Dataflow Architecture

Revisiting the Classic Software Architecture Paradigm
in the age of AI

Andrei Paleyes

SAP Inspiration Sessions, 07.06.2023

Who is Andrei?

Last year PhD student at ML@CL (Cambridge Computer Lab)
with lots of software engineering background

Who are you?

Software engineers?

Data scientists?

ML/data engineers?


ML deployment is hard!

Venture Beat, 2019

“Why do 87% of data science projects never make it into production?”

InfoWorld, 2021

“85% of AI and machine learning projects fail to deliver, and only 53% of projects make it from prototypes to production.”

Capital One and Forrester, 2022

“73% of respondents find transparency, traceability, and explainability of data flows challenging.”

But why?

Many reasons

e.g. model accuracy vs business value

or computational and labour costs

or skillset

"150 Successful Machine Learning Models: 6 Lessons Learned at", Bernardi et al., KDD 2019
"Challenges in Deploying Machine Learning: A Survey of Case Studies", Paleyes et al., ACM Computing Surveys, 2022

One more reason

Data management in modern software is often a mess

Scaling Big Data Mining Infrastructure: The Twitter Experience J Lin, D Ryaboy; ACM SIGKDD Explorations Newsletter, 2013

“Effective big data mining at scale doesn't begin or end with what academics would consider data mining”
“Data scientists expend a large amount of effort to understand the data available to them, before they even begin any meaningful analysis”
“Exploratory data analysis always reveals data quality issues”

Who's to blame?

Software Services

What is a service

A service is a piece of software,

that provides a function, or many functions,

known as interface or API,

that clients* can reuse,

together with policies to control its usage.

*A client can be anything: another software, a person, a hardware.

Service oriented architecture is

  • Scalable
  • Flexible
  • Modular
  • Reliable
  • Encourages ownership


Two services

Three services

Big ball of mud

Image credit: Ben Morris,

Scaling Big Data Mining Infrastructure: The Twitter Experience J Lin, D Ryaboy; ACM SIGKDD Explorations Newsletter, 2013

“Twitter is powered by many loosely-coordinated services.”
“Since a single user action may involve many services, a data scientist wishing to analyze user behavior must first identify all the disparate data sources involved.”
“Services are normally developed and operated by different teams, which may adopt different conventions for storing and organizing log data.”

What to do?

What else to do?

What else else to do?

Build software with data as the first priority!


  • Prioritise data while designing services - Götz et al., 2018
  • Split data storage to encourage ownership - Data Meshes, Dehghani, 2019
  • Cluster services by data domains - Domain-Oriented Microservice Architecture, Uber, 2020

Or... roll on dataflow!

What is dataflow architecture?

"Data flow schemas", Dennis et al., International Symposium on Theoretical Programming, 1974

Boring slide

Control flow: instructions are executed one after another, classic von Neumann architecture.

Dataflow: instruction is ready to execute as soon as all its inputs are available.

Intuition behind the boring slide

Control flow is about operations and their order

Dataflow is about data routes and transformations

Examples of dataflow in action

  • Flow-based programming
  • MapReduce
  • Data streaming

You are likely familiar with dataflow already!

Features of dataflow architecture

  • Separates data and logic
  • Data as a first class citizen
  • Complete dataflow graph
  • Decentralised

Benefits of dataflow architecture for (AI) systems

Benefit 1: data discovery and collection

"Towards better data discovery and collection with flow-based programming", Paleyes et al., DCAI Workshop, NeurIPS 2021
Image credit: Compex IT,

Benefit 2: deployment of ML

"An empirical evaluation of flow based programming in the machine learning deployment context", Paleyes et al., CAIN 2022
"Assuring the machine learning lifecycle: Desiderata, methods, and challenges", Ashmore et al., ACM Computing Surveys, 2021

Benefit 3: data governance

"Decision provenance: Harnessing data flow for accountable systems", Singh et al., IEEE Access, 2018
"Desiderata for next generation of ML model serving", Akoush et al., DMML Workshop, NeurIPS 2022
"A Primer on Provenance: Better understanding of data requires tracking its history and context.", Carata et al., ACM Queue, 2014
Image credit: redgate,

This is a causal graph!

  • Arrows show causal relationships
  • Complete by design
  • No post-hoc discovery
  • No expert knowledge
  • Interventions:
    • Software updates
    • Bugs
    • Input data shifts

"Dataflow graphs as complete causal graphs", Paleyes et al., CAIN 2023

Benefit 4: causal fault localisation

"Causal fault localisation in dataflow systems", Paleyes et al., EuroMLSys 2023

Benefit 4: causal fault localisation

"Causal fault localisation in dataflow systems", Paleyes et al., EuroMLSys 2023

Benefit 5: experimentation

Intellectual debt

"Intellectual Debt: With Great Power Comes Great Ignorance", Zittrain J, Medium: Berkman Klein Center Collection, 2019

So how do we build explainable (AI) systems?

  • Explicit dataflows
  • Data as a first priority
  • Causal reasoning

But also

  • Statistical emulation of systems
  • End-to-end system optimisation
  • Self-adaptive and continual learning systems
  • System FITness
  • Security and privacy

"Real-world Machine Learning Systems: A survey from a Data-Oriented Architecture Perspective", Cabrera et al., 2023
"Machine learning from innovation to deployment: A strategic research agenda for AutoAI", ML@CL, 2022


  • Services are good
  • But the world needs data as a priority
  • Dataflow architecture may help
  • Dataflow improves ML deployment
  • Dataflow enables causal inference on systems
  • Intellectual debt calls for a wide research agenda