Skip to content

Minecraft: Java Edition 26.3 rewritten from scratch in Rust, checked against vanilla with 281 differential harnesses. Get it with the Horizon launcher.

2.5 KB
Raw Blame

Shared startup release response

Startup previously fetched /releases?per_page=100 on two worker threads: one for game releases and one for updates. It now decodes that response once and parses game releases once, publishing them before older-launcher fallback lookups. Manual and four-hour update checks still request fresh data. Release page retries remain independent. Existing self-update guards still apply.

Run the committed benchmark against the actual backend implementation:

CARGO_INCREMENTAL=0 CARGO_PROFILE_DEV_DEBUG=0 CARGO_PROFILE_TEST_DEBUG=0 \
  cargo test --manifest-path launcher/Cargo.toml --locked --lib \
  backend::release_loading_tests::benchmark_shared_startup_response \
  -- --ignored --nocapture

Measured on 2026-10-14, Linux x86_64, AMD Ryzen 5 5600X, Rust 1.99.0, unoptimized Cargo test profile with debug information disabled. A shared CARGO_TARGET_DIR was used for local build reuse. Each case warms up 20 times and takes the median of 201 samples; outputs are validated for equivalence and passed to black_box.

Scenario Before median After median Ratio Fetches per startup
Serial parsing/decoding work, zero transport delay 14.701 ms 7.339 ms 2.003x 2 → 1
Concurrent startup workers, 5 ms per mock request 14.778 ms 13.642 ms 1.083x 2 → 1

The measured 201 samples make 402 mock release-list fetches before and 201 after. The concurrent case reduces elapsed time by 7.7%; the two old requests can overlap, so halving requests does not halve startup latency.

The synthetic 453,197-byte JSON contains 99 game releases (including a draft) with roughly 4 KiB of notes each and one launcher release with a platform asset. The mock fetch decodes identical bytes for each request; it does not exercise HTTP, TLS, GitHub rate limits, filesystem installation scans, GUI rendering or self-update downloads. The comparison covers backend orchestration, JSON decoding, sorting, filtering and update selection, not end-to-end launcher startup or release-build timing. Fallback timing is omitted from the benchmark and its ordering/error behavior is tested separately. No persistent cache or extra copy of the game-release vector is introduced.

Validation: 26 launcher library tests passed (benchmark ignored by default); the benchmark passed explicitly; full launcher all-target check and Clippy with -D warnings passed; root and launcher format checks and git diff --check passed. The GUI compiled but was not launched interactively.