Several replicas

Run rws on several machines behind a load balancer and keep runtime changes — the IP blocklist, maintenance mode, drained backends, feature flags and cache purges — the same on all of them through Redis.

Most of rws's configuration is the config file, so replicas agree by running the same file. Changes made at runtime through the admin API, rws ctl or the MCP tools are different: on their own they stay on the replica that took them. Point the replicas at one Redis (since 18.22.0) and those changes reach every replica within about a second:

[cluster]
redis  = "10.0.0.20:6379"     # defaults to [acme] redis when that's set
# redis_password = "…"        # or RWS_CONFIG_CLUSTER_REDIS_PASSWORD (never printed)
# prefix = "rws:cluster:"     # key prefix, when several clusters share a Redis
# node   = "web-1"            # this replica's name (default: the host name)

Shared: the IP blocklist, maintenance mode, drained upstream backends, feature flags, and purges of named caches (each replica purges its own). Not shared: anything in the config file (deploy the same file), ACME certificates (use [acme] redis for those, which can be the same Redis), rate-limit counters, sessions.

rws ctl --admin web-1:9901 block 203.0.113.9     # refused on every replica
rws ctl --admin web-2:9901 status
# cluster   node web-2 · Redis up · 3 node(s): web-1, web-2, web-3

How it works: one JSON document in Redis (<prefix>state) holds the shared state and a version number. A change applies on the replica that took it, then a Lua script applies the same change to the document inside Redis, so changes made at the same moment on different replicas don't overwrite each other. Every replica reads the document once a second and makes its own state match when the version moved. Each replica also writes a heartbeat (<prefix>nodes); GET /admin/api/v1/status lists the replicas seen in the last minute.

When Redis is unreachable, changes still apply on the replica that took them, rws ctl status shows them as waiting, and they're sent when Redis is back. A Redis that comes back empty is filled again from the replicas' own state, so nothing they agreed on is undone. A replica started later adopts the document. Audit-log entries name the replica (node).

Redis needs Lua scripting (every Redis since 2.6, Valkey, KeyDB, managed Redis services); run it on a private network or with a password.