quorate
+86 130 5181 3099 botonwa@gmail.com
Book a 30-min review
Book a 30-min review +86 130 5181 3099 botonwa@gmail.com
Cluster engineering · Silicon Valley

Any cluster.
Debian to exascale.
Built, hardened, and on-call.

We deploy, harden and operate the clusters your business actually runs on — the network fabric and firewalls underneath, then Linux and Kubernetes, databases, distributed storage, SAN, messaging, HPC and GPU on top. You keep ownership and control. We build it, prove it works, teach your team, and then answer the phone at 3am.

Book a 30-min cluster review See all 23 disciplines

No sales team. Talk to an engineer: +86 130 5181 3099 · botonwa@gmail.com

  • Build · Managed · Audit
  • 233 cluster categories indexed
  • Fixed-price audits from $2,500
  • You own the hardware and the code
23
cluster disciplines covered
233
cluster categories indexed
—
nodes under management
—
GPU cards deployed
< 15 min
median first response, managed tier

Placeholder 2 figures above need real numbers before launch. We publish measured values or nothing — we do not put estimates in a trust bar.

Capability matrix

Twenty-three cluster disciplines. One team.

Most infrastructure shops do one of these well. The problems that hurt are the ones that span several — a database that depends on storage, that depends on the SAN fabric, that depends on the network, that depends on a scheduler nobody tuned. We work across all of it, and we say so plainly when something is outside our lane.

M

Network clusters (L2/L3)

Firewall pairs that actually fail over on purpose, switch stacks and MLAG pairs that neither loop nor blackhole when a member dies, and routers whose routing convergence you have measured instead of assumed.

  • EVPN-VXLAN / MLAG
  • FRR + BGP · VRRP
  • Fortinet FGCP / PAN-OS HA
How we deliver it
N

SAN & storage networking

Dual-fabric SANs with zoning a human can read, multipathing proven by actually pulling a path, and NVMe-oF over lossless Ethernet that does not drop a frame under load.

  • FC zoning · dual fabric
  • NVMe-oF (RDMA / TCP)
  • dm-multipath · ALUA
How we deliver it
O

Foundation network services

The three services nobody capacity-plans for, and the first three that page the whole company at the same minute when they fail.

  • BIND 9 · Knot · PowerDNS
  • Kea DHCP HA
  • chrony · PTP grandmaster
How we deliver it
R

SD-WAN & multi-site

Every site is a small cluster with one big dependency: the link. We design for the link being gone, not for it being fast.

  • WireGuard overlay
  • FRR + BFD sub-second failover
  • Policy-based breakout
How we deliver it
A

Bare metal & OS layer

Twenty machines arrive. In fifteen minutes each they are identical, provisioned from Git, and ready to join a cluster — with nobody logging in by hand.

  • MAAS / iPXE
  • Ansible + AWX
  • Talos / Flatcar
How we deliver it
B

Kubernetes & container platform

A cluster you can upgrade without holding your breath, whose etcd backup you have actually restored, and which you could rebuild from code in under an hour.

  • kubeadm / RKE2 / Talos
  • Cilium + eBPF
  • Argo CD GitOps
How we deliver it
C

High availability & failover

Most HA clusters have never actually failed over. We build them so the failover works, and then we prove it by pulling the plug.

  • Pacemaker + STONITH
  • Keepalived + VRRP
  • DRBD 9 / LINSTOR
How we deliver it
D

Load balancing & ingress

Traffic enters your infrastructure once, and it enters through a design you can explain, test with real numbers, and renew certificates without anyone waking up.

  • LVS / IPVS
  • HAProxy + Envoy
  • Gateway API / Cilium
How we deliver it
P

Directory, identity & secrets

One directory that can answer "who has access to what", replicated across sites, with certificate and secret lifecycles that do not depend on a calendar reminder.

  • FreeIPA / 389 DS
  • Keycloak · OIDC / SAML
  • Vault + Consul
How we deliver it
E

Distributed storage

Ceph, BeeGFS and MinIO clusters designed around your real failure domains, benchmarked with fio, and proven to survive a disk, a node, and a rack going away.

  • Ceph + Cephadm
  • BeeGFS / Lustre
  • Rook on Kubernetes
How we deliver it
F

Database clusters

The cluster that keeps your business alive. Planned failover in seconds, backups you have actually restored, and upgrades performed while the application keeps writing.

