Expand description
The in-engine dispatch seam. Every backend (builtin, expression, wasm
container, llm) implements TransformBackend; the standalone
fv-compute library is what the pipeline runner, the derived-property materializer, the
Action runtime, or a REPL embeds to run transform X@v over this Arrow batch.
Scalable-pipeline shape (the whole point): the seam is batch-vectorized
(whole RecordBatches, never row-at-a-time) and Send + Sync, so DataFusion can call a
loaded Compute per partition across all cores; a backend that has per-instance cost (the
wasm Store pool) amortizes it behind TransformBackend::load, which is invoked once and
whose Compute is then called per batch.
Enums§
Traits§
- Compute
- A loaded, ready-to-invoke transform. Cheap to call per batch (any expensive setup happened in
TransformBackend::load).Send + Syncso it can be shared across partition workers. - Transform
Backend - A pluggable backend for one
ImplKind. Implemented by the wasm/container/expression/builtin backends.loaddoes the expensive work once (compile the wasm module, parse the expression, resolve the builtin); the returnedComputeis the hot per-batch path.