Stage 2 test report

Recorded 2026-11-04 after the commands below. GitHub Actions has not been run from this machine. Linux x86_64 is configured as a required decode job and is untested here.

Environment

The eight .f3 fixtures were not regenerated. Their SHA-256 values are still the ones in test/fixtures/manifest.json. small.f3 is 4168 bytes, SHA-256 6f1ff9d7202c3f0d8f910a4ef154ba195cd225804d261ca752339cc8fc123a13.

Toolchain spike

Stable rustc 1.97.1 cargo check of native/f3_native stopped in fastlanes on #![feature(generic_const_exprs)].

nightly-2025-02-24 (rustc 1.87.0-nightly) stopped before compiling Vortex. Rustler 0.38.0 declares rust-version = "1.91". A fresh resolve also selected croaring 2.9.0, which wants rustc 1.95.

rustc 1.101.0-nightly (2026-10-15) compiled Vortex's feature gates, then failed in vortex-fastlanes because ArrayChunks::into_remainder no longer returns Option. That signature changed in rust-lang/rust#149127, merged 2025-11-21. The same nightly also needed chrono = "=0.4.39" so arrow-arith 53.3.0 could compile. chrono 0.4.45 adds Datelike::quarter, which collides with Arrow's own quarter.

nightly-2025-11-15 is after Rust 1.91 and before that standard-library change. With croaring = "=2.1.0" (the upstream lock) and chrono = "=0.4.39", cargo check finished. F3_BUILD_NATIVE=1 mix compile --force produced an arm64 priv/native/f3_native.so. nm -gU on that file showed no fff-poc and no mimalloc symbols.

The first uncompressed int64 EncUnit of small.f3 (offset 0, size 1664) decoded to the values 1, 2, and 3.

Commands

mix format
mix credo --strict
mix test
66 passed (3 doctests, 63 tests), 4 excluded
mix test --cover
66 passed, coverage 92.05%
F3_BUILD_NATIVE=1 mix test --only native
4 passed, 65 excluded

mix test and mix test --cover compile without Rust. F3_BUILD_NATIVE=1 mix compile --force in the test environment is what loads the NIF for --only native. The default beam used by mix test does not contain it.

One local run of the pinned FileReaderV2 (fff-poc example dump_rows, built with nightly-2025-02-24) wrote test/fixtures/golden/*_rows.json for seven fixtures. dictionary.f3 panics in that reader with Incorrect type of indices. The native test compares those seven files to the golden JSON and all eight files to the manifest rows. CI does not regenerate the .f3 files.

Not measured

No throughput, latency, or RSS. The eight fixtures are smaller than one encoding unit, so they cannot show dirty-scheduler delay. The NIF is still scheduled DirtyCpu. Linux x86_64 and Linux arm64 were not executed. The required decode job in .github/workflows/ci.yml installs nightly-2025-11-15 on macos-14 and ubuntu-latest and runs mix test --only native. The informational native job stays on stable Rust and is expected to fail.

Implemented

F3.read/2 returns {:ok, [F3.Batch.t()]}. F3.stream/2 returns a stream and does not pread until the first pull. Enum.take/2 stops further reads. The caller owns close/1. Options are columns:, row_groups:, and batch_size: (default 8192). One NIF call decodes a whole EncUnit. Elixir slices it.

The eight fixtures round-trip, including nullable values, a nullable struct, a list of int32 with an empty list, a local dictionary, zstd, and lz4. CUSTOM_WASM and a shared dictionary are :unsupported_encoding and are not executed. Bool and plain binary are not in the fixtures.

A wiped EncUnit returns {:error, %F3.Error{category: :corrupt}}. The caller stays alive. Chunk checksums are checked only when the footer field is present. These fixtures store none. Postscript checksums stay reported and unverified.

Next

Stage 3 is Wasm. Execution stays off.