  • Patroni + etcd
  • Galera / Group Replication
  • ClickHouse + Keeper
How we deliver it
G

Caching & messaging

Kafka, Pulsar, NATS and Redis clusters that survive a broker loss without losing a record, and that tell you about consumer lag before your customers do.

  • Kafka KRaft
  • Strimzi on K8s
  • Redis Cluster
How we deliver it
H

Big data & lakehouse

Spark, Flink, Trino and an open table format on storage you control — so analytics stops being a separate, expensive, fragile universe.

  • Spark on K8s
  • Trino / Iceberg
  • Flink on K8s
How we deliver it
T

File & block services

The file services your users actually touch every day, made redundant without changing a single path or permission they are used to.

  • clustered Samba · CTDB
  • NFS-Ganesha · pNFS
  • CephFS / DRBD backends
How we deliver it
I

HPC & AI / GPU clusters

Your GPUs are expensive and mostly idle. We fix the part that wastes them: the scheduler, the fabric, the storage feeding them, and the way jobs queue.

  • Slurm + Slinky
  • NVIDIA GPU Operator + DRA
  • vLLM / KServe
How we deliver it
J

Observability clusters

Metrics, logs and traces that survive a node loss, alert on things that matter, and cost less than the infrastructure they watch.

  • Prometheus + Thanos
  • Loki + OpenTelemetry
  • OpenCost
How we deliver it
U

VDI & desktop virtualisation

Centralised desktops judged on one number: how long a login takes during the morning rush, when everyone arrives at once.

  • Proxmox VE + Kasm Workspaces
  • golden image pipeline
  • profile redirection
How we deliver it
S

Application & web clusters

Deployments that happen on a Tuesday afternoon because the application tier holds no state — which is the actual prerequisite, not the deployment tool.

  • nginx + PHP-FPM
  • Tomcat · WildFly · JVM clustering
  • Redis sessions · object storage
How we deliver it
V

CI/CD & artifact platforms

Your pipeline is production infrastructure now. It gets the same availability design, backup and monitoring as anything customer-facing.

  • GitLab Runner fleet
  • Harbor HA registry
  • Argo CD + Vault
How we deliver it
K

Virtualization & HCI

Licence costs that do not triple every renewal. Proxmox VE, Ceph, KubeVirt and OpenStack, with a migration plan that includes a tested way back.

  • Proxmox VE + Ceph
  • Harvester / KubeVirt
  • OpenStack Kolla
How we deliver it
L

Fleet, DR & security

The layer that holds every other cluster type together: fleets that deploy the same way everywhere, DR that has been rehearsed, and security that produces evidence.

  • Karmada / Fleet
  • Velero + witness DR
  • Vault + Cosign
How we deliver it
Q

Backup & disaster recovery

Recovery tiers with a measured RTO, an off-site copy that cannot be encrypted by ransomware, and a restore drill the board can read.

  • Proxmox Backup Server · Bareos
  • Velero · Kasten
  • Immutable off-site vault
How we deliver it
W

Security operations clusters

Security tooling you cannot query under load is security tooling nobody uses. Indexing and retention are capacity problems before they are policy problems.

  • Wazuh · OpenSearch
  • Suricata / Zeek
  • hardening baselines (CIS)
How we deliver it

The cluster atlas — all 233 categories, and where we stand on each

How we work

Five stages. Every one of them produces a document you own.

No stage ends with "and then we'll see". Each one has a defined output and an exit criterion, and you get the artifacts whether or not you continue to the next stage.

  1. 01

    Discover

    1–2 weeks

    We read your architecture, not your marketing deck.

    • Current-state risk map (including single points of failure)
    • Baseline measurements: latency, IOPS, GPU utilisation, cost
    • Prioritised findings with effort/impact
  2. 02

    Design

    1–3 weeks

    Every decision gets written down, with the alternative we rejected and why.

    • Reference architecture
    • Failure-domain analysis
    • RTO/RPO targets you can hold us to
    • Bill of materials with licence comparison
  3. 03

    Build

    2–10 weeks

    Everything is described in code. Nothing is configured by hand at 2am.

    • Ansible / Terraform / GitOps repository
    • Cluster built to the approved design
    • Handover runbooks
    • Team walkthrough session
  4. 04

    Harden

    1–2 weeks

    We break it on purpose so it doesn't break by accident.

