The last failed publish of every sink under wal: :none, for Ankusa.Health.Check.Sinks.

The direct ack path (Ankusa.Edge.Publish) has no queue and no retry, so a broker that refuses every publish shows nowhere but in the 503s. It calls record/3 with each sink's outcome: a failure stores the monotonic time under the sink's identity (its module and options, hashed), a success removes the entry, so a sink that came back reads healthy at once instead of after the window. Both are one ETS write in the request process, no message to a singleton.

One process per instance owns the table and sweeps entries older than listener.ready_sink_error_window_ms every window, so the table holds at most the sinks that failed within it. It is also capped at @max_entries: a failure of a sink beyond the cap is not recorded (a node with that many failing sinks is unready already, and the sweep makes room as they age out).