Size of the dimension.

• name: Name of the dimension.

The example below shows how a dim object can be defined.

Registering external model

A model can be registered by sending a POST request to the /api/v2/externalmodel endpoint. The request must include a model definition as primary data.

The request below shows an example of an external model registration.

As a response, the server will return a JSON object with complementary metadata, identifying a registered model version.

Response document structure

The response object from the external model registration request contains the following fields:

• id: Model version ID, uniquely identifying a registered model version within Hydrosphere platform;

• model: An object, representing a model collection, registered in Hydrosphere platform;

• modelVersion: Model version number in the model collection;

• modelContract: Contract of the model, similar to the one defined in the request section above;

• metadata: Metadata of the model, similar to the one defined in the request section above;

• monitoringConfiguration: MonitoringConfiguration of the model, similar to the one defined in the request section above;

• created: Timestamp, indicating when the model was registered.

Model object

model object represents a collection of model versions, registered in the platform. The response model object contains the following fields:

• id: ID of the model collection;

• name: Name of the model collection.

The example below shows, a sample server response from an external model registration request.

Training data upload

To let Hydrosphere calculate the metrics of your requests, you have to submit the training data. You can do so by:

Upload using CLI

Switch to the cluster, suitable for your current flow.

If you don't have a defined cluster yet, create one using the following command.

Make sure you have a local copy of the training data that you want to submit.

Submit the training data. You must specify two parameters:

• --model-version: A string indicating the model version to which you want to submit the data. The string should be formatted in the following way <model-name>:<model-version>;

• --filename: Path to a filename, that you want to submit.

Depending on the size of your data, you will have to wait for the data to be uploaded. If you don't want to wait, you can use the --async flag.

Upload using an HTTP endpoint

To upload your data using an HTTP endpoint, stream it to the /monitoring/profiles/batch/<MODEL_VERSION_ID> endpoint.

In the code snippets below you can see how data can be uploaded using sample HTTP clients.

Custom metrics assignment

This step is optional. If you wish to assign a custom monitoring metric to a model, you can do it by:

• using Hydrosphere UI

• using HTTP endpoint

Using Hydrosphere UI

Using HTTP endpoint

To assign metrics using HTTP endpoint, you will have to submit a JSON document, defining a monitoring specification.

Top-level members

The document must contain the following top-level members.

• name: The name of the monitoring metric;

• modelVersionId: Unique identifier of the model to which you want to assign a metric;

• config: Object, representing a configuration of the metric, which will be applied to the model.

The example below shows how a metric can be defined on the top level.

Config object

config object defines a configuration of the monitoring metric that will monitor the model. The model must contain the following members:

• modelVersionId: Unique identifier of the model that will monitor requests;

• threshold: Threshold value, against which monitoring values will be compared using a comparison operator;

• thresholdCmpOperator: Object, representing a comparison operator.

The example below shows, how a metric can be defined on a top-level.

ThresholdCmpOperator object

thresholdCmpOperator object defines the kind of comparison operator that will be used when comparing a value produced by the metric against the threshold. The object must contain the following members:

• kind: Kind of comparison operator.

The only valid options for kind are:

• Eq;

• NotEq;

• Greater;

• Less;

• GreaterEq;

• LessEq.

The example below shows, how a metric can be defined on the top level.

The request below shows an example of assigning a monitoring metric. At this moment, both monitoring and the actual prediction model should be registered/uploaded to the platform.

Analysis invocation

In the code snippets below you can see how analysis can be triggered with sample gRPC clients.

Metrics retrieval

A request must contain the following parameters:

• limit: how many requests to fetch;

• offset: which offset to make from the beginning.

An example request is shown below.

Calculated metrics have a dynamic structure, which is dependant on the model interface.

Response object structure

A response object contains the original data submitted for prediction, the model's response, calculated metrics and other supplementary metadata. Every field produced by Hydrosphere is prefixed with _hs_ char.

• _id: ID of the request, generated internally by Hydrosphere;

• _hs_request_id: ID of the request, specified by user;

• _hs_model_name: Name of the model that processed a request;

• _hs_model_incremental_version: Version of the model that processed a request;

• _hs_model_version_id: ID of the model version, which processed a request;

• _hs_raw_checks: Raw checks calculated by Hydrosphere based on the training data;

• _hs_metric_checks: Metrics produced by monitoring models;

• _hs_latency: Latency, indicating how much it took to process a request;

• _hs_error: Error message that occurred during request processing;

• _hs_score: The number of all successful checks divided by the number of all checks;

• _hs_overall_score: The amount of all successful metric values (not exceeding a specified threshold), divided by the amount of all metric values;

• _hs_timestamp: Timestamp in nanoseconds, when the object was generated;

• _hs_year: Year when the object was generated;

• _hs_month: Month when the object was generated;

• _hs_day: Day when the object was generated;

Apart from the fields defined above, each object will have additional fields specific to the particular model version and its interface.

