CATs: Content-Addressable Transformers

Description:

Content-Addressable Transformers (CATs) is a unified Data Service Collaboration framework for organizations implemented as an edge-computing service that establish a Data Mesh as a scalable self-serviced Data Platform of Data Products with Data Provenance. CATs connect collaborators between organizations on a Data Mesh via the Content-Addressed Storage (CAS) of interoperable and scalable data processing to enable Data Provenance. CAT data processing workloads (CATs) are deployable as parallelized and distributed processes at horizontal & vertical scale to support scalable (big) data processing microservices with Scientific Computing capabilities. CATs are also integration points which enable scaled data processing portability between client-server cloud platforms and mesh (p2p) networks with minimal rework or modification.

• Techncal Use-Case Specification: CATs' utilizes Ray as an execution middleware framework deployed on Kubernetes for interoperable & parallelized distributed computing applications / Big Data processing with Scientific Computing enabled ecosystem integrations such as Apache Spark, and PyTorch.

How are CATs workloads processed as Data Provenance Records?

CAT's (Data) Mesh is specified / reified by CATs executing Bills-Of-Materials (BOMs) as specifications used to chain CAT Nodes' content-addressed records into a verifiable lineage of Data Provenance. (* Details)

CATs are submitted as content-addressed Orders of data processes (transformers) which are Invoiced for verification and logged as BOMs that serve as Data Provenance records that are unique identifiers of CAT workloads and their content. BOMs are CATs' Content-Addressed Data Provenance records for verifiable data processing with URIs for transport over a Mesh network of CATs. BOMs are also used as CAT’ input & output that contain CATs’ means of data processing between CAT Nodes.

BOMs employ Content Identifiers (CIDs) for CAS to provide a means of location-agnostic data transportation / retrieval of based on its content / CAT processes for Data Verification. Therefore, the implementation of CATs' as content-addressed data processes establishes and self-services a scalable Data Platform as a Data Mesh network of interoperable distributed computing workloads deployable on Kubernetes as CATs execution paradigm.

• BOM CIDs can be used to verify the means of processing data (input, transformation / process, output, infrastructure-as-code (IaC)) they can also make CATs resilient by enabling re-execution via retrieval. CATs certifies the accuracy of data processing on data products and pipelines by enabling maintenance and reporting of data and process lineage & provenance as chains of evidence using CIDs.

How do CATs enable colaborative Data Processing for Data Initiatives?

CATs enables the continuous reification of Data Initiatives by cataloging discoverable, accessible, and re-executable workloads as Data Service Collaboration composable records between organizations. These records provide a reliable and efficient way to manage, share, and reference data processes via Content-Addressing Data Provenance records. Data Initiatives will be naturally reified as a result of Data Service Collaboration on CATs. CATs will be compiled and executed as interconnecting services on a Data Mesh that grows naturally when organizations communicate CATs provenance records within feedback loops of Data Initiatives.

What is Content Addressing & How does CATs use it?

Content-Addressing is a method of uniquely identifying and retrieving data based on its content rather than its location or address. CATs provides verifiable data processing and transport on a Mesh network of CATs interconnected by Content-Addressing Data Provenance records with IPFS CIDs (Content-Identifiers) as content addresses issued by IPFS client to identify and retrieve inputs, transformations, outputs, and infrastructure (as code [IaC]) for verifying transformation accuracy given CIDs.

• IPFS serves as CATs' Data Mesh's network layer to provide parallelized data ingress and egress for IPFS data. This network portability closes the gap between data analysis and operations by connecting the network planes of the cloud service model (SaaS, PaaS, IaaS) with IPFS. CATs connect these network planes by enabling the instantiation of FaaS with cloud services in AWS, GCP, Azure, etc. on a Data Mesh network of CATs. IPFS enables this connection as p2p distributed-computing job submission in addition to the client-server job submission provided by Ray.

Get Started!:

1. Install Dependencies (including uv, which manages CATs' Python interpreter, virtual environment, and locked dependencies)
2. Install CATs:

  git clone git@github.com:DynamicalSystemsGroup/cats.git
  cd cats
  uv python install   # installs the Python version pinned in .python-version
  uv sync             # creates .venv and installs locked dependencies from uv.lock

See ENV.md for the full environment workflow, including the ops and mac extras. 3. Demo: Establish a CAT Mesh 4. Test: CAT Mesh Verification 5. Experiments

Contribute!