    • Chaos drill report: real failover, real timings
    • Restore-from-backup proof, timed
    • CIS baseline scan results
    • Remaining risk register — the honest one
  5. 05

    Operate

    Ongoing

    The part most vendors skip.

    • 7×24 on-call with named engineers
    • Quarterly failover + restore drills
    • Capacity and cost review each quarter
    • Patch and upgrade windows you approve

Read the full process, stage by stage

Technology

We name versions, not categories.

"Container orchestration" means nothing. Cilium 1.18 with Gateway API, Rook-Ceph 1.17 and DRA on Kubernetes 1.34 means something. Here is a sample of what we work with weekly — 295 entries are indexed on the stack page.

  • Kubernetes
  • Ceph
  • EVPN-VXLAN
  • PostgreSQL
  • Patroni
  • Kafka
  • Slurm
  • FRR
  • Talos Linux
  • Cilium
  • Fortinet HA
  • NVMe-oF
  • Proxmox VE
  • Loki
  • TiDB
  • Fibre Channel
  • BeeGFS
  • vLLM
  • NetBox
  • Vault
  • Argo CD
  • Envoy
  • ClickHouse
  • MinIO
  • MLAG
  • NATS
  • Iceberg
  • Kyverno
  • HAProxy
  • Pacemaker
  • DRBD
  • dm-multipath
Browse the full stack index
Representative work

What changed, and by how much.

Not "we delivered a solution". Here is the situation, what we did, and the number that moved afterwards. Six of the 37 engagements we have documented, organised by cluster type.

Anonymised Representative engagements described under NDA: no client names, no logos, no invented metrics. Where a number appears, we measured it and the customer saw the report.

Research institute · North America

Forty machines, one Git repository, no one logging in by hand

Bare metal provisioning plane

Challenge
The team could not safely reinstall a node — the procedure existed only as tribal knowledge — so a failed disk meant a host down for days.
Approach
Replaced the wiki page with a provisioning plane and a machine inventory as the single source of truth. Rack-by-rack roll-through with two spare nodes as a buffer, starting with the least critical rack.
Delivered
  • MAAS as the authoritative machine inventory, synchronised nightly from Redfish
  • iPXE + preseed bring-up with automated RAID, bonding and firmware baseline
  • Ansible roles for OS baseline, network config, package policy and unattended upgrades
Measured result
  • 12 min per-node bring-up, measured on all 40
  • 100% of machine state now described in Git
  • 0 interactive logins required for a new node
Read the full case
B2B SaaS · North America

A cluster nobody dared upgrade, rebuilt so upgrades became routine

Kubernetes — self-managed HA control plane

Challenge
Nothing about the cluster was reversible. Every change was a one-way door, so the team made changes less and less often.
Approach
Built a staging cluster from an approved design first and proved the upgrade path there, then rebuilt production the same way — GitOps from the first commit so the cluster could be recreated from code.
Delivered
  • HA control plane across three failure domains, external etcd with timed restore
  • GitOps from day one — every workload, policy and add-on in a repository
  • Cilium network policies replacing a convention-based multi-tenancy README
Measured result
  • 49 min measured control-plane upgrade, zero dropped requests
  • 4 min measured etcd restore to a working control plane
  • 0 plaintext secrets remaining in Git
Read the full case
Manufacturing · Midwest US

Two nodes, real fencing, and a failover that finally happens without a human

Pacemaker + Corosync, DRBD replication

Challenge
An untested HA cluster is worse than no cluster: it produces confidence without capability.
Approach
Fixed the fencing path first — because everything else depends on it — then rehearsed failover under load, including the case where the primary is unresponsive rather than cleanly down.
Delivered
  • Fencing validated against both power and watchdog paths, with SBD timeouts tuned
  • DRBD replication verified under load with a written split-brain recovery procedure
  • Resource ordering and constraints rewritten so failover is deterministic
Measured result
  • 14s measured automatic failover under production load
  • 0 manual interventions required
  • 2 fencing paths validated per node
Read the full case
Media production · EU

