MayaScale Datasheet

Cloud Block Storage Caps Your Performance

Cloud disk APIs limit IOPS and throughput per instance — regardless of how much backend storage you attach. Databases stall, GPUs sit idle waiting on I/O, and teams overprovision expensive provisioned disk to work around the ceiling.

Kubernetes makes it harder: stateful workloads need fast, HA persistent volumes with per-tenant guardrails, but managed clusters offer no high-performance shared block tier that bypasses the per-instance disk limits.

MayaScale delivers disaggregated NVMe block storage over standard NVMe-over-TCP, with a Kubernetes CSI driver, flexible pools, and ephemeral-SSD protection — 2M+ IOPS per node at sub-millisecond latency, inside your own cloud account.

Benefits

2M+ IOPS per node, sub-millisecond latency
Network-attached — not limited by per-instance cloud disk caps
Kubernetes CSI for EKS, GKE, and AKS — zero client footprint
Ephemeral-SSD protected — HA + DR to cold object storage
Runs in your own cloud — full data sovereignty

Disaggregated NVMe Block, Kubernetes-Native

Clients — Databases · Vector DBs · Kubernetes Pods · GPU VMs

NVMe-oF / TCPiSCSI Standard kernel nvme-tcp — no proprietary client

↓ ↓ ↓ ↓

MayaScale Block Engine Millions of IOPS · sub-millisecond · per-volume QoS · active-active HA

Server-side RAID10 RAID1 mirror across nodes for HA · client RAID0 stripe for throughput

Bypasses cloud disk API caps Connects over the VM network — limited by bandwidth, not the per-instance disk API

Storage Node A — local NVMeRAID1 mirror

Storage Node B — local NVMeRAID1 mirror

Manage & Deploy

Kubernetes CSIEKS · GKE · AKS — NVMe-oF block volumes

TerraformDeploy as code · Marketplace

Web GUImayagui console

CLI & APImayacli · REST

High-performance block storage, Kubernetes-native

MayaScale presents standard NVMe-over-TCP block devices and a Kubernetes CSI driver, so databases, vector stores, and stateful workloads get sub-millisecond block storage on any cluster — without a proprietary client and without the per-instance disk caps of native cloud disk.

Kubernetes-Native — CSI for EKS, GKE, and AKS

A CSI driver provisions block volumes as PersistentVolumes on Amazon EKS, Google GKE, Azure AKS, and self-managed Kubernetes. A single StorageClass pool parameter names the backend pool; the volume kind is auto-derived. Dynamic provisioning, snapshots, clones, online expansion, and per-volume QoS.

NVMe-over-TCP — Zero Client Footprint, No-Touch Networking

Clients connect with the standard in-kernel Linux nvme-tcp initiator — no proprietary agent, no special NICs, no switch configuration. Because storage is served over the VM network rather than the cloud's block-disk service, aggregate throughput scales with network bandwidth instead of the fixed per-instance limits native cloud disks impose.

Flexible Storage Pools — ZFS, LVM, or RAID

Choose the backend that fits the workload: a ZFS zpool mirror (FSx-style data services), an LVM volume group (thin or thick), or vanilla mdadm RAID1 volumes — selected per StorageClass, with the volume kind derived automatically.

Sub-Millisecond, 2M+ IOPS per Node

Server-side RAID10 — RAID1 mirroring across nodes for HA, client-side RAID0 striping for throughput — delivers 2M+ IOPS per node at sub-millisecond latency, scaling with added nodes and clients.

Per-Volume Guardrails, Sovereign

IOPS and throughput limits set per volume via StorageClass and enforced below the application — noisy-neighbor isolation for multi-tenant clusters. Everything runs on standard cloud VMs inside your own account; your data never leaves your cloud.

Ephemeral SSD Protection — HA, DR to Object, and Drain

Local NVMe is fast but ephemeral. MayaScale protects it in layers: a 2-node RAID1 mirror for drive/node failure, automatic replication to a cold pool on standard object storage (S3 / Blob / GCS) for DR, and a CSI drain service that evacuates volumes to the cold pool before planned shutdown or scale-in — so data survives instance stop and termination.

