Cluster
Those are floor values for a working install, not a sizing exercise. Connector
workers run as pods in the same namespace and are the part that grows with your
workload — how many connectors, tables and subscriptions a single deployment
holds is covered in Sizing and limits.
Storage
An object store is not optional: the broker keeps no durable state anywhere
else, and there is no PostgreSQL retention — that backend was removed from
Kodansu and a
postgres:// engine is now refused. The alternative to S3 is
ADLS Gen2, which rules out only the optional stream browser — it has no ADLS
client. Setup is on the
Kubernetes install page.
Network
The control plane never reaches into your cluster: every exchange is initiated
outbound by the data plane. Control plane and data
plane documents exactly which metadata crosses
that boundary.
Reaching a source database that is not on the cluster’s network is a
separate problem, solved with a tunnel rather than a firewall rule — see
Tunnels and private connectivity.
From Popsink
Two things must be in your hands before you start:A registry token
A Google Artifact Registry service-account JSON. Every Popsink image is
pulled from a private registry and the control plane holds no copy of this
credential.
A control-plane account
On
https://control-plane.popsink.com, or on your own control plane. It is
what issues the deployment’s identity.Not on Kubernetes?
Single VM
One machine, no Kubernetes knowledge required. 4 vCPU, 16 GB RAM, ~30 GB
disk, plus an external database and S3 bucket.
Deployment options
SaaS, BYOC, on-premises and air-gapped compared.