The hub tests every player's sent-chunk set for every entity and every
column of the view each tick, and each player's tracked-entity map for
every entity: tens of thousands of SipHash lookups a tick with 50
players. Both now use the Fx hash. Their iteration order was random
before and carries no meaning.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every chunk packet scanned all 4,096 blocks of each section holding a
block entity block to list them, though the section's wire bytes were
already cached. The cache now keeps those blocks (index and state, in
scan order) beside the bytes, made in the same pass that encodes them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
encode_in asks whether each of the nine surrounding columns' 24 sections
holds anything, to decide which light layers exist: 216 cache lookups,
each locking the cache and SipHashing a section's address. The column's
sections now share one lock, and the cache uses the Fx hash.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The hub looked up an entity's tracking range (a binary search over type
names) for every player and every entity each tick: about 87,000 string
comparisons a tick with 50 players. The ranges are now worked out once
per entity per tick; the visibility test is the same.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The generation scan walked every area's square of columns, so with 50
players standing near each other it tested about 100,000 columns a tick
(each against every player's range) to find the few hundred around them.
The squares are now merged into one run per row first. The candidates
are the same: they were sorted and deduplicated before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tick_hoppers read the block under every block entity in the loaded world
each tick to find the hoppers. A hopper's block entity is always a plain
container (BlockEntity::Chest), so only those need the lookup.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PlayerAdvancements.flushDirty only works when it is the first packet or
progress changed. Horizon walked the whole advancement tree for every
network player every tick to find nothing new: a fifth of the server
tick with 50 players.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Interpolated animated sprites (`"interpolate": true`) mix two frames into
every mip level with `f32::round`. Baseline x86-64 has no rounding
instruction (it needs SSE4.1), so each byte was a call into libm's roundf:
7.8 to 8.5% of the render thread's samples on a Kaby Lake laptop.
Round with a truncation and an exact fraction test instead. The result
is bit-identical to `f32::round` for every value the blend can produce,
checked over all frame byte pairs at 144 blend times.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PlayerList.tick broadcasts UPDATE_LATENCY once ++sendAllPlayerInfoIn > 600;
the hub sent it every 100 ticks. The extra updates landed in the
dedicated_server harness's respawn window and failed it at random.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Online mode checks accounts since v2026.10.17.07; until #589, players added
before they join get offline UUIDs, with a workaround. Portals follow a
lone player since v2026.10.17.09. Copied game files need v2026.10.17.07 or
newer. Section 10 links the 2026-10-17 benchmark and says plainly that 25
moving players are too many for now. Links to other docs are absolute, so
they work both on GitHub and on the site.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Port CachedUserNameToIdResolver to the dedicated server, in place of the
in-memory name cache: usercache.json in the universe folder is loaded at
startup (bad entries skipped, a bad date warned about), players are added
when they join and kept for a month, and the file keeps the 1000 most
recently used, newest first, as Gson writes it.
The console's op, deop, ban, pardon and whitelist find players by name
through it, in any case. A name it doesn't hold is asked of the profile
service (profiles::find_by_name) on offline servers too, as vanilla does;
one the service doesn't know is the offline player of the lower-case name
on an offline server, and "That player does not exist" on an online one.
tools/test_user_cache_parity.py runs vanilla's server and Horizon's with a
mock profile service and compares the replies, the names looked up and the
files; the dedicated-server harness now compares usercache.json too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With oracle/facts.json, a mod the oracle did not load (a library it
needs is missing) lost every block: none of its ids were in the facts.
It now keeps the guesses from its blockstates, as without an oracle.
ATM10: Regions Unexplored (TerraBlender is CurseForge-only) gets its
805 blocks back.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
horizon-bots and tools/server_bench.py at 1, 10, 25, 50 and 100 bots,
walk and explore, Horizon (main at 123987c4) and vanilla 26.3 taking turns.
The Mac mini was overloaded by other work (load 55-141), so only the 1-
and 10-bot rows are a fair comparison; the README says so and gives the
load of every run. It also records where Horizon's server thread goes at
25 walking bots: 96 % in ChunkEncoder::encode_in.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
At render distance 32 the section arena holds about 1.5 GB of vertices
(2.1 GB of GPU memory in all while flying). Vulkan drivers fail an
allocation once video memory is full instead of paging like OpenGL
drivers, and wgpu's default error handler turns that into a panic: a
4 GB laptop GPU that also drives the desktop crashed a few seconds into
an RD32 world.
