azure-enterprise-scale-ml (ESML)

Non-breaking MLOps pipelines since 2019

Enterprise Scale ML (ESML) - AI Factory on Azure

• A solution accelerator, for Enterprise Scale Machine Learning & MLOps, based on best & proven practices for organizational scale, across projects.
  • Best practice: CAF/AI Factory: https://docs.microsoft.com/en-us/azure/cloud-adoption-framework/ready/azure-best-practices/ai-machine-learning-mlops#mlops-at-organizational-scale-ai-factories
  • Best practice: Microsoft Intelligent Data Platform: https://techcommunity.microsoft.com/t5/azure-data-blog/microsoft-and-databricks-deepen-partnership-for-modern-cloud/ba-p/3640280
    • Modern data architecture with Azure Databricks and Azure Machine Learning: https://docs.microsoft.com/en-us/azure/architecture/solution-ideas/articles/azure-databricks-modern-analytics-architecture
  • Best practice: Datalake design: https://docs.microsoft.com/en-us/azure/storage/blobs/data-lake-storage-best-practices
    • Datamesh: https://martinfowler.com/articles/data-mesh-principles.html
      • Credit to: Zhamak Dehghani
• ESML has a default scaling from 1-250 ESMLprojects for its ESML AI Factory.
  • That said, the scaling roof is on IP-plan, and ESML has its own IP-calculator (allocated IP-ranges for 250 is just the default)
• Enterprise "cockpit" over ALL your projects & models.
  • See what state a project are in (Dev,Test,Prod states) with cost dashboard per project/environment

NEWS TABLE

Date	Category	What	Link
2023-03	Networking	No Public IP: Virtual private cloud - updated networking rules	https://learn.microsoft.com/en-us/azure/machine-learning/v1/how-to-secure-workspace-vnet?view=azureml-api-1&preserve-view=true&tabs=required%2Cpe%2Ccli
2023-02	ESML Pipeline templates	Azure Databricks: Training and Batch inference pipeline templates updated to have 100% same support as Azure ML pipeline templates, ESML inner/outer loop	-
2022-08	ESML infra (IaC)	Bicep now support yaml as well	-
2022-10	ESML MLOps	ESML MLOps v3 engine has advanced mode, with added support for Spark pipelines ( Databricks notebooks as Azure ML pipeline steps )	-

TEMPLATES for PIPELINES (TRAINING & INFERENCE pipelines) is 1 of 5 template types in ESML:

THE Challenge

Innovating with AI and Machine Learning, multiple voices expressed the need to have an Enterprise Scale AI & Machine Learning Platform with end-2-end turnkey DataOps and MLOps. Other requirements were to have an enterprise datalake design, able to share refined data across the organization, and high security and robustness: General available technology only, vNet support for pipelines & data with private endpoints. A secure platform, with a factory approach to build models.

Even if best practices exists, it can be time consuming and complex to setup such a AI Factory solution, and when designing an analytical solution a private solution without public internet is often desired since working with productional data from day one is common, e.g. already in the R&D phase. Cyber security around this is important.

• Challenge 1: Marry multiple, 4, best practices
• Challenge 2: Dev, Test, Prod Azure environments/Azure subscriptions
• Challenge 3: Turnkey: Datalake, DataOps, INNER & OUTER LOOP MLOps Also, the full solution should be able to be provisioned 100% via infrastructure-as-code, to be recreated and scale across multiple Azure subscriptions, and project-based to scale up to 250 projects - all with their own set of services such as their own Azure machine learning workspace & compute clusters.

THE Strategy

To meet the requirements & challenge, multiple best practices needed to be married and implemented, such as: CAF/WAF, MLOps, Datalake design, AI Factory, Microsoft Intelligent Data Platform / Modern Data Architecture. An open source initiative could help all at once, this open-source accelerator Enterprise Scale ML(ESML) - to get an AI Factory on Azure

THE Solution - TEMPLATES & Accelerator

ESML provides an AI Factory quicker (within 4-40 hours), with 1-250 ESMLProjects, an ESML Project is a set of Azure services glued together securely.

• Challenge 1 solved: Marry multiple, 4, best practices
• Challenge 2 solved: Dev, Test, Prod Azure environments/Azure subscriptions
• Challenge 3 solved: Turnkey: Datalake, DataOps, INNER & OUTER LOOP MLOps ESML marries multiple best practices into one solution accelerator, with 100% infrastructure-as-code

ESML 4 main components:

ESML AI Factory - 4 step process:

ESML AI Factory "Oneslider": Dev,Test,Prod environments - Enterprise Scale LandingZones

• Easy to provision a new ESMLProject for Dev,Test,Prod with easy cost followup, since its own PROJECT resource groups for each Project team in the ESML AI Factory:
• Horisontally 3 COMMON environment (Dev,Test, Prod) and vertically ESMLProject 1-250

The Azure Devops/BICEP can optionally integrate with ITSM system as a "ticket" in ServiceNow/Remedy/JIRA Service Desk. The below info is needed for the ESML provisioning:

ESML Architecture - "Modern data analytics platform"

Based on this reference architecture: https://docs.microsoft.com/en-us/azure/architecture/solution-ideas/articles/azure-databricks-modern-analytics-architecture

Q: Is this for you? DataOps married with MLOps? Whats the benefits of the ESML Controlplane SDK?

