Bibliothèque de concepts

Un carnet de bord de ma réflexion sur chaque problème.

Plus de 50 entrées couvrant l'architecture des données, l'ingénierie des données, la BI et l'analyse, et l'IA agentique. Chaque fiche suit la même structure : ce que c'est, pourquoi c'est important, comment ça fonctionne, des comparaisons d'outils basées sur ce que j'ai réellement déployé, et mon avis synthétique.

ENLes entrées de la bibliothèque sont rédigées et maintenues en anglais.

Le système d'exploitation

Comment les pôles s'assemblent.

01
Ingénierie des Données
Pipelines, calcul & orchestration
02
Architecture des Données
Stockage, modèles & sémantique
03
BI & Décisionnel
Métriques, tableaux de bord & décisions
04
IA Agentique
Flux de travail, outils & orchestration

Architecture des Données

14 entrées

La forme que prennent les données lorsqu'elles doivent servir plusieurs consommateurs de manière fiable — lakes, warehouses, lakehouses, mesh, couches sémantiques et les contrats qui les lient.

Parcourir le pôle

Ingénierie des Données

12 entrées

Les disciplines qui permettent de faire circuler les données de manière fiable à grande échelle — orchestration, transformation, contrats, observabilité, coûts et modèle opérationnel associé.

Parcourir le pôle

BI & Décisionnel

8 entrées

La couche qui transforme les données en décisions — modèles sémantiques, tableaux de bord, intelligence décisionnelle et la rigueur de conception des indicateurs pour les directions générales.

Parcourir le pôle

IA Agentique

16 entrées

Bâtir des systèmes d'IA qui agissent, et pas seulement répondent — topologies d'agents, garde-fous, évaluation, gouvernance et les architectures d'agents viables en production.

Parcourir le pôle

Architecture des Données

14 entrées

Intermédiaire

Data Federation and Virtualization

Querying data where it lives across multiple systems through a unified virtual layer, instead of physically copying it into a central warehouse.

Intermédiaire

Data Lake vs Data Warehouse vs Lakehouse

Three storage and compute paradigms — schema-on-read, schema-on-write, and the convergent model that combines both — each fit for different organizational profiles.

Intermédiaire

Data Mesh

An organizational and architectural pattern that distributes data ownership to domain teams, treats data as a product, and federates governance — replacing the central data team with a platform-and-policy model.

Intermédiaire

Data Vault Modeling

An ensemble modeling pattern (hubs, links, satellites) designed for auditability and adaptability in environments with frequent source changes — the right answer for regulated, integration-heavy data warehouses.

Avancé

Lambda vs Kappa vs Unified Streaming

Three architectural patterns for processing batch and streaming data — preserving them as separate paths, collapsing them into one streaming-only path, or unifying them on a single engine.

Intermédiaire

Medallion Architecture

A three-zone data architecture — bronze, silver, gold — that separates raw, refined, and business-ready data within a single lakehouse.

Fondamental

The Modern Data Stack

The cloud-native, modular set of tools — ELT, warehouse, transformation, semantic layer, BI — that replaced monolithic data platforms in the 2015-2020 decade.

Avancé

Multi-Region Data Architectures

Patterns for distributing data across regions to satisfy residency laws, latency requirements, and disaster-recovery commitments — without losing analytical coherence.

Intermédiaire

OLTP vs OLAP vs HTAP

Three database workload patterns — transactional, analytical, and hybrid — each requiring different storage layouts, query patterns, and operational commitments.

Avancé

Open Table Formats — Iceberg, Delta, Hudi

Open metadata layers built on top of Parquet that bring transactional guarantees, schema evolution, time travel, and indexing to lakehouse storage.

Avancé

Reverse ETL and Operational Analytics

The pattern of pushing modeled data from the warehouse back into operational tools (CRM, marketing, support) — turning the warehouse into the system of record for data that drives action.

Intermédiaire

The Semantic Layer

A governed contract between physical data and consuming tools — defining metrics, dimensions, and access in one place so that every consumer (BI, ML, agents) reads the same business truth.

Intermédiaire

Slowly Changing Dimensions

A set of patterns (Type 0 through Type 7) for handling attribute changes in dimension tables — choosing how much history to preserve and how to expose it.

Intermédiaire

Star, Snowflake, and Galaxy Schemas

Three dimensional modeling patterns that organize fact and dimension tables for analytical consumption — each trading off normalization against query performance and complexity.

Ingénierie des Données

12 entrées

Essentiel

Batch vs Streaming vs Micro-Batch

Three processing modes — periodic batches, continuous streams, or small frequent batches — each with different latency, cost, and complexity profiles.

Essentiel

Change Data Capture (CDC)

Capturing every insert, update, and delete from an operational database in near real-time, instead of polling for changes on a schedule.

Essentiel

Cost Optimization for Data Platforms

The discipline of producing the same outcomes for materially less money — through better pipeline design, smarter compute selection, and operational discipline most teams skip.

Intermédiaire

Data as a Product

Treating each data asset like a product — with an owner, a contract, a roadmap, an SLA, and consumers who can give it the equivalent of an NPS rating.

Essentiel

Data Contracts

An explicit, versioned agreement between data producers and consumers — schema, semantics, SLA, ownership — enforced in CI before bad data reaches production.

Essentiel

Data Quality and Observability

The practice of catching bad data before it reaches a dashboard or a model — through assertions, anomaly detection, and end-to-end lineage.

Essentiel

Data Transformation: dbt, SQLMesh, Coalesce

The transformation layer turns raw warehouse data into modeled, tested, documented tables that the rest of the business consumes. dbt defined the category; SQLMesh and Coalesce are credible challengers.