CATs' Architectural Quantum:

Organizations and collaborators participating will employ CATs for rapid ratification of service agreements within collaborative feedback loops of Data Initiatives. CATs' apply an Architectural Quantum Domain-Driven Design principle described in Data Mesh of Data Products to reify Data Initiatives.(* Design Description)

The Action Plane is the Analytical Data Processing interface. The Action Plane orchestrates and supervises how virtual resources owned by the Data Product should be managed, routed, and processed and is stored “offmesh” (“offline”). It supervises the exchange of data between sub-Process components within the Data sub-Plane (Process) in adherence to Data Contracting Standards of organizations participating in a Data Mesh.

Quantum Architecture Description as a Minimal Federated Operating Model

• Function is a FaaS for scalable Data Processing and analytics executed as CAT Processes. Functions (FaaS) are deployed on Structure (PaaS) to execute Processes orchestrated by InfraFunctions (FaaS)
  • Processes are Functional Data Processors executable by InfraFunctions (FaaS) deployed on Structure (PaaS), and contextualized with pre and post processed data by InfraFunctions (FaaS). Processes (FaaS) are executed with and made orchestratable by InfraFunctions (FaaS) to support the following use-cases
    • The CAT Order is updated with the inclusion of resulting mutated Functions (FaaS) for execution processed by CATs Factory Client.
  • InfraFunction (FaaS) is a Data Processing orchestrator that employs a CAR for the configurable execution of scalable Processing operated by the Plant (SaaS)
    • The CAT Order is updated in alignment CATs Architectural Quantum’s Functionality. This Order will include the resulting updated of Structure (PaaS) with respect to the updated Plant (SaaS) and an updated Function (FaaS) with updated Ingress and Egress subProcesses (FaaS)
• Structure (PaaS as IaC) provisions and maintains the Plant (SaaS) as Function’s (FaaS) scalable execution environment.
  • The Plant (SaaS) is a dynamically scaled execution environment of Function (FaaS) as an IaC plugin(s)
    • The web application codebase is Content Addressed within CAT Orders as Data Contract metadata for Order registration.
  • InfraStructure (IaaS) supports the provisioning of dynamically scaled infrastructure for maintaining a Plant (SaaS).
    • The CAT Order is updated in alignment with event-driven functionality and operations with the resulting mutation of Structure (PaaS).

CAT Mesh: CATs Data Mesh platform with Data Provenance

CAT Mesh is a self-serviced Data Mesh platform with Data Provenance. CAT Nodes are CAT Mesh peers that enable workloads to be portable between client-server cloud platforms and p2p mesh network with minimal rework or modification.

Multi-disciplinary and cross-functional teams can use CAT Nodes to verify and scale distributed computing workloads. Workloads (CATs) executed by CAT Nodes interface cloud service model (SaaS, PaaS, IaaS) offered by providers such as AWS, GCP, Azure, etc. on a Mesh Network interconnected by IPFS.

CAT Nodes are Data Products - peer-nodes on a mesh network that encapsulate components (*) to function as a service providing access to a domain's analytical data as a product; * code, data & metadata, and infrastructure.

In the following image:

• Large ovals in the image above represent Data Products servicing each other with Data
• "O" ovals are Operational Data web service endpoints
• "D" ovals are Analytical Data web service endpoints
• Source: Data Mesh Principles and Logical Architecture - Zhamak Dehghani, et al.

Key Concepts:

• Data Verification - a process for which data is checked for accuracy and inconsistencies before processed
• Data Provenance - a means of proving data lineage using historical records that provide the means of pipeline re-execution and data validation
• Data Lineage - reporting of data lifecyle from source to destination
• Distributed Computing - typically the concurrent and/or parallel execution of job tasks distributed to networked computers processing data
• Bill of Materials (BOM) - an extensive list of raw materials, components, and instructions required to construct, manufacture, or repair a product or service

Image Citations:

• "Illustrated CAT"
  • Python logo
  • SQL logo
  • Terraform logo
  • IPFS logo
  • cat image
  • ray.io logo

Acknowledgments

CATs was developed by the Dynamical Systems Group (DSG) team.

Key contributions:

• Network Architecture & Verified Information Exchange:
  • Michael Zargham (mzargham)
  • David Sisson
• Lead Solutions Architect & Developer / Distributed Systems Engineer
  • Joshua E. Jodesty
• Testing: Danilo