We write kernel modules, consensus protocols and parallel filesystems, and nothing is considered done until it has survived the failures production will throw at it.
Sources: evidence ledger, below. Numbers from our current client work.
Evidence ledger
We show our work so you don't have to take our word for it. This is what we built, how it was checked, and in which engagement, so your engineers can review it as they would a design.
| Claim | What we built | How it was checked | Engagement |
|---|---|---|---|
| Data survives a node failure mid-write | Chunk-level replication, erasure coding and recovery for distributed block storage | fio writes CRC32c-verified blocks, a storage node is killed mid-write, recovery runs, every checksum is re-validated in CI | Niova |
| HA metadata survives a second failure | Fixes for metadata-server crashes and buddy-resync failures in mirrored deployments | Stress run: baseline, kill primary, kill secondary mid-resync, re-run, compare before/after graphs | BeeGFS / ThinkParQ |
| Upgrades without downtime | Kernel compatibility shim through Linux 6.12, idmapped mounts, a parallel kernel module for two major versions | Live-upgrade scenario in the release test suite: both versions mount side by side, rollback is safe | BeeGFS / ThinkParQ |
| Consensus holds under partition | Raft-replicated, RocksDB-backed metadata service with leader election and leases | Protocol logic model-checked with TLA+/TLC; partition scenarios encoded in multi-node system tests | Niova |
| Releases that build everywhere | 20 OS × arch release builds on native ARM and x86 runners, GPG-signed packages | Support matrix auto-discovered from upstream distro repositories, so it never goes stale | BeeGFS / ThinkParQ |
| Performance measured at scale | IOR / mdtest pipelines up to 6M files at 90% capacity; IO500 automated end to end | Runs in the same pipeline as fault injection, so throughput and recovery are measured together | BeeGFS / ThinkParQ |
| Operable from day one | 15+ Grafana dashboards and installable alerting rules shipped with the product | Metadata, client operations by node and user, and storage targets covered | BeeGFS / ThinkParQ |
| GPU data path validated | CI builds and tests against NVIDIA GPUDirect Storage | Automated DOCA compatibility matrix per distro and kernel | BeeGFS / ThinkParQ |
Method
The same four steps on every engagement, whether we're writing a storage engine or tuning a Lustre deployment. We spend the most time on the third step.
Protocol-critical logic is specified and model-checked before it becomes C.
TLA+ · TLC
Kernel modules, storage engines and control planes in C, C++ and Go.
io_uring · ublk · Raft
Real multi-node clusters, exact faults injected while I/O is in flight.
Ansible · pytest
Throughput and recovery measured in the same run, traced to the kernel function.
IOR · IO500 · eBPF
Where it was proven
Niova
BeeGFS / ThinkParQ
“We have had the pleasure of working closely with Manisha and the team at Paroscale Technologies for the past few months, going straight into forward development with very little time to adopt the development environment, as well as the code base for our parallel file system BeeGFS. The meticulous work that Paroscale has already provided, exceeded our expectations. It is a pleasure to have Paroscale as a partner.”
Start here
Topology, workload and one run is enough. An engineer replies with where the time is going and what we'd test first.