Batch execution

A batch query consumes finite inputs and produces a finite result. Krishiv uses DataFusion for SQL planning and Arrow batches for data exchange.

Lifecycle

1. A session registers tables, files, or providers in its catalog.
2. SQL or a DataFrame expression becomes a logical plan.
3. Optimization produces a physical plan.
4. The runtime chooses local or remote placement.
5. Operators emit Arrow RecordBatch values.
6. The caller collects, streams, prints, or stores the result.

Registration is explicit

Names in SQL resolve through the session catalog. A CLI --parquet registration lasts for that process; a long-lived service can own a longer-lived catalog.

Local and remote results

Embedded execution can return results directly. Remote execution uses public Flight/gRPC and coordinator control paths. Keep result size and lifetime in mind: large results should be consumed as batches or written to a sink instead of materialized into one client-side object.

Multi-statement SQL

The CLI accepts semicolon-separated statements. Earlier statements run for side effects; only the final statement result is printed.

krishiv sql --query \
  "CREATE TABLE t (id INT); INSERT INTO t VALUES (1); SELECT * FROM t"

SQL surface and DDL behavior are pre-release. Pin the Engine version when you store scripts outside the repository.