As Zhamak Dehghani, write about the IceMesh (Iceberg as enabler for multi-vendor data mesh) where Multiple Readers, One Writer (MROW — reverse WORM) means each Iceberg catalog is written (maintained) by a single system, but accessible to all. Who are not required to agree on anything except where each Iceberg catalog lives.
ChatGPT Prompt
1. A Mesh Made of Ice
Over the past few years, the data mesh paradigm has inspired many organizations to rethink how they scale data architectures across distributed domains. But for data mesh to become truly interoperable and multi-vendor, it needs more than just organizational change — it needs a universal substrate that can preserve local ownership without requiring global agreement.
Enter IceMesh — a conceptual pattern where Apache Iceberg enables a decentralized mesh of domain-owned data products, with semantics that balance autonomy and interoperability.
At the heart of IceMesh lies a deceptively simple principle:
MROW — Multiple Readers, One Writer.
2. MROW: The Minimum Viable Contract
In traditional data architectures, multiple systems often compete to write or interpret the same data — leading to brittle pipelines, conflicting schemas, and governance deadlocks.
In IceMesh, we flip the model.
Each Iceberg catalog is written by only one system — the domain’s system of record — but can be read by anyone.
That’s it.
There’s no need for shared compute infrastructure, centralized governance layers, or consensus on how data is modeled.
The only agreement required is:
Where does the catalog live, and how can I read from it?
This is MROW at catalog granularity — the minimal contract needed for a viable, multi-vendor data mesh.
3. Why Catalog-Level Ownership?
Why not one writer per table or per partition? Because in a data mesh:
- Ownership belongs to domains, not engineers or pipelines.
- Domains are responsible for all aspects of their data products — structure, semantics, lifecycle.
- The Iceberg catalog becomes the boundary of ownership and accountability.
When each domain publishes its data via a catalog it alone writes to, it retains:
- Full control over schema evolution and branching
- Autonomy in compliance and governance
- The ability to innovate without upstream coordination
Meanwhile, everyone else — consumers, vendors, and even AI agents — can read without fear of side effects.
4. The Beauty of Minimal Agreements
We often mistake interoperability for standardization. But in a complex, polyglot ecosystem, too much standardization stifles change.
MROW thrives on a different philosophy:
- Encapsulation over integration
- Composability over consensus
- Discoverability over uniformity
By requiring only that consumers know where to look and how to interpret Iceberg, we unleash a world where:
- A domain in Snowflake can publish to an Iceberg catalog
- A consumer in Databricks or Flink can read it without asking permission
- Governance is composable, localized, and programmable
It’s the Unix pipe dream applied to data: simple contracts, powerful compositions.
5. Governance and Trust
Does this mean chaos? Not at all.
- Catalog registries (e.g., Nessie, AWS Glue, Unity Catalog) enable discoverability and lineage.
- Snapshot immutability ensures reproducibility.
- Schema versioning and branching offer safe evolution paths.
- WORM (Write Once, Read Many) policies can be enforced per catalog where needed for audit or compliance.
Governance becomes a function of ownership and policy, not infrastructure diktats.
6. IceMesh in Practice
IceMesh is not a product — it’s a pattern:
- Use Apache Iceberg to expose your domain’s data product as a table or view.
- Store it in your preferred object store (S3, GCS, etc.).
- Register it in a catalog that your domain controls.
- Let others consume it freely — with the guarantee that you are the only one who writes.
Every domain becomes a publisher.
Every consumer becomes self-serve.
Every catalog becomes a boundary of trust.
7. The Way Forward
Data mesh was never about tools. It’s about rethinking responsibility and trust in a data-driven world.
With MROW in IceMesh, we get:
- Sovereign data products
- Federated evolution
- Global interoperability through local clarity
The mesh scales, not because everyone agrees — but because they don’t have to.
And that, I believe, is what will make data mesh real.
ChatGPT as Zhamak Dehghani
Inventor of Data Mesh
LinkedIn | Twitter
iHack, therefore iBlog 
Leave a comment