Intermédiaire

DataOps Maturity

The practices that make a data platform reliable: version control, CI/CD, environments, automated testing, observability, and clear incident response — borrowed from software engineering and adapted for data.

Intermédiaire

Idempotency and Exactly-Once Semantics

Designing pipelines so that running the same job twice produces the same result, and so that each input event is reflected in output exactly once — even after retries, failures, and replays.

Essentiel

Orchestration Frameworks: Airflow, Dagster, Prefect, Temporal

The control plane that schedules pipelines, tracks dependencies between tasks, retries failures, and gives engineers a single place to ask 'did the data run, and is it correct?'

Intermédiaire

Schema Evolution and Backwards Compatibility

How to change the shape of data over time — adding columns, renaming fields, deprecating tables — without breaking the consumers downstream.

Intermédiaire

Team Topologies for Data and AI

How to organize people around data and AI work — central platform team, embedded analytics engineers, domain ownership, agentic system squads — so the org structure matches the work.

BI & Décisionnel

8 entrées

Intermédiaire

Anomaly Detection and Forecasting

Statistical and ML techniques for spotting metric movements that warrant attention, and projecting where metrics are likely to land — applied carefully so that signals do not become noise.

Essentiel

Dashboard Design Principles

The visual and structural choices that make a dashboard usable: clear hierarchy, deliberate metric selection, restraint in chart types, and an explicit answer to the question 'what decision does this support?'

Intermédiaire

Decision Intelligence

Designing analytics around the decision being made — including the action, the decision-maker, and the feedback loop — rather than around the data being available.

Intermédiaire

Embedded vs Standalone Analytics

The choice between analytics inside the application your customer or employee already uses, versus analytics in a separate BI tool — with very different design, governance, and cost implications.

Intermédiaire

Executive Metric Design

Selecting, defining, and operationalizing the small set of metrics that drive executive decisions — distinct from the operational metrics teams use day-to-day.

Essentiel

The Modern BI Stack

The architectural pattern of warehouse + semantic layer + BI tool, replacing the monolithic BI platforms of the previous decade.

Intermédiaire

Reverse ETL and Activation

Pushing modeled data from the warehouse back into operational tools — CRM, marketing automation, support — so that insights become actions in the systems where work happens.

Essentiel

Self-Service Analytics

Letting business users answer their own data questions through governed, modeled, semantic-layer-backed tools — without going through a ticket queue.

IA Agentique

16 entrées

Essentiel

Agent Evaluation Frameworks

Systematic ways to measure whether an agentic system works — output evaluation, trace evaluation, regression suites, A/B testing — and the discipline that turns 'it seems to work' into 'we know it works'.

Essentiel

Agentic AI Architecture Patterns

The reusable shapes for organizing LLMs and tools into systems — from single-agent loops to orchestrator-worker hierarchies to graph-based state machines — each with different trade-offs in cost, latency, predictability, and capability.

Intermédiaire

AI Governance Frameworks

The policies, processes, and roles that govern how AI systems are built, deployed, and monitored — covering risk classification, approval gates, ongoing monitoring, audit trails, and incident response.

Essentiel

Cost and Latency Budgeting for AI

The discipline of designing AI systems within explicit cost-per-run, latency-per-step, and total-budget envelopes — and the architectural choices that make those envelopes hold under real traffic.

Essentiel

Deterministic Guardrails

The non-LLM checks — input validation, output validation, allowlists, budget caps, content filtering — that constrain what an agentic system can do regardless of what the model decides.

Essentiel

Fine-tuning vs RAG vs Tool Use

Three ways to make a model good at your specific problem — teach it (fine-tuning), give it the relevant content at runtime (RAG), or let it act (tool use) — each fitting different problem shapes.

Essentiel

LLM vs Agentic: The Difference

An LLM produces an answer. An agentic system uses LLMs in a loop with tools, memory, and decisions — taking actions, observing outcomes, and continuing until a goal is met.

Essentiel

MCP: Model Context Protocol

An open protocol that lets LLM-powered systems connect to external tools, data sources, and services through a standardized interface — making integrations portable across models, frameworks, and clients.

Essentiel

Memory in Agents

How agentic systems remember — within a single run (working memory), across runs (long-term memory), and across users (shared knowledge) — and what each layer is good and bad at.

Essentiel

Multi-Agent Orchestration

Coordinating multiple specialized agents to solve a problem together — through hierarchical orchestrator-worker patterns, peer-to-peer collaboration, or graph-based state machines.

Essentiel

Production Patterns for Agentic Systems

The architectural and operational patterns that separate agentic systems that survive in production from prototypes that break under load — covering observability, durability, deployment, evaluation, and incident response.

Intermédiaire

Prompt Engineering vs Architecture

The distinction between writing better prompts and designing better systems — and why most production failures are architectural, not linguistic.

Essentiel

RAG: Retrieval-Augmented Generation

Augmenting LLM context at runtime with relevant documents retrieved from an index — letting models answer questions about content they were never trained on, with citation discipline.

Essentiel

Tool Use and Function Calling

The mechanism that lets an LLM emit structured calls to external functions — query a database, send an email, call an API — turning a text generator into something that can act on the world.

Essentiel

Vector Stores and Embedding Strategy

The storage and retrieval layer for embedding-based search — and the design choices around embedding models, chunking, hybrid retrieval, and operational scale that determine whether RAG works in production.

Essentiel

What Is an Agent, Really

An LLM-powered system that decides what to do next based on its goal, takes actions through tools, observes the results, and continues until the goal is met or the budget is exhausted.

Document vivant

La bibliothèque est revue chaque trimestre. Si une comparaison d'outils est dépassée ou qu'une définition a vieilli, signalez-le moi et je mettrai à jour la fiche — avec mention de votre contribution si vous le souhaitez.