A licensing renewal the board would not sign, turned into 140 VMs on open source

VMware → Proxmox VE + Ceph migration

Challenge
The migration had to be reversible per wave, and no application owner would accept a weekend of unplanned downtime.
Approach
Modelled the replacement platform first, then migrated in four batches ordered by blast radius — starting with internal tooling so the team could make its mistakes where they were cheap.
Delivered
  • Hyperconverged platform with SSD and HDD pools and a documented failure-domain map
  • Per-wave rollback rehearsal: every VM restorable to the old platform within the window
  • Licence cost comparison table for the board, including support and training lines
Measured result
  • 4 waves, zero unplanned downtime
  • 100% of waves had a rehearsed rollback
  • 140 VMs migrated, all with rollback tested
Read the full case
Financial technology · North America

Eleven-minute outages became eight-second failovers, with restores nobody had ever timed

PostgreSQL — Patroni HA across three failure domains

Challenge
The database held the ledger. Both the availability story and the recovery story had to become measured facts rather than confident assumptions.
Approach
Rebuilt the consensus layer across three genuinely separate failure domains, then ran the first real restore into a scratch environment and published the time.
Delivered
  • Patroni-managed cluster with etcd quorum spanning three failure domains
  • Synchronous commit policy per workload class, documented trade-offs
  • pgBackRest with off-site storage, WAL archiving and retention policy
Measured result
  • 8s measured automatic failover
  • 14 min measured full restore from backup
  • 100% of failovers now automatic
Read the full case
Broadcast & media · North America

A Ceph cluster losing OSDs weekly, rebuilt around placement instead of hardware

Ceph — distributed storage cluster

Challenge
The problem was not failing disks. It was a CRUSH layout and recovery configuration that turned one disk failure into a cluster-wide performance event.
Approach
Rebuilt the failure-domain model first, tuned recovery and backfill against measured client latency, then rolled the layout out rack by rack without taking the cluster offline.
Delivered
  • Failure-domain map rebuilt: rack-level CRUSH rules replacing a flat layout
  • Recovery, backfill and scrub settings tuned to a client-latency budget
  • NVMe WAL/DB separation for the HDD tier, which removed a hidden write bottleneck
Measured result
  • weekly → 0 degraded-state events in the following quarter
  • — client latency under recovery, measured against budget
  • 1.1PB raw capacity with no data migration required
Read the full case

All 37 case studies, by cluster type

Managed & SLA

Published tiers. Published starting prices.

Most vendors in this space quote on request, which tells you the number depends on how much they think you can pay. Ours depends on cluster size, failure domains, and whether you want us on call. Here is the shape of it — the exact figure comes out of a short call.

Essential

99.9% uptime target

from $1,900 /cluster/month

  • Monitoring stack deployed and tuned
  • Business-hours incident response (PT)
  • Monthly patch cycle
  • Quarterly capacity report
Not included
  • Out-of-hours response
  • Failover drills
Most chosen

Business

99.95% uptime target

from $3,600 /cluster/month

  • Everything in Essential
  • 16×5 incident response
  • Quarterly failover drill with report
  • Restore-from-backup drill
  • Named engineer, direct line
Not included
  • Contractual 15-minute response

Mission-Critical

99.99% uptime target

from $7,900 /cluster/month

  • Everything in Business
  • 24×7 follow-the-sun response
  • Contractual 15-min response target
  • Monthly chaos drill
  • Dedicated runbook ownership
  • Capacity headroom guarantee

GPU / HPC node management

from $220 /GPU node/month

Priced per node, not per GPU card. Includes driver/fabric health, scheduler tuning, and utilisation reporting.

Retainer / emergency support

from $4,000 /month

A prepaid hour pool you can drain anytime. For teams who build themselves but want a senior engineer one message away.

Full pricing, including what changes it

Where you probably are

Six situations, six ways in.

Find the one that sounds like your week. Each links to the discipline that usually applies — though the honest answer often spans two of them.

The network is the thing nobody dares touch

A firewall pair that has never been failed over on purpose, a switch stack nobody fully understands, and no configuration backup in Git.

Measured failover drills, MLAG that survives a member loss mid-day, and every device config version-controlled and reversible in one command.

Network & firewall clusters

Your cloud bill is outgrowing your revenue

Kubernetes running core business on someone else's metal, $40k–$400k a month, and no clear path down.

Move to bare metal or colocation, keep the Kubernetes experience, cut the platform cost.

Bare-metal Kubernetes

You own racks, and the clusters keep breaking

Servers, VMware, a database with a single point of failure and no dedicated SRE to watch it.

