Skip to content
Alvorance

Architectures

Reference architectures.

Models for the most frequent kinds of project, with the technical decisions that support each design.

Data engineeringPractice

Modern data platform

Context

Data scattered across spreadsheets, production databases and manual exports. Every report is rebuilt from scratch and nobody can audit where a number came from.

  1. 01 · Sources

    • Transactional database
    • CRM and ERP
    • Third-party APIs
    • Product events

  2. 02 · Ingestion

    • CDC for databases
    • Scheduled batch
    • Streaming for events

  3. 03 · Layers

    • Raw: faithful copy of source
    • Cleansed: typed and validated
    • Serving: modeled by domain

  4. 04 · Transformation

    • Versioned models
    • Contract tests
    • Traceable lineage

  5. 05 · Consumption

    • BI and dashboards
    • Reverse ETL to CRM
    • Features for models

Design decisions

  • The raw layer is never mutated, and that is what lets you reprocess when a rule changes.
  • Every published model has a contract test: it breaks the pipeline before it breaks the dashboard.
  • Nobody queries the raw layer directly. Needing to is a sign a model is missing.
Analytics & trackingPractice

Server-side tracking with consent

Context

Blockers and cookie limits wiped out conversion volume. Platforms disagree with each other and legal is asking for privacy compliance.

  1. 01 · Browser

    • Site data layer
    • Consent banner
    • Consent Mode v2

  2. 02 · Own collector

    • First-party domain endpoint
    • Server-side container
    • Enrichment and validation

  3. 03 · Policy

    • Consent signal per purpose
    • Personal data redaction
    • Drop what wasn't consented

  4. 04 · Destinations

    • GA4
    • Ad platform Conversions APIs
    • CRM

  5. 05 · Warehouse

    • Raw event retained
    • Own attribution model
    • Reconciliation with finance

Design decisions

  • Consent denied by default. No tag fires before opt-in, including on this site.
  • The raw event reaches the warehouse before the ad platforms: that's how you audit disagreement between them.
  • Server-side recovers signal, but does not bypass consent. Anyone promising that is selling legal risk.
Software engineeringPractice

Delivery pipeline

Context

Deploy is a risk event scheduled for Friday night. There is no trustworthy test environment and the first news of failure comes from the customer.

  1. 01 · Commit

    • Short-lived branch
    • Mandatory review
    • Type and lint checks

  2. 02 · Integration

    • Unit tests
    • Integration tests
    • Reproducible build

  3. 03 · Ephemeral environment

    • One environment per pull request
    • Synthetic data
    • End-to-end test

  4. 04 · Production

    • Progressive rollout
    • Feature flag
    • One-command rollback

  5. 05 · Observability

    • Business metrics, not just CPU
    • Alerts with an owner
    • Blameless post-mortem

Design decisions

  • If rolling back isn't cheaper than fixing forward, the pipeline is wrong.
  • An alert without a named owner becomes noise in two weeks and nobody looks again.
  • An ephemeral environment per pull request costs less than one night of manual rollback.

Let's talk about your project.

An initial conversation to understand the context and assess how we can help.

Get in touch