The game logs into HORIZON_HOME and falls back to the player's own game folder when it is unset. Harnesses
that set only HORIZON_NO_AUDIO (and test_native_client, which set nothing) wrote logs, and on a world run
advancements, into ~/Library/Application Support/horizon.
lab.game_env gives every game a fresh HORIZON_HOME, HORIZON_LAUNCHER_HOME, XDG_DATA_HOME and XDG_CONFIG_HOME,
no audio and no launcher self-update; a folder the caller passes is kept. Every launcher of the game uses it,
parity.py gives each harness its own HORIZON_HOME, and tests/test_game_home.py fails for a script in tools/
that starts the game without it.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Preserve Java parity oracles and comparison benchmark instrumentation. Move their remaining dependency locks under tools/benchmark-deps and remove legacy product packaging, fixtures, and CI checks.
Co-Authored-By: OpenAI Codex gpt-6.1-sol <noreply@openai.com>
perf_bench ran every game in a 1280x720 window only. It now runs the
standard set at 1280x720 and 1920x1080 (--sizes picks others), each size
its own reports: <label>-rd<N>.json at 720p as before, <label>-1080p-rd<N>.json
at 1080p.
Horizon takes vanilla's --width and --height for its window's starting size
(1280x720 by default); perf_bench passes them to every game, and to Horizon
only off 720p, so builds before this one still run the 720p set.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
Without min_chunks (the old call) wait_loaded behaves as before, which tests/test_perf_bench.py's
old-builds case relies on.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- DEFAULT_ENDPOINT is https://telemetry.playhorizon.dev. The heartbeat, the launcher's live count and the
outbox upload all go there; HORIZON_TELEMETRY_URL still names another server, and HORIZON_TELEMETRY_URL=off
sends nothing at all. CI workflows and the harnesses that start the game or the launcher set off. Consent
is unchanged: the four switches stay opt-in, only the anonymous ping is always on.
- GET /v1/benchmarks/summary: players' benchmark uploads grouped by GPU, CPU, render distance and game, groups
of 3+ installs only, medians per install then across installs, no install ID, day, OS version, driver,
memory or screen. Cached 10 minutes, CORS for the website. Same rules as website/scripts/benchmarks.mjs,
tested on the same fixtures; a deleted install leaves its groups at once.
- /v1/aggregate no longer lists the newest 50 raw benchmark reports (hardware, driver, day, one install each):
uploads are published only through the summary.
- Behind Railway's proxy the rate limit counts X-Real-IP, else the last X-Forwarded-For entry (the one the
proxy adds; a client could choose the first). Still in memory only, never stored or logged.
- crates/horizon-telemetry-server/railway.json: Dockerfile build from the repository root, health check
/healthz, one replica in europe-west4 (one SQLite file on the /data volume).
- Docs: docs/telemetry.md, the server README (the summary's shape, Railway setup), the privacy page (the ping
is on, where the data lives, how benchmark uploads are shown).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Right after a teleport, and at spawn, Horizon v2026.10.22.04 reports nothing queued and steady chunk counts
for over a second before it starts loading. wait_loaded() took that for "loaded" and stopped the clock with
0 to 3 sections meshed (load 0.1 s), so the flights ran over empty terrain and the frame rates meant nothing.
Now the counts must change once since the wait began (or stay at the full view for 5 s), and the resident
chunks must reach 80% of the view circle (pi x RD^2), before a steady second counts as loaded.
Measured with it on the Mac mini and the ThinkPad for the website's Benchmarks page (#1274). At RD 32 a 4-thread
laptop needs --timeout above the default 300 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
lab.VERSION, lab.GAME (.local/minecraft/<version>) and lab.RESOURCES
(.local/horizon-client/<version>) come from the lock, and lab.version_info()
reads the pinned client jar's version.json. Every harness, generator and
tool that named 26.3 (resource folders, jar and source paths, result
records, the vanilla client's --version and asset index, the bare login
client's protocol, the benchmark's options.txt data version, the packet id
table's header) uses them, so a candidate worktree pinned to another version
runs them unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Horizon game with content mods builds its content root and block behaviour
snapshot on its first start, in its game folder, and reuses them afterwards.
perf_bench starts every game from an empty folder, so it only measured first
starts. --warm starts the game once to the title screen first, unmeasured.
