Plume

Plume executes data analytics queries on Dandelion leveraging DuckDB. It parses SQL queries using the standard DuckDB query parser and optimizer, then converts the resulting DuckDB LogicalPlan to a Plume PhysicalPlan (Dandelion composition DAG). The plan is executed using a set of operators built on top of DuckDB's internal compute kernels without using a full-fledged DuckDB instance to keep the functions lightweight.

Layout

The repository is structured as follows:

• core/ — the query engine + SQL compiler, built as the plume library shared by functions/ and client/.
• functions/ — the Dandelion-invoked function executables (plume_stage, plume_{csv,pq}_{prepare,stage}) and their host ABI.
• client/ — the SQL -> Dandelion-composition compiler plus the plume_export/ plume binaries
• benchmarks/ — the TPC-H harness + a simple micro-benchmark tool, built against client.
• cmake/ — shared CMake modules (DuckDB, cpr, nlohmann_json) and the tracked DuckDB extraction patches.
• dandelion/ — the Dandelion preload config listing the function binaries to deploy.
• scripts/ — setup and build scripts, including the two below.

Building the functions

scripts/build_dandelion.sh cross-compiles functions/ with the Dandelion SDK compiler and builds DuckDB for the target platform. Output binaries land in <build_dir>/functions/.

# (optionally) install LLVM, Cmake, libssl-dev, ninja
./scripts/setup_cloudlab.sh

# fetches the dsdk compiler if not provided (all flags have defaults)
./scripts/build_dandelion.sh -p <platform> -b <build_dir> -c <dsdk_compiler_path> -t <cpu_arch>

Building the client / benchmarks

scripts/build_client.sh builds client + benchmarks with the host compiler. Output binaries land in <build_dir>/plume/client/ and <build_dir>/plume/benchmarks/.

# (optionally) install LLVM, Cmake, libssl-dev, ninja
./scripts/setup_cloudlab.sh

# builds
./scripts/build_client.sh -b <build_dir> -d <duckdb_dir>