parallaxd: confirm an outage before alerting

A monitor in one location cannot tell the difference between “the service is down” and “my path to the service is broken.” parallaxd handles that ambiguity by checking normally from one place, then asking independent probers to corroborate a suspected failure. It alerts only when a quorum agrees.

Several viewpoints, one decision

Results are explicitly up, down, or unknown; an unknown result never becomes a vote. Checks declare whether they need a public or internal vantage, and optional provider-diversity rules stop several hosts at one provider from looking like independent evidence. A prober that loses contact with its peers is treated as isolated, so its results cannot create a misleading outage.

The probers only collect evidence. A coordinator evaluates quorum, persists the incident timeline, and alerts on state transitions instead of repeating the same notification on every failed check. Supported probes include HTTP, TCP, TLS, DNS, SMTP, gRPC, NTP, ICMP, and banner or request checks.

Failure modes are part of the design

Requests and results are signed with Ed25519, with an encrypted control network still recommended for confidentiality. An external watcher detects a silent coordinator, and an optional warm standby can be promoted only after the old primary is fenced. That last step is intentionally manual: losing contact with a primary is not proof that it has stopped.

The project includes an operator dashboard, durable webhooks, native Tintwire alerts with fallback, a redacted status export, Ansible deployment, and documented acceptance and failover drills. A useful production layout starts with one coordinator and at least three probers in genuinely different failure domains.

Source and documentation: github.com/kilo666mj/parallaxd.