Severity: LOW
Component: crates/cgka-engine
Summary
Engine::transport_group_id_index (HashMap<Vec<u8>, GroupId>) is insert-only. There is no .remove(...)/.retain(...)/.clear(...) anywhere in the crate, so it accumulates entries with no ceiling on a long-running device. It sits on the per-inbound-event routing hot path.
Location
crates/cgka-engine/src/engine.rs:206 — field definition.- Inserts:
engine.rs:718, engine.rs:756, group_lifecycle.rs:302, group_lifecycle.rs:632.
- Read on the inbound routing path:
message_processor/ingest.rs:1746 (group_id_for_transport_group_id).
engine.rs:686 — reindex_transport_group_id is documented as additive on routing rotation: prior nostr_group_id → group_id entries are retained for the overlap window and never expire.
grep -rn "transport_group_id_index" crates/cgka-engine/src | grep -E '\.remove|\.retain|\.clear' returns nothing.
Two independent growth vectors
- Routing rotation: each
reindex_transport_group_id adds a new transport_group_id → group_id mapping and keeps the prior one (overlap window), which never expires. A single long-lived group that rotates its Nostr routing component repeatedly accretes stale entries forever.
- Group churn: no engine path removes a group's entries on leave / self-eviction / removal / wipe.
realize_self_eviction and the removal seams set removed = true but keep the index entries; do_join_welcome only re-inserts.
Impact
Steady, unbounded memory growth on always-on clients that churn groups or rotate routing. Stale mappings also outlive a group whose local record was later deleted at the storage layer (delete_local_group_data, #756); an inbound event on the stale routing id then resolves to a now-deleted group id and does an extra get_group miss per event on the hot inbound-routing path (wasted work, no crash).
Suggested fix
Prune the account's transport_group_id_index entries on group leave/removal/wipe (mirror the group-lifecycle removal seams), and drop superseded routing mappings once the rotation overlap window closes.
Dedup
Same fix shape as #868 (EngineMetrics::last_applied grows one entry per group, never pruned), but a distinct site and trigger: #868 is diagnostic telemetry keyed by group_id on the metrics path; this is a routing-critical index keyed by transport_group_id on the inbound hot path, with a second growth vector (routing rotation) that has no analogue in #868. No open/closed issue covers transport_group_id_index.
Filed from a meticulous automated code-review pass; verified against source and deduplicated against the existing open/closed issue set.
Severity: LOW
Component:
crates/cgka-engineSummary
Engine::transport_group_id_index(HashMap<Vec<u8>, GroupId>) is insert-only. There is no.remove(...)/.retain(...)/.clear(...)anywhere in the crate, so it accumulates entries with no ceiling on a long-running device. It sits on the per-inbound-event routing hot path.Location
crates/cgka-engine/src/engine.rs:206— field definition.engine.rs:718,engine.rs:756,group_lifecycle.rs:302,group_lifecycle.rs:632.message_processor/ingest.rs:1746(group_id_for_transport_group_id).engine.rs:686—reindex_transport_group_idis documented as additive on routing rotation: priornostr_group_id → group_identries are retained for the overlap window and never expire.grep -rn "transport_group_id_index" crates/cgka-engine/src | grep -E '\.remove|\.retain|\.clear'returns nothing.Two independent growth vectors
reindex_transport_group_idadds a newtransport_group_id → group_idmapping and keeps the prior one (overlap window), which never expires. A single long-lived group that rotates its Nostr routing component repeatedly accretes stale entries forever.realize_self_evictionand the removal seams setremoved = truebut keep the index entries;do_join_welcomeonly re-inserts.Impact
Steady, unbounded memory growth on always-on clients that churn groups or rotate routing. Stale mappings also outlive a group whose local record was later deleted at the storage layer (
delete_local_group_data, #756); an inbound event on the stale routing id then resolves to a now-deleted group id and does an extraget_groupmiss per event on the hot inbound-routing path (wasted work, no crash).Suggested fix
Prune the account's
transport_group_id_indexentries on group leave/removal/wipe (mirror the group-lifecycle removal seams), and drop superseded routing mappings once the rotation overlap window closes.Dedup
Same fix shape as #868 (
EngineMetrics::last_appliedgrows one entry per group, never pruned), but a distinct site and trigger: #868 is diagnostic telemetry keyed bygroup_idon the metrics path; this is a routing-critical index keyed bytransport_group_idon the inbound hot path, with a second growth vector (routing rotation) that has no analogue in #868. No open/closed issue coverstransport_group_id_index.Filed from a meticulous automated code-review pass; verified against source and deduplicated against the existing open/closed issue set.