Connect a network
Quickstart gets one connector running. This page covers doing it properly.
Egress requirements#
The connector needs outbound access only:
| Destination | Port | Purpose |
|---|---|---|
*.edge.proxility.scailabs.ai |
UDP/443 | Data plane. Preferred |
*.edge.proxility.scailabs.ai |
TCP/443 | Data plane fallback where UDP is blocked |
api.proxility.scailabs.ai |
TCP/443 | Enrolment and certificate renewal |
No inbound rules. If your egress filter is name-based, allow the two names
above; if it is address-based, proxility edge ranges prints the current
prefixes, and they are also published for automated allow-lists.
UDP is strongly preferred — the TCP fallback puts a reliable transport under a
reliable transport, and under loss the two congestion controllers interact
badly. If you are on the fallback, you will see it in proxility connector list and should fix the filter rather than live with it.
Run two connectors#
A production network gets two connectors on separate hosts, ideally in separate failure domains:
1 2 | |
Publish each service through both:
1 2 3 4 | |
Sessions distribute across available connectors. Losing one drops its in-flight connections — clients reconnect and land on the survivor. There is no state to fail over, because there is no shared session state.
Installation#
| Platform | Method |
|---|---|
| Debian/Ubuntu | apt install proxility-connector from the ScaiLabs repo |
| RHEL/Rocky | dnf install proxility-connector |
| Container | scailabs/proxility-connector:<version>, NET_ADMIN not required |
| Kubernetes | Helm chart scailabs/proxility-connector, one DaemonSet or Deployment per cluster |
| Windows | MSI from /downloads |
Builds are published on the ScaiLabs downloads page under Standalone apps & clients.
Upgrades#
Upgrade one connector of a pair at a time:
1 2 3 | |
A connector restart drops its established sessions. With a redundant pair and clients that reconnect, this is invisible; with a single connector it is a brief outage, so schedule it.
Retiring a connector#
1 | |
Revocation takes effect at the next handshake — the identity stops being trusted immediately rather than when its certificate expires. Remove the connector from any services it publishes first, or those services lose an endpoint.
Monitoring#
Expose the connector's metrics endpoint to your own scrape target and alert on:
| Signal | Why |
|---|---|
| Certificate renewal failure | The connector will stop working within 24 hours |
| Edge sessions below two | You are one failure from an outage |
| Transport = TCP fallback | UDP is being blocked somewhere |
| Policy-deny rate rising | Either a misconfiguration or someone probing |
Forward the same events to ScaiLog and let ScaiSentry alert on them if you do not want to run your own pipeline.