• ESML "marries" DataOps + MLOps, with templates for both Azure Data factory and Azure machine learning pipeline templates - ESML autogenereated Azure ML Pipelines
• ESML Auto-provisions the AI Factory, with 100% BICEP, where you can 1-click a new ESMLProject in Azure Devops, serviecs glued together with private endpoints (network & identity)
• ESML has MLOps: both INNER and OUTER LOOP (can talk across Dev,Test, Prod Azure ML workspaces)
• ESML controlplane can compare scoring from model in DEV workspace with TEST workspace, and register the model in an external workspace (this with also network security: vNets & private endpoints, NSG's, FW)
  • See image:

Contributing to ESML?

This repository is a push-only mirror. Pull requests are ignored. See LICEENCE file (open source, MIT license)

• Ping Joakim Åström for contributions / ideas. Speaking of: Credit to Kim Berg and Ben Kooijman for contributing! (kudos to the ESML IP calculator and Bicep for ESML Project)

INTRO - Is this for you: refine data? AutoML or manual ML? R&D phase?

Q1:I want to use Azure AutoML, with MLOps ready to be turned ON , with datalake design automatically generated for me, including BRONZE, SILVER, GOLD concept

• A: Yes. ESML is AutoML first, and have married this with MLOps, and an AutoLake™ for Azure ML Studio.
• There is 22 DEMO notebooks End-2-End MLOps, with Azure ML Pipelines, using Azure datalake GEN 2 all the way - from Azure datafactory, in Azure ML Pipelines/Datasets.

Q2:I want to do ML, but only R&D phase - I don't need MLOps or DEV,TEST, PROD environments. Can I still get benefits of ESML - get a quick DEV env & AutoLake?

• A: Yes. ESML is meant for quick R&D ( and if successful PoC -> quickly turn ON, to full enterprise scale MLOps solution
  • Quick setup: You can setup ESML for 1 environment only (have same subscriptionID for all 3).
    • Copy the settings & notebook_demo folder (but no need to copy MLOPS folder)
  • R&D Mode: Run ESML SDK with ESMLProject.rnd=True, and dataset-versioning will be turned off, but you still get a AutoLake with bronze, silver, gold concept.

Q3:I want to do ML, but NOT AutoML - just scikit learn, my own model. Can I still leverage ESML, besides training step?

• A: Yes, you can wrap your TRAIN-step code, in an Azure ML Pipeline, as a Python step, or Databricks step, and still leverage the AutoLake and other ESML accelerators.
  • You have multiple options for your steps in this pipeline, besides automl_step, you have python_script_step, Databricks_script_step, estimator_step, synapse_spark_step, ...
  • Full list: https://docs.microsoft.com/en-us/python/api/azureml-pipeline-steps/azureml.pipeline.steps?view=azure-ml-py
• A: With that said, ESML has a AutoML first approach.
  • Using this accelerates more, and enables easier & cheaper governance (unattended retraining with auto-hyperparameter tuning)

Q: How was this accelerator born, and what is it based on? It this for me?

• A:Working with multiple enterprise customers (aviation, manufacturing, space, energy and retail industry), we noticed common non-industry-specific challenges, to scale across projects, that ESML solves - an organizational scalability.

  • ESML extends Azure Machine Learning via accelerators, organizational agnostic - since the project/teams concept in ESML.
  • It extends at specific purposes: data refinement/datalake/machine learning to build faster.
  • Also adds enterprise grade solution design & scalability (dev,test, prod environments) - across subscriptions.
  • `An Enterprise Datalake, with ADLS GEN2, and logical evendriven DataMesh, with private endpoints security

Note: You can use this for any enterprise grade solution in need of single or multi-subscription solutions, with an enterprise datalake need, DEV only need, or DEV->TEST->PROD need.

• ESML was born out of these needs. Based on both Microsoft and open source best practices and customer proven practices
• Disclaimer:Although this have accelerated others, there is no guarantee it will accelerate your situtation. Read the MIT LICENCE file & Happy coding.

Q6 ESML AI Factory: Can I just use the Azure ML SDK directly? Instead of the ESML SDK?

• Yes, You can bypass ESML SDK 100% (the 5th ingredience) and only take advantage of the other ingredients (See ) the templates

EDUCATION & Prerequisites (good to have in the backpack)

• Learning by doing is probably the best thing, but below some Azure certificates are listed good to have in the backpack
• DP 100 https://docs.microsoft.com/en-us/learn/certifications/exams/dp-100
• There are also 6 ESML videos (in editing room), about ~1h each, hopefully up soon.

