fix(api): batch the message lookups behind message search (#2640)

deploy/self-hosting/.env.example

@@ -365,10 +365,16 @@ FLUXER_DISCOVERY_ENABLED=true
 #FLUXER_ERLANG_SCHEDULERS_MAX=16
 
 # In-flight request ceiling for the four services Compose forwards it to: the
-# users and messages routers and their shards. The Rust built-in defaults are 192
-# for messages, 320 for snowflakes and 64 elsewhere, and they govern every service
-# Compose does not forward this to.
-#FLUXER_SVC_MAX_CONCURRENT_REQUESTS=20
+# users and messages routers and their shards. Leave it unset and each service
+# uses its own built-in default, which is what the numbers below describe. Set it
+# and the one value replaces the built-in default on all four, so size it for the
+# busiest of them rather than for the smallest. The built-in defaults are 192 for
+# messages, 320 for snowflakes and 64 elsewhere, and they govern every service
+# Compose does not forward this to. A router holds a slot for the whole round
+# trip to its shard, so this is a ceiling on requests in flight at once and not a
+# rate: too low a value does not slow requests down, it rejects them, and the api
+# turns that rejection into a 503.
+#FLUXER_SVC_MAX_CONCURRENT_REQUESTS=192
 
 # The api and the Rust services name their fixed Postgres statement shapes so the
 # server can reuse their plans. Named prepared statements require a session that