MayaScale Architecture

Architectural Attribute	Description
Disaggregated NVMe Block	NVMe-over-TCP block storage served from local NVMe on standard cloud VMs — compute and storage scale independently.
Kubernetes CSI Driver	Provisions block volumes as PersistentVolumes on Amazon EKS, Google GKE, Azure AKS, and self-managed Kubernetes.
No Proprietary Client or Agent	Uses the standard in-kernel Linux nvme-tcp initiator; the CSI node plugin is the only Kubernetes component.
Network-Attached Performance	Serves storage over the VM network (NVMe-oF/TCP) rather than the cloud block-disk service, so throughput scales with network bandwidth instead of the fixed per-instance IOPS/throughput limits native cloud disks impose.
Server-side RAID10	RAID1 mirror across storage nodes for high availability; client-side RAID0 stripe for aggregate throughput.
Flexible Storage Pools	ZFS zpool mirror (FSx-style), LVM volume group (thin/thick), or mdadm RAID1 — selected per StorageClass, volume kind auto-derived.
Cold Object Tier (DR)	Automatic replication to a cold pool on standard object storage (S3 / Blob / GCS) for disaster recovery and ephemeral-SSD durability.
CSI Drain to Cold Pool	A drain service evacuates volumes to the cold object pool before planned shutdown or scale-in, so data survives instance stop and termination.
Per-Volume QoS	IOPS and throughput guardrails set per volume via StorageClass, enforced in the storage layer.
Active-Active HA	Both nodes serve traffic; automatic failover with multipath for volume continuity.
Performance	2M+ IOPS per node at sub-millisecond latency; scales by adding storage nodes and clients.
Deploy on VM or Cloud	Runs as software on standard cloud VMs in AWS, Google Cloud, and Microsoft Azure.
Data Sovereignty	Runs in your own cloud account; data stays on your storage nodes, never a vendor backend.
Powers Zettabranch	MayaScale is the storage engine beneath Zettabranch branchable Postgres — a database per agent.

Kubernetes CSI & Data Services

CSI / Kubernetes	Description
Managed & self-managed	Amazon EKS, Google GKE, Azure AKS, and self-managed / k3s Kubernetes.
StorageClass pool parameter	Names the backend pool (ZFS mirror, LVM thin/thick, mdadm RAID1, or cold object); the volume kind is auto-derived.
Dynamic provisioning	StorageClass-driven creation of NVMe-oF block PersistentVolumes on demand.
Snapshots & clones	CSI VolumeSnapshots and writable clones for backups, dev/test, and fast restores.
Online expansion	Grow volumes in place without downtime.
Per-volume QoS	IOPS / throughput limits set per StorageClass, enforced below the application.
Access modes	Raw block and filesystem volumes; multipath for HA continuity across failover.

Data Protection & Durability	Description
RAID1 mirroring	Cross-node mirror for high availability and drive/node failure tolerance.
Active-Active HA	Both nodes serve traffic with automatic failover and multipath continuity.
DR to cold object pool	Automatic replication to a cold pool on standard object storage (S3 / Blob / GCS) for disaster recovery.
CSI drain to cold pool	Evacuates volumes to the cold object pool before planned shutdown or scale-in — ephemeral NVMe data survives instance stop and termination.
Encryption	At rest and in transit.

Workloads	Examples
Databases	PostgreSQL, MySQL, MongoDB, Cassandra — consistent sub-millisecond block.
AI / ML	Vector databases, KV-cache tiers, and training scratch / checkpoint volumes.
Kubernetes & platforms	Persistent volumes for any stateful workload on EKS / GKE / AKS; the engine beneath Zettabranch branchable Postgres.

Programs & Partnerships

Available on Amazon Web Services, Microsoft Azure, and Google Cloud Marketplace, and via Terraform. sales@zettalane.com