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>
Loading the eight carried-over mods in the game (not only the probe) showed
two limits real packs reach. Block textures past 2048 layers are another PR;
this one handles block states and measures a pack.
- Chunk sections store a block state in 16 bits. Vanilla 26.3 has 35,723
states; Macaw's four mods add 36,154, the eight 41,389. Past 65,536 ids
wrapped onto other blocks (a trapdoor became a hanging sign).
mod_content::build now refuses a set of mods past MAX_STATES and names the
states per mod; the game then starts without content mods, as for any
refused content.
- tools/mod_port_harness.py probes each mod on its own against vanilla, as a
player adding it would load it.
- tools/perf_bench.py --game neoforge: Minecraft 1.21.1 on NeoForge 21.1.259
with --mod-jars and --deps, driven by a dev-only NeoForge port of the bench
mod (mods/horizon/src/neoforge-bench); --port-work loads a work folder's
carried-over mods into Horizon. tools/neoforge_oracle.py installs the
client once under .local/neoforge-oracle (the installer, then the vanilla
libraries it inherits, sha1-checked) and builds the driver. A NeoForge
start that crashes before the driver loads (FML reading fml.toml while
creating it) is retried once.
Measured (docs/modpacks.md): six ported mods in Horizon against the same
mods on NeoForge, Mac mini M4 Pro, 2 games per row on a loaded machine. The
world is ready 3.1x sooner at 8 and 16 chunks, the flights run 1.7x and
3.6x the frames, and Horizon's RSS is not lower (+1.1 GB with the mods).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Builds that report status.streaming are loaded once chunk counts are
stable and no stage (generation, decoration, light, meshing) has held work
for half a second. Meshing going quiet also waited through the remeshing
that follows a load (fluids settling, random ticks: 1500 falling to 60
meshes/s for ~150 s at RD16), which is simulation, not loading. Older
builds keep the meshing-rate rule.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sections wait for their neighbours' light before they are meshed, so the
chunk counts settled while edge sections were still undrawn.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each game runs --repeats times (3) per render distance and writes one v1
report per distance (benchmarks/perf/<label>-rd<N>.json): figures as means
with standard deviation and 95% CI, per-flight frame entries with the fixed
histogram, the hardware without the host name. frames() and summary() match
crates/horizon-bench value for value (tested on the same inputs).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tools/perf_bench.py runs Horizon, the vanilla 26.3 client and vanilla with
Sodium, Lithium, FerriteCore and ImmediatelyFast through the same steps:
fresh world (seed 1234, creative) per render distance, start to title, start
to in-world, then per scene teleport, chunk load, settle and a fixed 192-block
flight. Every frame time is kept and one function computes average FPS, 1% and
0.1% lows, p50/p95/p99/p99.9 and a frame-time histogram for every game; RSS is
sampled from outside the process for every game. One options.txt (vanilla's
keys, Fancy preset values with a Custom preset so the render distance holds)
is written for all of them.
Vanilla is driven by a dev-only Fabric mod (mods/horizon/src/bench) compiled
against the user's own client jar: it creates the world with
WorldOpenFlows.createFreshLevel, teleports with /tp and places the player on
the path every frame as horizon-ctl fly does. Nothing of the game is
redistributed and the performance mods are fetched from Modrinth with the
hashes pinned in packs/mods.lock.json.
Horizon side: `fly ... PITCH raw` also replies with every frame time, and
--offscreen drives frames from the event loop, so a window nobody sees (a
locked macOS session) still renders.
docs/benchmark.md describes what is measured, how fairness is kept, the
caveats and how to reproduce it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ludovic Marie
rdbtCVS