Cluster rebuild, real high availability, managed operations, and training for the team you already have.

HA & failover engineering

You bought GPUs and the utilisation is embarrassing

A cabinet of accelerators, unstable multi-node training, and no way to prove where the idle time went.

Slurm or Kubernetes, fabric and storage that keep up, and a measured utilisation report.

HPC & AI clusters

Your virtualisation renewal just doubled

VMware costs your CFO will not approve, and a migration that still sounds like a science project.

Assessment, then Proxmox VE + Ceph or KubeVirt in migration waves — each with a rehearsed rollback.

VMware migration

Storage is slow and the array is taking the blame

Two fabrics on paper, hosts quietly single-pathed, multipath never failed over, and a p99 latency nobody can explain.

Fabric and path audit, zoning generated rather than hand-typed, and a failover drill that produces a number instead of an argument.

SAN & storage networking
Who you'll actually talk to

Engineers, not account managers.

The person who scopes your engagement is the person who builds it. There is no handover to a delivery team you have never met, and no one between you and the engineer who understands your cluster.

Network engineer

Spine-leaf fabrics, firewall clusters, routing, network automation

Storage network engineer

Fibre Channel, NVMe-oF, multipath, lossless Ethernet

Platform engineer

Bare metal, OS layer, provisioning, Kubernetes lifecycle

Database engineer

PostgreSQL, MySQL, distributed SQL, backup and restore

Storage engineer

Ceph, parallel filesystems, object storage, failover drills

HPC / AI engineer

Slurm, GPU scheduling, fabric, inference serving

Placeholder Named engineers with photos, certifications and public contributions go here before launch. We're not inventing faces to fill the space — this section stays as roles until there are real people to put in it.

Free · no call required

The 38-point Cluster Health Checklist

The exact list we work through during a paid audit: quorum and fencing, storage failure domains, backup restorability, upgrade paths, certificate expiry, GPU allocation idleness, and the eleven things that most often turn out to be wrong.

  • 38 items, each with what "good" looks like and how to verify it
  • Written for engineers, not for procurement
  • Includes the questions we ask that usually make people go quiet

Four fields, that's it. Or skip the form and email botonwa@gmail.com directly.

Questions we get asked

The ones that usually decide it.

How are you different from Rancher, Spectro Cloud or Rafay?

They sell software platforms. You install the platform and you're on your own for the hard parts — that's the single most common complaint in their own reviews. We're the team that does the hard parts, on your hardware, with your constraints. We'll happily deploy their platform if it fits; we just don't stop at the licence email.

We already have racks and a VMware estate. Can you work with what we own?

That's most of our work. We start with an audit of what exists — hardware inventory, firmware, network topology, failure domains, licence exposure. You get a risk map before you commit to anything. If the honest answer is 'keep most of it and fix four things', that's what we'll tell you.

Where does our data live?

On your infrastructure. We don't proxy your traffic and we don't copy your data to our cloud. Our access is through infrastructure you grant, logged, and revocable. For managed customers we need monitoring and log access — we'll document exactly what and why, and you can scope it down.

How fast do you respond when something breaks?

Managed tiers carry contractual response targets: business hours on Essential, 16×5 on Business, 24×7 with a 15-minute target on Mission-Critical. Median first response on the managed tier is tracked and published on our status page. We don't publish an SLA we haven't measured.

Does our team need to know Kubernetes already?

No, but we won't pretend it's magic either. Every engagement ends with runbooks and a walkthrough session. If you want your people trained, the 3-day cohort on your own hardware is the fast path — we'd rather leave you capable than dependent.

What does a project actually cost, and what makes it cost more?

Audits start at $2,500 fixed. Build engagements start at $15,000. Managed starts at $1,900 per cluster per month. Things that move the number: failure domains (multi-site costs more than multi-rack), regulated environments, air-gapped networks, and migrations with a live cutover. We list those multipliers on the pricing page rather than hiding them behind a quote form.

Do you take equity or work for deferred payment?

No. We're a services business and we're paid in money. If you need the audit but can't fund the build yet, take the audit report and shop it around — it's yours.

Book a 30-min cluster review

Bring your architecture diagram, or just the problem. We'll tell you what we'd do, what it would cost, and whether we're the right people for it.

Book a 30-min review See pricing

botonwa@gmail.com · +86 130 5181 3099