ESML Accelerator benefits

ESML has MLOps embedded, and adds NEW concepts to enrich Azure ML Studio:

• EMSL enables enterprise CONCEPTS (Project/Model/Dev_Test_Prod)` - able to scale across Azure subscriptions in DEV, TEST, PROD for a model.
• ESML includes accelerators for data refinement, with CONCEPTS: Bronze, Silver, Gold, able to share refined data ACROSS projects & models
• ESML Pipeline factory automatically generates Azure ML pipelines of 7 types, with the data model IN->Bronze->Silver-Gold (we will refer to this as IN_2_GOLD)
• ESML includes efficiency accelerators for ML CONCEPTS such as SCORE vs INFERENCE,ESMLPipelieFactory (auto-creates pipeline), Auto-Split to TRAIN,VALIDATE, TEST (auto-register).
• ESML marries MLOps with AutoML - you get working MLOps template with support for Azure AutoML.
• You don't need to remember folder paths - since the ESML Datalake design and automapping of Azure ML Datasets, if you work with the ESML SDK (Python, Pyspark)

• These datasets are automapped/autogenerated by ESML at p.split_to_gold()
• Same thing at feature engineering, at esmldataset.Bronze.Save(dataframe_state) - the Bronze dataset will be created, and a new version (if not p.rnd=True) is created for you.

ESMLPipelineFactory

• This scoring pipeline is automatically ESML-generated, via only 2 lines of code!! (This is possible due to the 4 ingrediences in ESML)
• If you have your data in IN in "GOLD" state, it will work as-is, but probably : ) you want to add your data wrangling per IN_TO_SILVER step, in the 1-M auto-generated ds_name_by_config.py scripts

Azure ML is great, it improves pipeline creation with 90% fewer lines of code to https://azure.microsoft.com/en-us/services/machine-learning/#features

I love when I get asked to push the boundries, and asks where dropping in from multiple places:

• Q: Azure Machine learning is great, but can ESML accelerate that even more? to make it even easier, less code, to create Azure mahcine learning pipelines?
• A: Hm, lets try. Below is the result: (0.1% of the code, of the already 90% acceleration, to get the same batch scoring pipeline)
  • Currently in ESML there are 7 Azure ML pipeline types, that can be generated in the same manner - with 2 lines of code, for scoring, retraining, or just refine data, etc.

WHAT is ESML Autolake™ ( Azure Datalake Storage GEN2 accelerator)

• It is based on Azure Datalake Storage GEN2, but includes a turnkey lake-design "skeleton" with concepts for ML (train, inference) and data refinement (Bronze, Silver, Gold), and enterprise scale concepts (incremental load, versioning, dev/test/prod).
• It also has MASTER vs PROJECT concept, able to support both DeltaLake on Azure datalake GEN2 and Azure ML pipelines with Azure Datalake GEN 2 Datastore.
• It is also automated for Azure ML Studio, to automatically register data as Azure ML Studio Datasets
  • Connected "per project & model". You see only your projects data.
• And it contains automated enterprise security ( uses Azure keyvault for secrets for you etc)

Automapping Data to Azure ML Datasets - only possible due to the ESML datalake

MLOps - Example: compare model in DEV subscription with TEST subscription

TEST_SET Scoring to Azure ML Studio, as TAGS

• You can use ESML to automatically calculate TEST-SET scoring, 1 line of code (works for classification or regression), and this will be TAGGED on the Azure ML Dataset GOLD_TEST and also on the Model
• 
• 

Scoring Drift / Concept Drift to promote newly trained model (also as step in ESML MLOps pipeline)

• We can adjust WEIGHTS, and definition of what a BETTER model is, scoring wise.
• Example below, the newly trained model in DEV SCORED worse, than TARGET model in TEST environment, Promote=False.
• 

Settings

• Besides there green circles, you have a config per environment (dev,test,prod) for COMPUTE power & and HYPERPARAMETER tuning needed (e.g. in DEV you might wanna have cheaper training runs)
  • Defaults enteprise settings (dev,test,prod), is usually set & decided once, by an enterprise architect, and all ESML Projects inherits these, but can override them also, if use case needs that.
    • To override default enteprise settings, a projects sets the project specific settings.

Project/Model settings

The BEST Model - according to YOU": Model_settings

• What: YOU define what "the best model" is. When ESML are comparing and promoting models, its based on YOUR Model settings with your weights that will decide last registered model = best
  • And, you can alwauys override this, to register an model of your choice manually - it'll become "the latest" and hence "the best" in the eyes of the ESML (when include in deployemnt for scoring pipleline )
• Purpose: For SCORING-DRIFT to know what metrics to use when COMPARING compare_metrics that YOU control, and can put WEIGHTs on also.
• See also the "docs1", "docs2","docs3" text in image
• All else, you can set to 0.0 to have no WEIGHTS when comparing scoring for model A and B, to see if we want ot promote model A
• 

DEPLOY to AKS - realtime scoring

• You can deploy a model to AKS with 2 lines of code. All ESML projects has their own private attached AKS cluster to Azure ML (BICEP)
• ESML will also save the credentials & url directly to the ESML Projects keuvault.
• ESML environments: If you are in DEV environment, the default enteprise settings is a Dev_Test (1 node AKS-cluster), if TEST or PROD environment an autoscale cluster decided by ESML core team
  • Defaults enteprise settings settings, is usually set & decided once, by an enterprise architect in the ESML core team, and all ESML Projects inherits these, but can override them also, if use case needs that.