Where DFO Online can use fewer resources.

A capsule probe is approximated with 27 rays: three points along the capsule, with nine rays at each point. An obstacle check can cast twice, and a fully blocked detour can try ten directions. Using the current enemy AI and collider against a small room with four blocking walls reproduced 540 ray queries in one detour call.

This establishes a possible blocked case, not the cost of every corner encounter. It does not establish how many milliseconds those queries take in the live game.

Worth testing: reuse obstacle results within a step, budget detour searches across updates, and compare the ray bundle with a suitable swept-volume query. Thin walls, slopes, ledges, and corners need regression checks.

Removing a door’s collision bucket invalidates the collider-wide broad-phase spatial index. The next relevant query walks the remaining buckets to rebuild it, although their geometry did not change. A direct check confirmed that the dungeon’s triangle data and other buckets remain intact.

Worth testing: incrementally update the affected bucket’s index entries, or separate static geometry from frequently changing objects. “Rebuilding the entire dungeon’s collisions” overstates what currently happens.

The ordinary dungeon path submits its entire static batch and unbatched draw list without camera visibility culling. The building path has the same general shape. GPU clipping does not eliminate CPU submission and vertex-processing work.

Worth testing: partition large levels into spatial chunks or rooms and cull the camera’s draw set. Shadow visibility needs its own set; something outside the view can still cast a visible shadow.

Shadow images are cached, but _staticSignatures still scans recorded geometry against selected lights every frame to decide whether those images changed. Candidate lists are rebuilt when needed in that frame. Additional indoor lights also perform signature scans.

Worth testing: retain spatial candidate sets between frames and invalidate them with geometry and light change generations. Preserve the existing fixes for shadow flicker, moved lights, and newly visible casters.

Visible ordinary billboard batches are sorted by texture key, then each batch receives a draw call. Shared buffers and sorting reduce state changes, but do not combine those submissions. The source has a separate instanced low-poly-tree path; this observation concerns ordinary billboards.

Worth testing: merge compatible batches across pixels, or use texture arrays and instancing where appropriate. Transparency order, animation, and shadow behavior constrain which batches can share a draw.

Clients retrieve shared town state on their own schedules. Homes have a 60-second client cache, while homesInTown reconstructs town data alongside tenancy, guild, vote, and offer information. Different players pay much of the same query work to see the same town.

Worth testing: separate versioned shared snapshots from personal access information, refreshing the shared portion on changes. Personal responses cannot simply be cached for everybody, and authoritative checks must remain on actions.

regionGroundRead retrieves witness records, groups them, and interprets them in JavaScript. Today’s caching fix reduces repetition, but the cache is per isolate, expires after ten minutes, and does not retain empty results. Cold isolates can repeat the aggregation.

Worth testing: maintain a durable, incrementally updated summary of established regional facts. Ordinary reads can use that summary instead of reconstructing it from reports.

Each incoming pose scans the room’s peers before selecting recipients. Enemy-state delivery also scans the room. Above 32 listeners, nearest-player selection adds another linear pass and temporary arrays. Across moving players, this is approximately quadratic work within a room.

These are counts from the current fan rule and an all-in-range crowd model. They are not CPU timings or a production load test. Batching reduces WebSocket messages; it does not remove the earlier recipient-selection work.

Worth testing: maintain spatial interest groups and reusable recipient sets. Update membership when location meaningfully changes. Dense events also need explicit update priorities; a spatial index alone cannot thin a crowd where everyone is nearby.

Shards keep outgoing polls open while holding sockets; the receiving bus holds a request and timer. Where that topology is used, both objects can remain non-hibernating even when doing little CPU work. Existing sleep changes help, so current savings depend on the deployed topology and traffic.

Worth testing: measure active duration by room type, consolidate quiet rooms, and investigate hibernation-compatible delivery. Removing polling without replacing its delivery guarantees would break behavior.