Source: containerjournal.com (original coverage, August 2020)
Overview
Hewlett Packard Enterprise (HPE) extended its container platform — rebranded as HPE Ezmeral — to support a data fabric that spans multiple Kubernetes clusters distributed across geographically separated data centers. The announcement centers on HPE Ezmeral Data Fabric 6.2, a release that adds significant capabilities for enterprises building large-scale, distributed containerized workloads, particularly those involving machine learning.
Background: MapR Acquisition and the Ezmeral Brand
The data fabric technology at the core of HPE Ezmeral originates from HPE’s acquisition of MapR Technologies. MapR had long offered a distributed data platform with strong support for files, tables, and event streams over a unified namespace. After the acquisition, HPE integrated this technology into its broader container and data management portfolio and rebranded the platform under the Ezmeral name.
What HPE Ezmeral Data Fabric Does
At its core, the HPE Ezmeral Data Fabric creates a global namespace accessible through standard APIs. This means applications — whether containerized or non-containerized — can read and write data without needing to know the physical location of that data. Data can be stored and accessed as:
- Files — using a POSIX-compatible file system interface
- Tables — for structured data with NoSQL-style access
- Event streams — for real-time, pub/sub messaging patterns
A data mirroring capability enables movement of data within or between clusters. This uses bi-directional, multi-master table replication and event stream replication, making it suitable for active-active distributed deployments where data consistency across sites is required.
New Features in HPE Ezmeral Data Fabric 6.2
According to Nanda Vijaydev, distinguished technologist and lead data scientist at HPE, version 6.2 enables IT organizations to manage fleets of Kubernetes clusters spread across multiple geographic locations. The specific new capabilities in this release include:
- Snapshot restore — the ability to roll back data volumes to a previous point-in-time state
- Fine-grained data placement — allows administrators to pin or tier data to specific classes of heterogeneous storage devices within and across nodes
- Policy-based security — enforces data governance rules automatically, reducing the burden on individual application developers to implement access controls
- Last access tracking — provides visibility into when data was last read, useful for lifecycle management and cost optimization
- Flash storage enhancements — improved performance and reliability when using NVMe or SSD-backed storage tiers
- Global file check enhancements — improved consistency verification across distributed file system volumes
- Improved erasure coding performance and rebuild — faster recovery of data after node or drive failures without full replication overhead
- Better management metrics — expanded observability for administrators monitoring cluster health and data operations
- External key management support — integration with external KMS (Key Management Service) systems for enterprise encryption workflows
Architecture: A Stackable Platform on Kubernetes
Vijaydev describes the HPE Ezmeral platform as a stackable services model built on top of a core Kubernetes cluster. Rather than deploying a single monolithic platform, IT teams can add capabilities incrementally as their container application environment matures. The data fabric layer provides the persistent, distributed storage foundation, while additional services — including ML tooling and cluster management — can be layered on top.
This approach is intended to lower the initial barrier to adoption while giving organizations a clear upgrade path as workload complexity grows.
MLOps as a Primary Driver
A significant portion of HPE’s positioning around Ezmeral targets machine learning operations (MLOps). Training large machine learning models requires access to massive datasets. Containers make it practical to decompose what would otherwise be monolithic training pipelines into modular, independently updatable components. Managing training data, model artifacts, and inference results across multiple clusters and locations requires exactly the kind of distributed, consistent data layer that Ezmeral aims to provide.
HPE’s view is that as ML workloads scale, enterprises will need a cohesive software stack addressing both the Kubernetes management layer and the data management layer — rather than assembling those capabilities from disparate open-source tools.
Deployment Options: Self-Managed vs. GreenLake
HPE offers the Ezmeral platform through two routes:
- Self-managed — the organization’s internal IT team deploys and operates the platform on their own infrastructure
- HPE GreenLake — a managed service model in which HPE handles infrastructure operations and, where applicable, updates to select third-party packaged application environments
The GreenLake option reflects a broader industry trend toward consumption-based IT, where enterprises prefer to pay for outcomes rather than own and operate infrastructure outright. HPE is positioned to benefit if organizations decide to offload Kubernetes and data fabric management as those environments grow in complexity.
Implications for Enterprise Kubernetes Strategy
Running Kubernetes in a single cluster is operationally challenging enough. Running fleets of clusters across multiple data centers — with shared, consistent data — raises the complexity significantly. Capabilities like global namespaces, bi-directional replication, and policy-based governance directly address the pain points that arise at that scale.
For DevOps and platform engineering teams evaluating multi-cluster data strategies, the key questions to assess against any solution — including HPE Ezmeral — include:
- How does the platform handle network partitions between geographically separated clusters?
- What is the replication latency for event streams and tables under realistic WAN conditions?
- How does policy-based security integrate with existing identity providers (e.g., LDAP, Active Directory, OIDC)?
- What is the operational overhead of running the data fabric control plane alongside existing Kubernetes operators?
HPE Ezmeral Data Fabric 6.2 represents a meaningful step toward making geo-distributed Kubernetes data management a production-ready capability, with MLOps use cases providing the commercial urgency to accelerate development.