Forty machines, one Git repository, no one logging in by hand
Bare metal provisioning plane
- 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
Work
What changed, and by how much
Case studies in this industry are usually a logo strip and a paragraph about "digital transformation". These are organised by cluster type — 37 engagements across 22 capability domains — and each one ends with the numbers that moved. The client is anonymised rather than the number softened.
37 engagements documented
22 domains covered
11 with a measured before → after
0 client-named and authorised
How to read these
Every case carries a badge. Representative means it describes an engagement
pattern we deliver, anonymised and without invented metrics — we are not claiming a specific
named client. Client-verified means a customer has authorised publication
and you can ask for a reference call. We will tell you which is which before you ask.
4 engagements
Container orchestration
A cluster nobody dared upgrade, rebuilt so upgrades became routine
Kubernetes — self-managed HA control plane
- 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
Three major versions crossed, one workload restarted on purpose
Kubernetes — version upgrade programme
- 3 major hops completed in planned windows
- 1 workload restarted deliberately, none unexpectedly
- 0 rollbacks required
Nine teams, one cluster, and tenant boundaries that are enforced rather than agreed
Kubernetes — multi-tenancy & policy
- 4 enforcement layers, all producing audit evidence
- 0 shared admin credentials after migration
- 100% of namespaces with default-deny network policy
4 engagements
Virtualisation & VDI
A licensing renewal the board would not sign, turned into 140 VMs on open source
VMware → Proxmox VE + Ceph migration
- 4 waves, zero unplanned downtime
- 100% of waves had a rehearsed rollback
- 140 VMs migrated, all with rollback tested
Four nodes to eleven, with the storage rebalancing planned rather than survived
Hyperconverged cluster (Proxmox VE + Ceph / Harvester)
- 82% → 54% cluster utilisation after expansion
- 0 user-visible incidents during rebalancing
- 2 nodes found sharing a single power feed
A private cloud that tenants can self-serve, built to a failure-domain map
OpenStack private cloud
- 6 tenants migrated off departmental islands
- 0 API outages during three control-plane patches
- — per-tenant utilisation, now chargeable internally
300 desktop sessions without a per-seat licence, and a login time under 20 seconds
VDI — open source / self-hosted
- 17s p95 login time under simulated login storm
- 180 concurrent sessions supported on 4 nodes
- 0 per-seat virtual desktop licences purchased
1 engagement
High availability
Two nodes, real fencing, and a failover that finally happens without a human
Pacemaker + Corosync, DRBD replication
- 14s measured automatic failover under production load
- 0 manual interventions required
- 2 fencing paths validated per node
1 engagement
HPC & scheduling
Queue time cut from days to hours by fixing the scheduler, not by buying nodes
Slurm — HPC batch cluster
- days → hours p90 queue time for mid-size jobs
- 60 → 89% node utilisation
- 0 compute nodes purchased to achieve it
2 engagements
GPU & AI
Expensive accelerators sitting idle, made visible and fixed without buying more
GPU training cluster — utilisation & scheduling
- 31 → 78% GPU utilisation, before → after
- 0 additional GPUs purchased
- 1 fabric misconfiguration found by scaling tests
A model served at the latency the product needed, on the GPUs they already owned
LLM inference cluster
- 900/s requests per second sustained at target p95
- — time to first token, measured and published internally
- 0 additional GPUs purchased
3 engagements
Relational & distributed SQL
Eleven-minute outages became eight-second failovers, with restores nobody had ever timed
PostgreSQL — Patroni HA across three failure domains
- 8s measured automatic failover
- 14 min measured full restore from backup
- 100% of failovers now automatic
A single MySQL primary became a synchronous three-node Galera cluster with a tested cutover
Galera cluster (MariaDB / Percona XtraDB)
- 3 nodes synchronous, single-node failure tolerated
- 0 acknowledged data loss in the new design
- 2.4s measured cutover, executed in the window
Sharding exhausted at 4TB, replaced by a horizontally scalable SQL cluster
Distributed SQL — TiDB
- 12 → 1 application shards replaced by one logical cluster
- 0 application releases required to add capacity
- 6TB in production with no resharding work
3 engagements
NoSQL, cache & search
A replica set that outgrew one machine, split into a sharded cluster with a controlled balancer
MongoDB — sharded cluster
- 45k writes per second sustained after migration
- 0 user-visible slowness during chunk migration
- 1 shard key chosen against real query patterns, not defaults
Cache outages stopped taking the site down, on a cluster sized for real failover
Redis Cluster / Valkey
- 1 shard loss tolerated with no user-visible impact
- 8% of traffic traced to a single hot key, now sharded
- 0 sessions lost during failover drills
A search cluster that fell over at 60% load, rebuilt around shard planning
Elasticsearch / OpenSearch cluster
- 3,000+ → 240 shards after reindexing
- — p95 query latency measured before and after
- 0 index deletions by hand since the policy was set
2 engagements
Messaging & streaming
ZooKeeper retired without a single consumer message lost
Kafka — KRaft mode migration
- 0 messages lost across the migration
- 120 consumer groups verified by offset diff
- 1 dress rehearsal before production was touched
An event backbone across three sites, with leaf nodes that survive a site losing its link
NATS JetStream super-cluster
- 0 messages lost during a six-hour link outage test
- 20k/s telemetry messages sustained at peak
- 3 sites operating autonomously
1 engagement
Analytics & lakehouse
Twelve-hour reports became four-second queries on a real ClickHouse cluster
ClickHouse — sharded and replicated cluster
- 12h → 4s flagship report runtime, before → after
- 0 analytical load reaching the transactional replica
- 22B rows queryable interactively
4 engagements
Distributed storage
A Ceph cluster losing OSDs weekly, rebuilt around placement instead of hardware
Ceph — distributed storage cluster
- 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
An S3 API on your own hardware, with erasure coding and site replication that was actually tested
S3-compatible object storage (MinIO)
- 480M objects under management
- 1 site loss survived with no data loss, tested
- — cost per terabyte per year, modelled against cloud
NFS saturated at 4GB/s, replaced by a parallel filesystem that holds 38GB/s
Parallel filesystem (BeeGFS / Lustre class)
- 4 → 38 GB/s aggregate read throughput
- 64 compute clients served concurrently
- 1 storage target failure survived, job continued
The array was blamed for slow writes. The fabric was the problem.
