Roadmap

What is not implemented is listed in parity, generated from a run against the image. This page is about which of those absences are next, and why.

Next

MERGE. Ordinary Snowflake, and what an incremental dbt model emits. No model in the Contoso product needs it today, which is why it has waited.

OBJECT_CONSTRUCT. The one common semi-structured constructor still missing.

REMOVE. The counterpart to PUT, which shipped. A stage you can write and never clear grows for the life of the container.

COPY INTO from a prefix. INFER_SCHEMA describes a whole feed now, but COPY INTO still resolves one name, so a paged feed is a loop in the caller. The two ought to take the same reference.

Stored procedures. CALL and SYSTEM$WAIT are refused. Listed here rather than left unclassified: this page previously named neither these nor REMOVE in either section, which for a project whose whole claim is that a gap can be planned around is the documentation version of the defect it guards against.

Done since this page was written

Typed loading, via INFER_SCHEMA. The entry that used to head this list. COPY INTO fills a table that already exists, so with nothing to ask, both Contoso Snowflake leaves grew their own sampler — read some rows, classify each cell, widen, emit DDL. Two copies of a type-inference heuristic living in the product because the warehouse would not answer the question. It answers now, and the answer carries Snowflake’s own type names, which meant teaching the emulator to accept them in DDL as well: CREATE TABLE t (a NUMBER(38,0)) was refused by its own inference until that landed. Measured, not assumed — see docs/06-stages-and-copy.md.

Deliberately not planned

DATEADD with month, quarter or year. Not effort — a limit. Every DuckDB spelling widens a DATE to a TIMESTAMP, and SQL unifies a CASE to one type, so no rewrite can return the type Snowflake returns. Answering with the wrong type is worse than refusing.

Time Travel and CLONE. They need a versioned store; DuckDB has none.

GRANT and roles. An access model that is accepted and not enforced is worse than one that is absent, because a consumer would test against it.

Snowflake SQL compatibility as a claim. The result names dialect: duckdb and always will. What this project promises is that a statement a real account accepts is either answered or refused by name — never quietly ignored.

The rule this project keeps relearning

Three separate defects this year were the same shape: something reported success while doing nothing. A CLI’s exit code trusted, a stream that would have missed an update, a COPY INTO that loaded no rows and said ok. Each was found by measuring rather than by reading, and each is now guarded by a test that fails on the old behaviour.

If you are adding to this emulator: make the failure honest first, then implement. The order matters — until refusals were honest, eighteen missing constructs were invisible, including several nobody had thought to ask about.