Urgent.News

600+ sources. One page. See who else covered it.

Editions

Tech

The IR Is a sqlglot AST

Every data tool has an intermediate representation, whether it admits to one or not. It is the thing a model becomes after parsing and before execution, and it quietly decides what the tool can do. dbt's IR is templated text, which is why ref() is a string and why a macro can produce SQL no analyser can reason about. An earlier iteration of Interlace effectively had a pandas DataFrame as its IR —…

The canonical intermediate representation (IR) for Interlace 2.0 is a sqlglot AST combined with Arrow schema. This IR sets the ceiling for what the system can do. The materialization process occurs once at the sink as a single native SQL statement executed inside the owning engine. This design decision led to the removal of Ibis, a data plane and expression builder library that added heavyweight dependencies and governance risks for no additional capability.

Remote engines now connect over ADBC, a narrower and better-specified contract than Ibis backends. The sqlglot AST allows strategies to be AST builders rather than string templates, with dialect appearing only at transpile(). The IR enables dependencies to be parsed rather than declared, and FROM/JOIN clauses to be structure rather than text.

References that match another model become DAG edges, with CTEs excluded. Semantic change classification is possible due to the ability to compare ASTs rather than strings. The wire format is an Arrow RecordBatchReader, allowing data to be streamed with bounded memory usage. The core install has nine runtime dependencies, with optional extras for pandas and an HTTP daemon.

The IR consists of a sqlglot AST, Arrow RecordBatchReader wire format, dialect-only-at-transpile, and DuckLake storage defaults. The control plane uses SQLite in WAL mode, with streams in their own durable log and exactly-once materialization. The rebuild eliminates Jinja, ref() as text, and pandas in the core, resulting in a system with five lines: IR is a sqlglot AST, wire format is an Arrow RecordBatchReader, strategies are AST builders with dialect at transpile, storage defaults to DuckLake, and the control plane uses SQLite in WAL mode.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in Tech

Beyond the Demo: Building Production-Ready AI Agents — A Guide to Benchmarking, Cost Optimization, and Tooling in 2026

Originally published on tamiz.pro . Most AI agents ship from a notebook, impress in a demo, and quietly fail in production.

  • Production AI agents need observability, evaluation rigor, and cost discipline beyond demos.
  • Benchmarks should include task-level tests, edge cases, and a weighted scoring rubric.
  • Cost optimization techniques include tiered model routing, prompt compression, and token budgeting.

More from Saturday 15 August →