Arena pages are now created inside an out-of-memory error scope. A page
that doesn't fit in video memory is created mappable, which the
allocator places in host-visible system memory, so the terrain still
draws (more slowly) as OpenGL drivers do. A page that can't grow keeps
its contents and a smaller new page is tried; with no memory left at
all, the section is skipped rather than the game ending.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Vanilla skips a resource with an invalid path, a data file that does not
parse, one at a time; Horizon's content loader refuses the whole mod
set. carry now keeps them in outside/ with the reason (report
`unreadable`): upper-case paths, JSON that does not parse, shared
tags with "replace": true. A byte-order mark is dropped.
ATM10's 42 carried mods: 19 such files in 7 mods (Mekanism Tools'
ry_UA lang, CTOV's empty tag, Towns and Towers' replacing tags, broken
template pools, recipes and a lang file), and the 42 load as carried.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
lan/chunks.rs: keep the per-section encode cache and the cached
heightmaps, applied to the level's section range from #668. Heightmaps
use client_heightmaps_in with the level's min_y and height, and
set_dimension drops cached heightmaps. The light layers go over the
level's sections and drop sky light where the level has none.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- perf_bench.py samples the physical footprint on macOS as well as
RSS. RSS counts the pages the allocator freed and marked reusable,
so the six mods' "+1.1 GB" is about 0.5 GB of footprint.
- examples/resource_memory.rs: live heap, peak, RSS and footprint after
each resource loader, for a prepared or a content-mod root.
- BlockRegistry indexes ids by a hash of the state, checked against
the states, instead of keeping every state twice: -22 MB live for
vanilla, -124 MB for ATM10's 42 mods (193,762 states).
- docs/modpacks.md: the per-loader breakdown and the next cut (the
per-state tables parsed whole at startup).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Port 26.3's ServerWaypointManager for the players of a served world: each
player who receives waypoints gets ClientboundTrackedWaypointPacket for
every other player who transmits one, as a block (in view), a chunk (one
chunk out) or an azimuth (really far), with the update thresholds,
breaking rules and team colours vanilla uses. Crouching, invisibility and
a carved pumpkin or mob head hide a player; spectators see everyone and
are seen by no one; the locator_bar game rule turns it all off and on.
tools/test_locator_bar_parity.py runs vanilla's server and Horizon's with
two vanilla-codec players and compares every waypoint packet each receives.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
AvatarRenderer.getArmPose gives THROW_TRIDENT to the hand using a
trident: HumanoidModel lifts that arm over the head (xRot * 0.5 - pi)
and leaves the other arm's pose alone (affectsOffhandPose). Horizon
used the plain item pose, so the held trident, drawn with its throwing
model, pointed down through the body. An aggressive drowned raises its
trident the same way (DrownedRenderer), and zombie arms leave that arm
alone (AnimationUtils.animateZombieArms).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
FloatGoal's constructor lets the navigation float, so Mob.jumpInLiquid
adds 0.04F a tick; Horizon set that flag only for a few species, and
every other FloatGoal user (pigs, cows, wolves, cats, horses, witches,
illagers...) got the non-floater 0.3 and bounced out of the water.
Register it from the goal (and from the constructors that set it:
allay, armadillo, camel, goat, parrot, wither), and add 0.04 as the
float vanilla adds.
Mobs that drop back into water now splash (Entity.doWaterSplashEffect
from updateFluidInteraction, skipped on the first tick), with their
swim splash sound. The new mob_ai deep_water scenarios also caught:
horses register addBehaviourGoals last (none for the skeleton horse, no
mount panic for the zombie horse), RandomStandGoal waits the horse
ambient interval (400 for every AbstractHorse), and the cat's sit-on-
block goal draws its delay only for tame cats once the last ran out.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With the wide state ids, ATM10's 42 data-tier mods (193,762 states)
reach the world once these loaders stop refusing them:
- Block sprites: eight texture arrays of 2048 layers instead of four
(9,790 block textures for the 42).
- Per-state tables read as NBT: the table limits cover 2^20 states, and
the collision table in level data uses them.
- Worldgen block tags: 2^20 blocks instead of 8,192 (mineable/pickaxe
holds every stone block of a pack).
- As vanilla's loaders do, a blockstate definition that does not bind,
a missing or broken block model of a mod, and a recipe that does not
parse are logged and skipped instead of dropping every mod (or
stopping the game).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Chunk sections keep block states as u32 and size the global palette
to the registry: 16 bits for vanilla's 35,723 states, wider when
content mods take the registry past 65,536. The content-mod limit
goes from 2^16 to 2^20 states.
The LAN host writes chunks at the registry's width, so with vanilla
(or mods within 16 bits) the wire is unchanged. Remote chunks are read
at the server's width, 16 from a vanilla server, and widened to ours.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>