SAN fabric — Fibre Channel, multipathing
- ~2× p99 write latency improvement
- 21 hosts found single-pathed by accident
- 0 storage purchased to fix it
1 engagement
File, block & object
A file server everyone depended on, replaced by a cluster with no visible failover
SMB / Samba scale-out cluster (CTDB)
- 0 user-visible interruptions during migration
- 3 nodes, any one of which can be lost without impact
- 60TB migrated with permission fidelity verified
1 engagement
Backup & DR
Backups that had never been restored, turned into a recovery time you can quote to the board
Backup cluster + off-site DR programme
- 0 → 14 systems with a measured and tested restore
- — measured RTO per tier, replacing a policy number
- 1 quarterly drill cycle, calendared and reported
1 engagement
Network devices & fabric
A firewall pair that had never failed over, and 41 devices with no configuration backup
Firewall HA cluster + network configuration management
- 2.8s measured stateful failover
- 41 devices under version control for the first time
- 9 devices years behind validated firmware
1 engagement
Load balancing & ingress
The load balancer was the single point of failure everyone had forgotten
L4/L7 load balancing — highly available pair
- < 1s measured VIP failover under load
- 140 backends now discovered, not hand-listed
- 0 services unreachable during the migration
1 engagement
Foundation network services
One DNS server was the whole company single point of failure. Now four, over two sites.
Anycast DNS cluster + DHCP HA
- 1 → 4 resolver instances, across two sites
- 0 resolution outages in the following two quarters
- 1 documented zone change process, replacing a phone call
1 engagement
Identity & secrets
Every server had local users. Now one directory, four replicas, and no shared root.
Identity & directory cluster (FreeIPA)
- 320 servers enrolled, no local admin accounts remaining
- 4 directory replicas across three sites
- 0 login outages during the batch rollout
1 engagement
Application & web
A PHP application that could not be deployed without downtime, now deployed any afternoon
LAMP / LEMP application cluster
- monthly → weekly release cadence in business hours
- 0 maintenance windows required for deployment
- < 5 min measured rollback to the previous release
1 engagement
CI/CD & artifacts
From a shared Jenkins box to a runner fleet, with builds that no longer queue at 9am
CI/CD platform + artifact registry
- 3,400/day pipelines with no queue at peak hour
- — median build duration, measured before and after
- 0 long-lived static credentials in pipeline config
1 engagement
Observability
Four monitoring tools, three of them lying. One cluster that everyone trusts.
Observability cluster (metrics, logs, alerting)
- 4 → 1 monitoring stacks after consolidation
- 9M active series with long-term retention
- — page volume per week, measured before and after
1 engagement
Security operations
A SIEM that could actually be queried during an incident, on infrastructure they own
SIEM / security analytics cluster
- 400 days retention, requirement-driven rather than disk-driven
- — query time over a 90-day window, measured before and after
- 2,400 endpoints reporting reliably
1 engagement
Edge & SD-WAN
Eighteen sites on one overlay, where a failed link is an event rather than an outage
SD-WAN overlay / multi-site interconnect
- 4 WAN links per site, all in active use
- < 1s measured link failover with no session loss
- 18 sites on one templated configuration
1 engagement
Multi-cluster & platform
Twenty-three clusters, one repository, and drift that gets corrected rather than reported
Multi-cluster fleet management
- 23 clusters under one declarative control plane
- 1 new cluster provisionable as a pull request
- — drift findings per week, trending down after the baseline
Before launch
If you have written authorisation to publish real customer names and logos, they get added
here — along with a link to what that customer was willing to say on the record. Until
then, this page stays anonymised. We do not invent clients, and we do not use stock
photographs of people who do not work here.
Want a reference call?
For a scoped engagement, we will arrange a call with a customer running something similar. They will not be hand-picked for enthusiasm — you can ask them what went wrong.