Introduction

TRAC D.A.P. is an open-source model orchestration toolkit, maintained by Fintrac Limited in association with FINOS. Designed to work alongside any modern analytics stack, TRAC D.A.P. creates a highly controlled environment for running analytics.

This project contains:

• The TRAC platform services (metadata, data, orchestrator and gateway)

• The Python model runtime, also available to download and install from pypi.org.

See CONTRIBUTING.md for the project’s governance and how to get involved.

Metadata-driven architecture

TRAC D.A.P. is built around a single metadata model, which catalogues and describes every asset and every traceable action on the platform. This is what makes it possible to enforce consistency and produce an audit trail automatically, rather than as a separate, manually-maintained process.

See Metadata Model for a full description including objects, tags, versioning and how to query them.

Self-describing models

Models are built using a standard API (see Model API (Python)) and can be deployed to the platform without code change or manual deployment activity. The model schema (inputs, parameters and outputs) is declared in code, which allows jobs to be compiled at runtime, rather than via mutable environment configuration.

For more on writing and running models, see Modelling Tutorial.

Immutability and repeatability

Model code stays in an external repository and is only fetched when a job runs. TRAC D.A.P. data service also enforces an append-only data model.

Because the underlying objects (models and data) are immutable resubmitting a job will always produce the same result, and old jobs can be loaded, tweaked and resubmitted, for example to run last year’s models against this year’s data. Runs are isolated and stateless, so many jobs - using different versions of the same model or dataset - can execute in parallel.