diff --git a/valkey/roles/valkey/handlers/main.yaml b/valkey/roles/valkey/handlers/main.yaml
index 3324657..0b32f7d 100644
--- a/valkey/roles/valkey/handlers/main.yaml
+++ b/valkey/roles/valkey/handlers/main.yaml
@@ -30,6 +30,12 @@
 # A new drop-in changes the generated unit, so Valkey has to be recreated to join
 # the network — Quadlet has no live equivalent of `podman network connect`.
 #
-# The cost is different from Postgres's: Valkey is a cache, so every attached app
-# sees a cold cache rather than a period without its database. Correctness is
-# unaffected; the first requests after the restart are simply slower.
+# The cost is lighter than Postgres's in one way and heavier in another. Lighter:
+# no app is left without its data. RDB persistence is on by default, the recreate
+# is a graceful stop, and valkey-server writes a final `dump.rdb` on the way down
+# that the new container loads back from the bind mount, so keys survive. Heavier
+# to reason about: two clients do not treat this as a pure cache — Nextcloud keeps
+# its PHP sessions and `memcache.locking` here, and Yamtrack uses it as its Celery
+# broker — so the seconds the container is down are seconds in which logins and
+# queued tasks are unavailable. Only a hard kill or a power loss costs changes
+# since the last save point. See readme.md, "The trade-off".