Skip to main content

Architectural Decisions (ADRs)

ADR-IDTitleStatusDate
ADR-001Federated Data AccessAccepted2025-01-15
ADR-002Asynchronous ComputeAccepted2025-01-15
ADR-003Zero Trust SecurityAccepted2025-01-20
ADR-004Secure Processing EnvironmentsAccepted2025-01-20
ADR-005Standardized Workflow ExecutionAccepted2025-01-20

ADR-001: Federated Data Access​

  • Context: GDPR constraints (L01/L02) prohibit moving sensitive genomic data out of national jurisdictions to a central cloud1.
  • Decision: We will NOT build a central data lake. All data resides in National Nodes. Users query metadata centrally but execute analysis locally (Bring Compute to Data).
  • Consequences: High complexity in orchestration. Requires robust federation API standards (Beacon, WES). Legal sovereignty is preserved.

ADR-002: Asynchronous Compute​

  • Context: Genomic analysis pipelines (e.g., Variant Calling) take hours or days to run. Synchronous HTTP requests would timeout.
  • Decision: All compute interfaces (WES) must be Asynchronous. Users submit a jobID and poll for status2.
  • Consequences: UI must support "Job History" and "Notification" patterns. Fire-and-forget logic required.

ADR-003: Zero Trust Security Model​

  • Context: The system spans untrusted networks (public internet) and multiple administrative domains. We cannot rely on perimeter security (firewalls) alone1.
  • Decision: We adopt a Zero Trust model based on GA4GH Passports.
    • Identity is federated via Life Science AAI (LS AAI).
    • Authorization is decoupled and carried by the user as cryptographically signed "Visas" (JWTs).
    • Every component (PEP) must verify the signature of the Visa before granting access, assuming no network is safe.
  • Consequences: High security, decoupled authz allows scaling. Requires PKI infrastructure (Trust Anchor) and complexity in managing key rotation.

ADR-004: Secure Processing Environments (SPEs)​

  • Context: GDPR Data Protection by Design and Default (DPbDD) requires strict isolation of processing3.
  • Decision: All computation on personal data must occur within a Secure Processing Environment (SPE) / Trusted Research Environment (TRE).
    • Isolation: The compute nodes must have NO direct internet access.
    • Airlock: All ingress (code/data) and egress (results) must pass through a monitored "Airlock" or Proxy.
    • Five Safes: The implementation must map to the Five Safes framework (Safe Data, Projects, People, Settings, Outputs).
  • Consequences: GDPR compliance, reduced risk of data exfiltration. Limits user freedom (cannot pip install internet packages on the fly).

ADR-005: Standardized Workflow Execution​

  • Context: Researchers use diverse workflow languages (Nextflow, Snakemake, Galaxy). Nodes have diverse backends (Slurm, Kubernetes, AWS)2 3.
  • Decision: We mandate the GA4GH WES (Workflow Execution Service) API and/or WfExS (Workflow Execution Service) backend as the standard interface for job submission.
    • Workflows must be packaged as RO-Crates (Research Object Crates).
    • Containers (Docker/Singularity) are mandatory for reproducibility.
  • Consequences: Interoperability, reproducibility, abstraction of underlying hardware. Learning curve for researchers to containerize and RO-Crate their pipelines.

Footnotes​

  1. Bridging the European Data Sharing Divide in Genomic Science. (https://doi.org/10.2196/37236) ↩ ↩2

  2. GDI Deliverable D8.8 - Evaluation of distributed analysis and federated learning infrastructure solutions. (https://zenodo.org/records/10887366) ↩ ↩2

  3. B1MG / 1+MG - IT infrastructure requirements based on a data protection by design and default approach. (https://zenodo.org/records/8399393) ↩ ↩2