Skip to content

README.md

Latest commit

 

History

History
89 lines (69 loc) · 4.36 KB

README.md

File metadata and controls

89 lines (69 loc) · 4.36 KB
Raw

Fuzzing

Coverage-guided fuzzing of the untrusted dataset parsers in the vector_db_benchmark library crate, using cargo-fuzz / libFuzzer.

These parsers ingest attacker- or corruption-controlled bytes and allocate / index based on header values, so they are the highest-risk surface. The goal is that no malformed input ever panics, integer-overflows, or OOMs — malformed input must return Err.

Targets

Target Parser File
sparse_reader read_sparse_matrix — CSR binary matrix src/readers/sparse_reader.rs
npy_reader read_npy_vectors — NumPy .npy header + f32 data src/readers/npy_reader.rs
jsonl_reader read_jsonl_vectors — JSONL arrays of floats src/readers/jsonl_reader.rs
compound_payloads read_payloads_jsonl — payloads.jsonl → metadata src/readers/compound_reader.rs
metadata parse_metadata_from_json — JSON → MetadataItem src/readers/metadata.rs
compound_queries read_compound_queries — tests.jsonl (query + conditions + ground-truth ids) src/readers/compound_reader.rs
compound_data read_compound_data — vectors.npy + payloads.jsonl directory src/readers/compound_reader.rs
parse_info parse_info — raw server INFO text → nested JSON src/parsers.rs
datetime datetime_to_epoch_secs — ISO-8601 filter value → epoch src/parsers.rs
ft_search parse_ft_search_response — FT.SEARCH RESP reply → hits src/parsers.rs
sparse_roundtrip read_sparse_matrix(write_sparse_matrix(x)) == x src/readers/sparse_reader.rs
npy_roundtrip read_npy_vectors(write_npy_vectors(x)) == x src/readers/npy_reader.rs

The parse_info / datetime / ft_search parsers ingest untrusted server responses (any Redis-wire server: Redis / Valkey / Dragonfly / ElastiCache / MemoryStore). They were extracted from the engine binary into the library crate (src/parsers.rs) so coverage-guided fuzzing can reach them; the binaries call the exact same implementations. The ft_search target frames arbitrary bytes into a redis::Value with the redis crate's own wire parser, then runs our extractor — fuzzing both RESP framing and our (id, score) extraction.

The two *_roundtrip targets are differential: they assert_eq! that a value survives write → read, catching silent corruption / offset bugs that crash-only fuzzing cannot. NaN is canonicalized (NaN != NaN would spuriously fail) and sizes are bounded via arbitrary so the writer never over-allocates.

Dictionaries

fuzz/dictionaries/*.dict seed libFuzzer with format tokens (NPY magic, JSON / RESP / INFO / CSR / datetime tokens). Each matrix entry in fuzz.yml maps a target to a dictionary, passed as -dict=.... Control bytes in a dict must use \xHH escapes (libFuzzer only understands \\, \", and \xHH).

Requirements

  • Nightly toolchain (libFuzzer needs it): rustup toolchain install nightly
  • cargo install cargo-fuzz
  • System deps for the library crate: libhdf5-dev pkg-config

Run locally

# from the repo root
cargo +nightly fuzz build                      # build all targets
cargo +nightly fuzz run sparse_reader -- -max_total_time=60 -rss_limit_mb=2048

Each target writes crash reproducers to fuzz/artifacts/<target>/. Reproduce a crash with:

cargo +nightly fuzz run sparse_reader fuzz/artifacts/sparse_reader/crash-XXXX

Corpus

A few valid, minimal inputs are committed under fuzz/corpus/<target>/valid* so fuzzing starts from valid structure. Everything else the fuzzer generates locally is git-ignored; CI persists the accumulated corpus via actions/cache.

Adding a target

  1. Add fuzz/fuzz_targets/<name>.rs (#![no_main] + fuzz_target!), writing the bytes to a tempfile::NamedTempFile and calling the reader (or feeding bytes straight to a JSON parser).
  2. Register a [[bin]] entry in fuzz/Cargo.toml.
  3. Add the target name to the matrix in .github/workflows/fuzz.yml.

When the fuzzer finds a crash

Harden the reader so the input returns Err (checked arithmetic for size math, validate offsets before slicing, cap allocations by file size), then pin the exact crashing bytes as a unit test in the reader module (assert!(... .is_err())) so the fix stays fixed in the normal cargo test suite.