A transient failure pushed the next attempt out by the full refresh interval. Failures now retry in five minutes, with a server Retry-After still honoured as a ceiling.
The four-way GROUP BY over node_short_ids ran once per 1000-route batch, up to 750 times a run, and could eat the whole 5s batch budget. The task now fetches the set once and passes it to each batch.
Periodic ticks now start five minutes off the task tickers so the hourly reconfirm trigger stops colliding with the cooldown, any capture clears a deferred periodic, subdirectories no longer stop the recorder, and the metadata sidecar is published before the profile.
Sessions are clean and client IDs random, so deferring acks never bought redelivery; it only capped delivery at the broker's inflight window and pushed backlog into the broker queue, where drops are silent.
MarkSent ran after the side effects, so a broker copy on another worker could win it and be streamed as a repeat of the first hearing. It now runs before any further DB work, unconditionally so the set is warm before anyone opts in, and never for TRACE, whose path bytes are per-hop SNR.
srv.Shutdown never cancelled the request context, so a SIGTERM mid-export left the plaintext dump and pg_service.conf in TMPDIR. The handler now derives its context from the app context and the router removes staging dirs older than 20 minutes at startup.
Duplicate observations are dropped before the hub today, so a live map only sees each packet's first hearing per observer. Clients that configure includeRepeats also get later hearings whose path is new; nothing is stored, repeats skip the count query, and they are dropped first when the broadcast channel is half full.
Storing a full candidate snapshot on every observation was the wrong call. The same node lists got copied thousands of times, about 22 GB on prod. Resolve endpoints live like path hops and drop the column.
MrAlders0n