A · Single node
Docker Compose
- One VM, all containers
- Pilot, demo, UAT
- Minutes to stand up
Deployment & scale
Four topologies from one image set, and an honest account of what grows as the programme does.
Deployment options
Four ways to run it. One image set.
The same containers run everywhere. Choose the topology that matches the stage — not a different product.
B · Container cluster
Kubernetes / OpenShift
- Pods autoscale (HPA)
- Rolling, zero-downtime
- Production default
C · High availability
Multi-AZ cluster
- No single point of failure
- Database primary + replica
- Enterprise scale
D · Sovereign
On-prem / air-gapped
- No public internet
- Private registry
- Regulated workloads
At a glance
| A · Single node | B · Container cluster | C · High availability | D · Sovereign | |
|---|---|---|---|---|
| Nodes | 1 | 3+ | 6+ (multi-AZ) | 3+ (private) |
| Scaling | vertical | horizontal (HPA) | horizontal + replicas | horizontal |
| Availability | single | rolling updates | active/active + failover | site-local |
What actually grows — and what doesn't
Petabytes stay in the source systems. Discovery samples and profiles data in place. Only the classification result, the lineage and the legal record are persisted, so platform storage stays in terabytes, not petabytes.
Volume model at organisational scale
- Consent records~20M subscribers × several purposes≈ 100M+ rows
- Audit logevery action on every record — the fastest grower≈ billions / years
- Classification resultssources × tables × columns≈ millions
- Rights requestsstatutory workload≈ 10⁴–10⁵ / year
- Evidence & exportsPDFs, certificates, packets≈ TBs (object store)
Scale-out playbook
- PostgreSQL, not SQLiteMandatory at this scale — SQLite is single-writer.
- Partition the audit logMonthly partitions, archived to object storage.
- Read replica for reportingDashboards and exports never touch the primary.
- Queue the heavy workScans and classification run on the worker tier.
- Autoscale on CPUWorkers scale with backlog; the API scales with traffic.
See your PDPL posture in a working system
Forty-five minutes, your own sample source, and a bilingual evidence pack you keep at the end of it.