rbt — Rust lake build tool

Medallion SQL DAGs for filesystem / object-storage lakes: bronze files → silver → gold, with dbt-shaped models, frontmatter contracts, and in-process DataFusion execution.

Status: 0.3.8. One package: library + CLI binary rbt (rbt-datalake on crates.io). Spine works (compile / run / test / --select). Default materialization is streaming full-refresh Parquet (no full-result RAM retention); default ref() is lake Parquet re-read. Multi-root paths, path_glob, opaque protobuf bronze. Filesystem Iceberg-style tables via --format iceberg (data + metadata layout; not multi-catalog OCC).

Why rbt


Niche	Bronze → silver → gold on a lake, not warehouse-dbt
Stack	Rust + Arrow + DataFusion + optional Iceberg-style FS tables + jshift
UX	Models, `ref` / `source`, frontmatter tests, CLI select
Claim	Replace ad-hoc scripts / Spark for team-scale medallion jobs

Install

cargo install rbt-datalake
rbt --help

Library (crate name differs from import path):

[dependencies]
rbt-datalake = "0.3.8"

use rbt::RbtProjectConfig; // lib name is still `rbt`

Git / from source:

cargo install --git https://github.com/shan-alexander/rbt --locked

From this repo:

git clone https://github.com/shan-alexander/rbt && cd rbt
cargo build -p rbt-datalake --release
# binary: ./target/release/rbt

Requires a matching rustc/cargo pair (see rust-toolchain.toml).

crates.io notes

Package name: rbt-datalake (the short name rbt is an unrelated “Rust bot toolkit”). Binary and lib import path remain rbt.

Early 0.0.1 internals (rbt-core, rbt-engine, …) are orphaned. Each has a 0.0.4 deprecation stub README pointing here. Do not depend on them.

Quick start (smoke fixture)

Tiny JSONL bronze → staging → dim:

cargo build -p rbt-datalake --release
./target/release/rbt compile -p examples/smoke_fixture --bronze-check fail
./target/release/rbt run -p examples/smoke_fixture --select dim_ticker --format parquet
./target/release/rbt test -p examples/smoke_fixture --select dim_ticker

Full market example (large Arrow IPC bronze):

./target/release/rbt compile -p examples/full_e2e_rbt_example --bronze-check fail
./target/release/rbt run -p examples/full_e2e_rbt_example --format parquet

See examples/smoke_fixture/README.md and examples/full_e2e_rbt_example/README.md.

CLI

Command	Purpose
`rbt compile -p <proj> [--select …]`	DAG + bronze path checks
`rbt run -p <proj> [--select …] [--format parquet\|iceberg\|…]`	Execute subgraph (ancestors always included)
`rbt test -p <proj> [--select …]`	Run subgraph + frontmatter tests
`rbt bench`	In-memory throughput microbench

`--select` (dbt-like)

Spec	Meaning
`dim_ticker`	That model + all ancestors on `run`/`test`
`stg_trades+`	Model + descendants (+ ancestors on execute)
`+fact_x`	Model + ancestors
`a,b`	Union of selectors

`--format`

Value	Output
`parquet`	Single `.parquet` file per model (default)
`iceberg`	Table dir: `data/part-*.parquet` + `metadata/v1.metadata.json`
`parquet-and-iceberg`	Flat parquet and sibling `.iceberg/` dir
`jsonl` / `csv`	Line/CSV files

Iceberg here is a filesystem table layout written by rbt (full refresh), not Glue/REST multi-writer production yet.

Project shape

my_project/
  rbt_project.yml
  models/staging|transforms|marts/*.sql   # YAML frontmatter + SQL
  lake/bronze/…                           # external landing
  lake/silver/…                           # rbt outputs
  lake/gold/…

Staging frontmatter: scan_path, source_format, path_glob, partition_by, grain, tests, columns.*.description / context.
Multi-root lakes, absolute targets, glob semantics, protobuf bronze, and when path_glob disables DataFusion listing pushdown: docs/MULTI_ROOT_AND_PATH_GLOB.md.

Materialize + `ref()` (optional)

By default rbt streams model results to disk (execute_stream → atomic Parquet publish) and re-reads the lake file for downstream {{ ref() }} (lake-as-truth; lower peak RSS).

materialize:
  mode: stream                 # default; use collect only for emergency/debug
  # max_row_group_rows: 1000000
  # max_row_group_bytes: 134217728   # 128 MiB flush hint
  ref_strategy: parquet        # default lake re-read for ref()
  # ref_strategy: memtable
  # memtable_max_rows: 50000

Env: RBT_MATERIALIZE_MODE=collect forces legacy collect path.
Tradeoffs: docs/REF_STRATEGY.md, docs/STREAMING_MATERIALIZE_PLAN.md.

Package layout

Single workspace member: crates/rbt (crates.io: rbt-datalake) — lib modules (core, engine, scan, json, materializer, testing) + bin/lib name rbt.

Docs

CONTRIBUTING.md — priorities and positioning
thesis.md — product thesis
docs/README.md — ADR index, archive, publishing
docs/adr/ADR_003_UDF_RSMODELS.md — polyglot roadmap

Benchmarks

cargo bench -p rbt-datalake --bench pipeline
# see crates/rbt/benches/README.md — uses smoke + full_e2e bronze when present

CI

bash scripts/smoke.sh

GitHub Actions: .github/workflows/ci.yml.

Version / publish

See CHANGELOG.md and docs/PUBLISHING.md.

License

Apache-2.0 — see LICENSE.

rbt-datalake 0.3.8