When the success of a marketplace creates more chaos than the failure it was meant to prevent

In 2016, BNP Paribas CIB launched an IT services marketplace. The idea was simple and the ambition modest: give internal teams a structured catalog of cloud resources, a single entry point to consume services without manual tickets and endless delays.

The bet paid off. Far too well.

Within a few months, its success outgrew the boundaries of the initial project. The scope extended to every business line, every internal service, every entity of the group. Thousands of employees consumed its services from Paris, New York and Singapore. Millions of artifacts were provisioned, modified, deprecated and recorded in 15 separate reference data repositories which, without coordinated governance, began to diverge silently.

That is the price of success: the faster a platform is adopted, the faster its data quality debt piles up. Each new participating provider brings its own naming conventions, its own life cycles, its own definitions of shared entities. Nobody is lying. Everyone is right within their own scope. And yet the 15 repositories no longer tell the same story.

Nobori Partners works at two distinct levels. First within the leadership of the team in charge of defining, building and running the marketplace products, at the heart of product and technical decisions. Then, when the scale of the data drift becomes impossible to ignore, as lead of the audit and the remediation plan.

The audit starts with archaeology. The goal is not to list anomalies; any ETL tool can do that. The goal is to understand why the system structurally produces them: map the interactions between providers through meta-orchestration, reconstruct the lineage of every artifact, and identify the friction points where repositories fall out of sync. The modeling reveals an architecture of complexity: cascading dependencies, life cycles that do not talk to each other, golden records that exist nowhere.

Then comes remediation. For every critical entity, a golden record is defined: a single reference version toward which all repositories converge. A unified data lake absorbs the multi-source complexity on S3 and Spark. NiFi and Airflow pipelines automate reconciliation. But the most tangible gain comes from the controlled removal of obsolete artifacts: thanks to a multi-step decommissioning process (identification, qualification, business validation, deletion), several thousand virtual machines are removed safely. Without this rigorous process, nobody dared to touch these artifacts for fear of breaking something.

But remediation is only a temporary victory if nobody watches what comes next. The last phase industrializes observability: QlikSense dashboards cover all 15 repositories, every quality indicator is visible the next day, and every synchronization anomaly is detected before it affects users in Paris, New York or Singapore.

The result: several thousand virtual machines removed through a controlled process, €2.5M in costs cut, data governance that holds over time, and thousands of people on 3 continents finally working from the same truth.