Cargo.toml

[workspace]
resolver = "2"
members = ["crates/*"]

[workspace.package]
version = "0.2.0"
edition = "2024"
# The lockfile's floor (Ballista's bundled AWS SDK crates need 1.94.1), not
# the 1.88 of the direct dependencies; verified by `make msrv` / the CI job.
rust-version = "1.94.1"
license = "Apache-2.0"
authors = ["Vyncint Ng <vyncint@icloud.com>"]
repository = "https://github.com/vyncint/oxidelake"
homepage = "https://github.com/vyncint/oxidelake"
documentation = "https://docs.rs/oxidelake-runtime"
readme = "README.md"
# `gpu` is how someone searching for this searches; the rest say what it is.
keywords = ["gpu", "arrow", "datafusion", "lakehouse", "cuda"]
categories = ["database-implementations", "science"]
# Nine crates publish together (docs/RELEASING.md). The kernels live INSIDE
# oxidelake-device so the published crate is self-contained — they used to sit
# at the repository root, which `cargo publish` cannot see.
exclude = [".github/*", ".claude/*", "docs/*", "target/*"]

[workspace.dependencies]
# internal crates
oxidelake-core = { path = "crates/oxidelake-core", version = "0.2.0" }
oxidelake-memory = { path = "crates/oxidelake-memory", version = "0.2.0" }
oxidelake-device = { path = "crates/oxidelake-device", version = "0.2.0" }
oxidelake-compute = { path = "crates/oxidelake-compute", version = "0.2.0" }
oxidelake-storage = { path = "crates/oxidelake-storage", version = "0.2.0" }
oxidelake-planner = { path = "crates/oxidelake-planner", version = "0.2.0" }
oxidelake-runtime = { path = "crates/oxidelake-runtime", version = "0.2.0" }
oxidelake-tui = { path = "crates/oxidelake-tui", version = "0.2.0" }
oxidelake-api = { path = "crates/oxidelake-api", version = "0.2.0" }

# data plane — Ballista 54 is the root of this chain
ballista = "54"                      # client: SessionContext::remote / standalone
ballista-core = "54"
ballista-scheduler = "54"
ballista-executor = "54"
datafusion = "54"
datafusion-proto = "54"              # PhysicalExtensionCodec for Gpu*Exec; always identical to datafusion
arrow = { version = "58", features = ["prettyprint"] }
parquet = "58"                       # lake format: statistics, page index, Bloom filters
object_store = "0.13"                # ObjectStore trait; io_uring local store implements it
# async
tokio = { version = "1.52", features = ["full"] }
futures = "0.3"
bytes = "1"
async-trait = "0.1"                  # only to implement object_store's ObjectStore (foreign trait)
# parallelism
rayon = "1"
# syscalls the standard library does not expose: installing a SIGUSR1 handler
# in the io_uring signal test (#27). Already in the graph transitively; this
# only names it so a dev-dependency can use it directly.
libc = "0.2"
# errors + observability
thiserror = "2"
anyhow = "1"
tracing = "0.1"
tracing-subscriber = { version = "0.3", features = ["env-filter"] }
# serialization (codec payloads, config)
serde = { version = "1", features = ["derive"] }
postcard = { version = "1", features = ["use-std"] }
# storage fast path
io-uring = "0.7"                     # Linux target-gated, behind the io-uring feature
# GPU (all optional)
cudarc = { version = "0.19", default-features = false, features = ["std", "driver", "nvrtc", "dynamic-loading", "cuda-12080"] }
objc2 = "0.6"                        # macOS target-gated, behind the metal feature
objc2-metal = "0.3"
objc2-foundation = "0.3"
# UI + CLI
ratatui = { version = "0.30", features = ["crossterm_0_29"] }
crossterm = "0.29"
clap = { version = "4", features = ["derive"] }
# in-database inference (optional, behind oxidelake-compute/predict)
#
# default-features = false is load-bearing, not tidiness. oxmera registers its
# CUDA backend from a `#[ctor]` that runs before `main` and calls
# CudaContext::new(0) — so the default build would have every `oxide` process,
# scheduler and worker included, dlopen libcuda and create a CUDA context at
# load time, beside the one OxideLake creates itself, even for a pure-CPU
# query. oxmera 0.4.0 made that a feature; this is the reason to use it.
oxmera = { version = "0.5", default-features = false }
safetensors = "0.8"                  # reading a model's header to recover its shape

# test-only
insta = "1"
# `serde` is on because the PTY suite round-trips a screen through JSON — the
# exact bytes `TERMLENS_ARTIFACT_DIR` writes when a wait fails in CI, so the
# artefact a failure leaves behind is a format this suite has read back. It
# also makes those artefacts `.screen.json` rather than `.screen.txt`.
termlens = { version = "0.11", features = ["serde"] }
assert_cmd = "2"
tempfile = "3"
rand = "0.10"
serde_json = "1"                     # reading a serialized termlens Screen back

[workspace.lints.rust]
unsafe_op_in_unsafe_fn = "deny"
missing_docs = "warn"

[workspace.lints.clippy]
unwrap_used = "deny"
expect_used = "deny"
todo = "deny"
unimplemented = "deny"
dbg_macro = "deny"

[profile.release]
opt-level = 3
# Thin, not fat. With fat LTO each binary's link was a 13.6 GB whole-program
# pass (measured 2026-09-07, Apple Silicon): Cargo links the three binaries
# in parallel, which killed the 16 GB Linux release runners (exit 143 on
# v0.1.0 and v0.1.2) and left the 7 GB macOS runners swapping for four hours
# per target. Thin LTO with one codegen unit peaks at 7.4 GB and links in
# half the time for 9% more binary; a 10M-row aggregate ran in the same
# 40 ms under both. Sixteen codegen units would peak at 2.6 GB but cost 75%
# in size, so the single unit stays.
lto = "thin"
codegen-units = 1
strip = "symbols"
# Deliberately no `panic = "abort"`: unwinding must stay sound across PyO3/FFI
# boundaries and workers must keep backtraces.

[profile.dev.package."*"]
opt-level = 2      # arrow/datafusion are unusably slow at -O0 in tests

# `cargo build --profile profiling`: release code generation with symbols and
# line tables kept, so perf / Instruments can attribute time to functions.
[profile.profiling]
inherits = "release"
debug = 1
strip = false