CUSTOM BUILD · STORAGE

Storage clusters that survive in production.

Ceph, all-flash, hybrid, or pure capacity tier — we design storage clusters around your workload, not around a vendor's reference architecture. Hardware, software, network fabric, integrated and tuned end-to-end. From a single 24-bay node to multi-petabyte clusters running thousands of VMs.

WHY US

Three things that matter at petabyte scale.

Storage is where the wrong decision shows up two years later, not two weeks. We've operated clusters long enough to know what the second-year failure modes look like.

▸ 01 / EXPERIENCE
Production Ceph & large-scale storage
Multi-petabyte Ceph clusters serving thousands of live VMs. We've handled the scrub storms, the OSD reweights, the rebalances that don't end. We know what survives.
▸ 02 / DEBUG
Hardware + software debug capability
When latency spikes mysteriously, when a disk firmware causes silent corruption, when network microbursts cause OSD flapping — we troubleshoot the full stack instead of swapping random parts.
▸ 03 / HA DESIGN
High-availability by design
Failure domains spanning racks, replication or erasure coding sized to actual recovery time, dual fabric, monitoring that catches problems before they become incidents. HA designed in, not bolted on.
WHAT WE BUILD

Storage architectures we have shipped.

Every architecture below has a reference design we've put into production. We start from a known-good baseline and tune for your workload — performance, capacity, durability, or the right blend.

▸ CEPH BLOCK / OBJECT
Multi-PB Ceph clusters for VM and object workloads
RBD for hypervisor pools, RGW for S3-compatible object. Replication or erasure coding sized for your durability target. NVMe metadata, HDD or SSD data tier as workload demands.
▸ CAPACITY-DENSE NODES
24-bay / 36-bay storage servers
Supermicro or equivalent chassis with 22T / 24T enterprise HDDs, NVMe boot and metadata. Designed for cold/warm capacity tiers and large-scale archive use cases.
▸ ALL-FLASH PERFORMANCE
NVMe-only performance tier
Full-flash nodes for IOPS-bound workloads — VM boot pools, database tiers, GPU training data feeds. PCIe topology mapped so NVMe drives don't share bottlenecks.
▸ HIGH-SPEED FABRIC
100G / 200G storage networks
Storage networks designed for the cluster, not retrofitted around a generic data center fabric. Dual-plane redundancy, RDMA where it pays off, congestion control tuned for storage traffic.
SHIPPED PROJECTS

What we've put into production.

Two storage clusters operating live today — one running thousands of VMs in a major Hong Kong IDC, one a multi-petabyte cluster scaled with us through two generations of fabric.

▸ CASE STUDY · AKARI · HONG KONG IDC

Ceph cluster running 4,000+ production VMs in a major HK datacenter.

Akari operates one of Hong Kong's larger IDCs. Their virtualization platform requires a storage backend that can handle thousands of VMs concurrently — boot storms, mixed read/write profiles, snapshots, live migrations — without latency excursions that translate into customer-visible problems.

Scope: Ceph cluster design, hardware spec across multiple node types (block, object, metadata), failure-domain layout aligned to physical rack topology, performance tuning for hypervisor I/O patterns, ongoing capacity planning as the VM fleet grew.

4,000+
Production VMs running on cluster
Ceph
Block + object backend
Major IDC
Hong Kong production
▸ CASE STUDY · 39AI · LARGE-CAPACITY CLUSTER

Multi-petabyte capacity cluster, scaled across two fabric generations.

39ai's storage cluster grew with the AI workload. We built the original 12-node deployment on a 25G fabric, then scaled it out and upgraded the fabric to 100G as throughput demands climbed.

Today's footprint: 14 × Supermicro 24-bay chassis, mixed disk loadout — 120 × 22TB drives plus 9 × 24 × 24TB drives for the high-density tier. Full 100G inter-node fabric. The same team designed the original cluster, the scale-out, and the fabric migration.

▸ HARDWARE FOOTPRINT 14× Supermicro 24-bay storage nodes  ·  120× 22TB enterprise HDDs  ·  9 × 24× 24TB drives (high-density tier)  ·  100G inter-node fabric  ·  upgraded from 12-node 25G baseline (also our build)
~5+ PB
Raw capacity
14 nodes
Current cluster size
25G → 100G
Fabric upgrade in place

Storage that needs to actually work in year three?

Cluster design proposal within 5 days · Reference architecture available on request