Is it a fit?

radiusward makes one trade on purpose: accounting is acknowledged only when it is safe on a majority of three members. That buys durable records and clean failovers, and it costs a cluster to run. Whether that trade is right depends on what your accounting records are worth to you.

Consider radiusward when

• Accounting records matter. They feed billing, usage limits, reconciliation or compliance, and losing the records of the last seconds before a crash — or counting them twice after a failover — is a real problem.
• You want to verify before you trust. The accounting contract, the failure scenarios and the tests behind them are published, and you can run them yourself.
• You can run three members and two gateways. On Kubernetes that is a Helm release; on hosts, five services.
• You run Java infrastructure, or want extension points in Java: identity stores and export destinations are plug-ins.

Evaluate carefully when

• You need the simplest possible installation. A single RADIUS server process is simpler to run than a cluster. If losing a few seconds of accounting on a crash is acceptable, you may not need replication.
• You need device certifications or a support contract today. radiusward has interoperability tests with other implementations, not certifications, and no commercial support yet.
• You depend on another RADIUS server’s modules or configuration format. radiusward has its own configuration and is not compatible with them; a migration is a re-configuration, best done in a lab first.
• Your members would sit on high-latency or unreliable links. Every acknowledgement waits for a majority, so the round trip between members is part of the answer time.

What it costs to run

Processes	three cluster members and at least two gateways
Memory	about 1 GiB per member in the small profile, which holds two million sessions
Disk	each member keeps the replicated log and snapshots; fsync is on
Network	members reach each other directly; each NAS is configured with two gateways
Operations	metrics, dashboards and alert rules come with it; a continuous backup runs outside the cluster

Not sure?

Run the quickstart on a laptop, kill the leader while it carries load, and look at what your NAS would have seen. That is what the project does every night on Kubernetes.