Skip to content

Security: vyncint/oxidelake

SECURITY.md

Security policy

Reporting a vulnerability

Please report suspected vulnerabilities privately through GitHub's Security → Report a vulnerability form on this repository rather than in a public issue. Include the commit or version, a reproduction, and the impact you believe it has. You will get an acknowledgement, and a fix or a written assessment, as fast as a small project can manage.

Posture

What the project does continuously, enforced by required CI on every change:

  • unsafe is confined and justified. Allocator/FFI internals only (oxidelake-memory, oxidelake-device, the io_uring worker in oxidelake-storage); every block carries a // SAFETY: comment stating the invariant it relies on (clippy.toml requires it), every raw resource is owned by an RAII type, and unsafe_op_in_unsafe_fn is denied. oxidelake-core is #![forbid(unsafe_code)].
  • No panics in library code. unwrap, expect, todo! and unimplemented! are denied by the workspace lints; unimplemented paths return EngineError::Unsupported. A plan payload from the network cannot drive the predicate evaluators arbitrarily deep (MAX_PREDICATE_LEAVES).
  • Untrusted input is validated before it becomes a node. The plan codec checks magic, version, column indices, types and dimensions; generated SQL quotes every identifier; a model path given to predict must be a constant, and the layer shapes in the file it names must fit the tensor data the file holds before anything is allocated (#96). A Parquet file's footer is checked against the file before any read is sized by it, so a corrupt or hostile file is an EngineError::Format rather than a panic or an allocation of whatever the footer declares (ParquetScanGuard; what it does not cover is in STATUS.md). An Arrow IPC spill or cache file is checked against its own length before arrow-ipc decodes it, so a corrupt one is an error, not a panic or an allocation the file cannot back (#110).
  • Dependency policy (deny job): RUSTSEC advisories, a license allow-list, banned crates and crates.io-only sources on every PR; weekly grouped Dependabot updates that leave the DataFusion/Ballista chain's majors alone. Every accepted advisory in deny.toml carries a written reason.
  • Workflow security (zizmor job at --persona=pedantic): every GitHub Action pinned to a full commit SHA, checkouts that do not persist credentials, a read-only workflow token by default. Accepted findings are in .github/zizmor.yml.
  • Build caches are dependency-only and never restored for a release. Three jobs use Swatinem/rust-cache with cache-bin: false and cache-workspace-crates: false, so no workspace crate and no tool binary is ever served from a cache; what returns is third-party dependency builds keyed by the lockfile. Every such step is guarded by github.workflow == 'CI' && !inputs.clean, so a workflow_call from release.yml restores nothing and builds from scratch — a published artifact never comes from a cache (ADR-0017, measured in docs/ci-cache-measurement.md).
  • Protected history. main takes only squash-merged pull requests with required-green and commit-policy green — direct pushes are rejected, the maintainer's included. v* tags can only be created by an admin, and releases publish through crates.io Trusted Publishing from a release environment that deploys from v* tags alone.
  • GPU correctness is asserted, not assumed. Every device operator is conformance-tested against stock DataFusion, and the Metal suite asserts device execution through transfer counters so a silent CPU fallback cannot pass as a GPU result.

Known, accepted advisories

Each is a dependency advisory that cannot be fixed inside the pinned DataFusion/Ballista chain, accepted with a written reason in deny.toml and re-examined at every chain bump. None is silent: they are listed here, in deny.toml, and in STATUS.md.

  • quick-xml (RUSTSEC-2026-0194/0195) — reached only by object_store's cloud backends, which OxideLake never registers.
  • paste (RUSTSEC-2024-0436) — a compile-time proc-macro dependency of the metal crate, which oxmera pulls on macOS under the predict feature (arrow-flight moved to pastey in 59); no runtime code.
  • chrono, chrono-tz and humantime (RUSTSEC-2026-0543, -0555, -0542) — "unmaintained" notices, not vulnerabilities, with no maintained release to move to. chrono and chrono-tz are arrow 59's date/time and timezone support, under every DataFusion and Ballista crate; humantime is object_store 0.13's duration parsing.

Supported versions

Only the main branch (and the most recent tagged release, if any) receives security fixes.

Trust model

OxideLake is an analytical engine intended to run on infrastructure you control. Its security boundary is the machine or private network it runs on, not the query interface.

  • Cluster mode has no authentication or transport encryption yet, so the network is the boundary, and the binaries hold it. oxide-scheduler and oxide-worker speak plain gRPC and Arrow Flight and will execute any plan a connected client submits, including reading any file the worker process can read. Both bind to 127.0.0.1 by default, accept a private address, and refuse a public one, or every address at once, unless --allow-insecure-bind is given (#30).

    Production deployments. The cluster's ports (the scheduler's one port for gRPC, REST and probes, each worker's Flight and gRPC ports, and any --metrics-port) have no authentication and no encryption: whoever can connect can run queries, and a query can read any file a worker can read. Keep them on a network that only the cluster's own hosts and trusted clients can reach, such as a private subnet or a VPN, behind a firewall. oxide-scheduler and oxide-worker refuse to listen on a public address, or on every address (0.0.0.0, ::), unless --allow-insecure-bind says that something else, such as that firewall, already decides who can connect, and they log a warning whenever they listen beyond loopback. TLS and authentication in OxideLake itself are tracked in #30.

  • A lost worker fails its queries rather than risk wrong rows (#148). Ballista 55's recovery from a lost worker can return duplicated rows with exit 0 (fixed upstream by apache/datafusion-ballista#2677, unreleased), so oxide-scheduler fails a query that held output on a worker it loses, and --stage-max-failures defaults to 1. Raising that limit turns Ballista's stage retries, and the bug, back on.

  • Workers trust their scheduler. Physical plans arrive over the network and are decoded by OxidePhysicalCodec. Payloads are validated (magic, version, column indices, types, dimensions, a cap on predicate size) so a malformed payload yields an error rather than a panic, but a malicious scheduler is outside the threat model: it can already run arbitrary SQL.

  • A scheduler's build check is not authentication. oxide-scheduler refuses a worker whose plan fingerprint differs from its own, so a mismatched deployment fails when the worker registers rather than in a query. The worker declares that fingerprint itself (in its executor id), so the check stops accidents, not an attacker who can reach the port.

  • The oxide CLI reads any Parquet path you give it with your privileges; identifiers you pass (--table NAME=PATH) are quoted before they are used in generated SQL.

  • GPU kernels bounds-check every global access; kernel sources are compiled from strings embedded in the binary at build time, never from user input.

  • Unsafe code is confined to allocator and FFI internals and every block documents its invariant (// SAFETY:); oxidelake-core is #![forbid(unsafe_code)].

Dependencies

cargo deny check runs in CI against the RustSec advisory database. Advisories that cannot be fixed by an upgrade are ignored only with a written reason in deny.toml explaining why the affected code path is unreachable in OxideLake; each ignore is revisited whenever the DataFusion/Ballista chain moves.

There aren't any published security advisories