• _hs_<field_name>_score: The number of all successful checks calculated for this specific field divided by the total number of all checks calculated for this specific field;

• <field_name>: The value of the field.

Raw checks object

_hs_raw_checks object contains all fields, for which checks have been calculated.

The example below shows, how the _hs_raw_checks_ object can be defined.

Check object

check object declares the check, that has been calculated for the particular field. The following members will be present in the object.

• check: Boolean value indicating, whether the check has been passed;

• description: Description of the check that has been calculated;

• threshold: Threshold of the check;

• value: Value of the field;

• metricSpecId: Metric specification ID. For each check object this value will be set to null.

The example below shows, how the check object can be defined.

Metrics object

_hs_metrics_checks object contains all fields for which metrics have been calculated.

The example below shows how the _hs_metrics_checks object can be defined.

Metric object

metric object declares the metric, that has been calculated for the particular field. The following members will be present in the object.

• check: Boolean value indicating, whether the metric has not been fired;

• description: Name of the metric that has been calculated;

• threshold: Threshold of the metric;

• value: Value of the metric;

• metricSpecId: Metric specification ID.

The example below shows how the metric object can be defined.

The example below shows a fully composed server response.

Overview

In this tutorial, you will learn how to create a custom anomaly detection metric for a specific use case.

Let's take a problem described in the previous tutorial as a use case and as a data source. We will monitor a model that classifies whether the income of a given person exceeds $50.000 per year.

By the end of this tutorial you will know how to:

• Train a monitoring model

• Deploy of a monitoring model with SDK

• Manage сustom metrics with UI

• Upload a monitoring model with CLI

Prerequisites

For this tutorial, you need to have Hydrosphere Platform deployed and Hydrosphere CLI (hs) along with Python SDK (hydrosdk) _**_installed on your local machine. If you don't have them yet, please follow these guides first:

This tutorial is a sequel to the previous tutorial. Please complete it first to have a prepared dataset and a trained model deployed to the cluster:

Train a Monitoring Model

We start with the steps we used for the common model. First, let's create a directory structure for our monitoring model with an /src folder containing an inference scriptfunc_main.py:

To make sure that our monitoring model will see the same data as our prediction model, we are going to apply the training data that was saved previously for our monitoring model.

This is what the distribution of our inliers looks like. By choosing a contamination parameter we can adjust a threshold that will separate inliers from outliers accordingly. You have to be thorough in choosing it to avoid critical prediction mistakes. Otherwise, you can also stay with 'auto'. To create a monitoring metric, we have to deploy that IsolationForest model as a separate model on the Hydrosphere platform. Let's save a trained model for serving.

Deploy a Monitoring Model with SDK

First, let's create a new directory where we will store our inference script with declared serving function and its definitions. Put the following code inside the src/func_main.py file:

Next, we need to install the necessary libraries. Create a requirements.txt and add the following libraries to it:

Just like with common models, we can use SDK to upload our monitoring model and bind it to the trained one. The steps are almost the same, but with some slight differences. First, since we want to predict the anomaly score instead of sample class, we need to change the type of output field from 'int64' to 'float64'.

Secondly, we need to apply a couple of new methods to create a metric. MetricSpec is responsible for creating a metric for a specific model, with specific MetricSpecConfig.

Managing Custom Metrics with UI

Go to the UI to observe and manage all your models. Here you will find 3 models on the left panel:

• adult_monitoring_model - our monitoring model

• adult_model_metric - a model that was created by Automatic Outlier Detection

During the prediction, you will get anomaly scores for each sample in the form of a chart with two lines. The curved line shows scores, while the horizontal dotted one is our threshold. When the curve intersects the threshold, it might be a sign of potential anomalousness. However, this is not always the case, since there are many factors that might affect this, so be careful about your final interpretation.

Uploading a Monitoring model with CLI

Just like in the case with all other types of models, we can define and upload a monitoring model using a resource definition. We have to pack our model with a model definition, like in the previous tutorial.

Inputs of this model are the inputs of the target monitored model plus the outputs of that model. We will use the value field as an output for the monitoring model. The final directory structure should look like this:

From that folder, upload the model to the cluster:

Now we have to attach the deployed Monitoring model as a custom metric. Let's create a monitoring metric for our pre-deployed classification model in the UI:

That's it. Now you have a monitored income classifier deployed on the Hydrosphere platform.

Overview

In this tutorial, you will learn the basics of working with Hydrosphere. We will prepare an example model for serving, deploy it to Hydrosphere, turn it into an application, invoke it locally, and use monitoring. As an example model, we will take a simple logistic regression model fit with randomly generated data, with some noise added to it.

By the end of this tutorial you will know how to:

• Prepare a model for Hydrosphere

• Serve a model on Hydrosphere

• Create an Application

• Invoke an Application

• Use basic monitoring

Prerequisites

For this tutorial, you need to have Hydrosphere Platform deployed and Hydrosphere CLI (hs) along with Python SDK (hydrosdk) _**_installed on your local machine. If you don't have them yet, please follow these guides first:

To let hs know where the Hydrosphere platform runs, configure a new cluster entity:

Before you start

In the next two sections, we will prepare a model for deployment to Hydrosphere. It is important to stick to a specific folder structure during this process to let hs parse and upload the model correctly. Make sure that the structure of your local model directory looks like this by the end of the model preparation section:

• train.py - a training script for our model

• requirements.txt - provides dependencies for our model

• model.joblib - a model artifact that we get as a result of model training

• src/func_main.py - an inference script that defines a function for making model predictions

• serving.yaml - a resource definition file to let Hydrosphere know which function to call from the func_main.py script and let the model manager understand model’s inputs and outputs.

Training a model

While Hydrosphere is a post-training platform, let's start with basic training steps to have a shared context.

First, create a directory for your model and create a new train.py inside:

Put the following code for your model in the train.py file:

Next, we need to install all the necessary libraries for our model. In your logistic_regression folder, create a requirements.txt file and provide dependencies inside:

Install all the dependencies to your local environment:

Train the model:

As soon as the script finishes, you will get the model saved to a model.joblib file.

Model preparation

Every model in the Hydrosphere cluster is deployed as an individual container. After a request is sent from the client application, it is passed to the appropriate Docker container with your model deployed on it. An important detail is that all model files are stored in the /model/files directory inside the container, so we will look there to load the model.

To run our model we will use a Python runtime that can execute any Python code you provide. Model preparation is pretty straightforward, but you have to create a specific folder structure described in the "Before you start" section.

Provide the inference script

Let's create the main file func_main.pyin the /src folder of your model directory:

To do inference you have to define a function that will be invoked every time Hydrosphere handles a request and passes it to the model. Inside that function, you have to call a predict (or similar) method of your model and return your predictions:

Inside func_main.py we initialize our model outside of the serving function infer.This process will not be triggered every time a new request comes in.

The infer function takes the actual request, unpacks it, makes a prediction, packs the answer, and returns it. There is no strict rule for naming this function, it just has to be a valid Python function name.

Provide a resource definition file

To let Hydrosphere know which function to call from the func_main.py file, we have to provide a resource definition file. This file will define a function to be called, inputs and outputs of a model, a signature function, and some other metadata required for serving.

Create a resource definition file serving.yamlin the root of your model directorylogistic_regression:

Inside serving.yaml we also providerequirements.txt andmodel.joblib as payload files to our model:

At this point make sure that the overall structure of your local model directory looks as shown in the "Before you start" section.

Serving a Model

Now we are ready to upload our model to Hydrosphere. To do so, inside the logistic_regression model directory run:

If you cannot find your newly uploaded model and it is listed on your models' page, it is probably still in the building stage. Wait until the model changes its status to Released, then you can use it.

Creating an Application

Invoking an application

Invoking applications is available via different interfaces. For this tutorial, we will cover calling the created Application by gRPC via our Python SDK.

To install SDK run:

Define a gRPC client on your side and make a call from it:

Getting Started with Monitoring

Hydrosphere Platform has multiple tools for data drift monitoring:

1. Data Drift Report

2. Automatic Outlier Detection

3. Profiling

In this tutorial, we'll look at the monitoring dashboard and Automatic Outlier Detection feature.

Provide training data

To provide training data users need to add the training-data=<path_to_csv> field to the serving.yaml file. Run the following script to save training data used in previous steps as a trainig_data.csv file:

Next, add the training data field to the model definition inside the serving.yaml file:

Upload a model

Now we are ready to upload our model. Run the following command to create a new version of the logistic_regresion model:

For each model with uploaded training data, Hydrosphere creates an outlier detection metric, which assigns an outlier score to each request. This metric labels a request as an outlier if the outlier score is greater than the 97th percentile of training data outlier scores distribution.

Update an Application

Let's send some data to our new model version. To do so, we need to update our logistic_regression application. To update it, we can go to the Application tab and click the "Update" button:

Send data to Application

After updating our Application, we can reuse our old code to send some data:

Monitor data quality

You can monitor your data quality in the Monitoring Dashboard:

The Monitoring dashboard plots all requests streaming through a model version as rectangles colored according to how "healthy" they are. On the horizontal axis, we group our data by batches and on the vertical axis, we group data by signature fields. In this plot, cells are determined by their batch and field. Cells are colored from green to red, depending on the average request health inside this batch.

Check data drift detection

To check whether our metric will be able to detect data drifts, let's simulate one and send data from another distribution. To do so, let's slightly modify our code:

You can validate that your model was able to detect data drifts on the monitoring dashboard.