fastflowtransform.executors.duckdb

DuckExecutor

Bases: SqlIdentifierMixin, SnapshotSqlMixin, BaseExecutor[DataFrame]

Source code in src/fastflowtransform/executors/duckdb.py

class DuckExecutor(SqlIdentifierMixin, SnapshotSqlMixin, BaseExecutor[pd.DataFrame]):
    ENGINE_NAME = "duckdb"

    _FIXED_TYPE_SIZES: ClassVar[dict[str, int]] = {
        "boolean": 1,
        "bool": 1,
        "tinyint": 1,
        "smallint": 2,
        "integer": 4,
        "int": 4,
        "bigint": 8,
        "float": 4,
        "real": 4,
        "double": 8,
        "double precision": 8,
        "decimal": 16,
        "numeric": 16,
        "uuid": 16,
        "json": 64,
        "jsonb": 64,
        "timestamp": 8,
        "timestamp_ntz": 8,
        "timestamp_ltz": 8,
        "timestamptz": 8,
        "date": 4,
        "time": 4,
        "interval": 16,
    }
    _VARCHAR_DEFAULT_WIDTH = 64
    _VARCHAR_MAX_WIDTH = 1024
    _DEFAULT_ROW_WIDTH = 128
    _BUDGET_GUARD = BudgetGuard(
        env_var="FF_DUCKDB_MAX_BYTES",
        estimator_attr="_estimate_query_bytes",
        engine_label="DuckDB",
        what="query",
    )

    def __init__(
        self, db_path: str = ":memory:", schema: str | None = None, catalog: str | None = None
    ):
        if db_path and db_path != ":memory:" and "://" not in db_path:
            with suppress(Exception):
                Path(db_path).parent.mkdir(parents=True, exist_ok=True)
        self.db_path = db_path
        self.con = duckdb.connect(db_path)
        self.schema = schema.strip() if isinstance(schema, str) and schema.strip() else None
        catalog_override = catalog.strip() if isinstance(catalog, str) and catalog.strip() else None
        self.catalog = self._detect_catalog()
        self._table_row_width_cache: dict[tuple[str | None, str], int] = {}
        if catalog_override:
            if self._apply_catalog_override(catalog_override):
                self.catalog = catalog_override
            else:
                self.catalog = self._detect_catalog()
        if self.schema:
            safe_schema = _q(self.schema)
            self._execute_sql(f"create schema if not exists {safe_schema}")
            self._execute_sql(f"set schema '{self.schema}'")

    def _execute_sql(self, sql: str, *args: Any, **kwargs: Any) -> duckdb.DuckDBPyConnection:
        """
        Central DuckDB SQL runner.

        All model-driven SQL in this executor should go through here.
        The cost guard may call _estimate_query_bytes(sql) before executing.
        This wrapper also records simple per-query stats for run_results.json.
        """

        def _exec() -> duckdb.DuckDBPyConnection:
            return self.con.execute(sql, *args, **kwargs)

        def _rows(result: Any) -> int | None:
            rc = getattr(result, "rowcount", None)
            if isinstance(rc, int) and rc >= 0:
                return rc
            return None

        return run_sql_with_budget(
            self,
            sql,
            guard=self._BUDGET_GUARD,
            exec_fn=_exec,
            rowcount_extractor=_rows,
            estimate_fn=self._estimate_query_bytes,
        )

    # --- Cost estimation for the shared BudgetGuard -----------------

    def _estimate_query_bytes(self, sql: str) -> int | None:
        """
        Estimate query size via DuckDB's EXPLAIN (FORMAT JSON).

        The JSON plan exposes an \"Estimated Cardinality\" per node.
        We walk the parsed tree, take the highest non-zero estimate and
        return it as a byte-estimate surrogate (row count ≈ bytes) so the
        cost guard can still make a meaningful decision without executing
        the query.
        """
        try:
            body = self._selectable_body(sql).strip().rstrip(";\n\t ")
        except AttributeError:
            body = sql.strip().rstrip(";\n\t ")

        lower = body.lower()
        if not lower.startswith(("select", "with")):
            return None

        explain_sql = f"EXPLAIN (FORMAT JSON) {body}"
        try:
            rows = self.con.execute(explain_sql).fetchall()
        except Exception:
            return None

        if not rows:
            return None

        fragments: list[str] = []
        for row in rows:
            for cell in row:
                if cell is None:
                    continue
                fragments.append(str(cell))

        if not fragments:
            return None

        plan_text = "\n".join(fragments).strip()
        start = plan_text.find("[")
        end = plan_text.rfind("]")
        if start == -1 or end == -1 or end <= start:
            return None

        try:
            plan_data = json.loads(plan_text[start : end + 1])
        except Exception:
            return None

        def _to_int(value: Any) -> int | None:
            if value is None:
                return None
            if isinstance(value, (int, float)):
                try:
                    converted = int(value)
                except Exception:
                    return None
                return converted
            text = str(value)
            match = re.search(r"(\d+(?:\.\d+)?)", text)
            if not match:
                return None
            try:
                return int(float(match.group(1)))
            except ValueError:
                return None

        def _walk_node(node: dict[str, Any]) -> int:
            best = 0
            extra = node.get("extra_info") or {}
            for key in (
                "Estimated Cardinality",
                "estimated_cardinality",
                "Cardinality",
                "cardinality",
            ):
                candidate = _to_int(extra.get(key))
                if candidate is not None:
                    best = max(best, candidate)
            candidate = _to_int(node.get("cardinality"))
            if candidate is not None:
                best = max(best, candidate)
            for child in node.get("children") or []:
                if isinstance(child, dict):
                    best = max(best, _walk_node(child))
            return best

        nodes: list[Any]
        nodes = plan_data if isinstance(plan_data, list) else [plan_data]

        estimate = 0
        for entry in nodes:
            if isinstance(entry, dict):
                estimate = max(estimate, _walk_node(entry))

        if estimate <= 0:
            return None

        tables = self._collect_tables_from_plan(nodes)
        row_width = self._row_width_for_tables(tables)
        if row_width <= 0:
            row_width = self._DEFAULT_ROW_WIDTH

        bytes_estimate = int(estimate * row_width)
        return bytes_estimate if bytes_estimate > 0 else None

    def _collect_tables_from_plan(self, nodes: list[dict[str, Any]]) -> set[tuple[str | None, str]]:
        tables: set[tuple[str | None, str]] = set()

        def _walk(entry: dict[str, Any]) -> None:
            extra = entry.get("extra_info") or {}
            table_val = extra.get("Table")
            schema_val = extra.get("Schema") or extra.get("Database") or extra.get("Catalog")
            if isinstance(table_val, str) and table_val.strip():
                schema, table = self._split_identifier(table_val, schema_val)
                if table:
                    tables.add((schema, table))
            for child in entry.get("children") or []:
                if isinstance(child, dict):
                    _walk(child)

        for node in nodes:
            if isinstance(node, dict):
                _walk(node)
        return tables

    def _split_identifier(
        self, identifier: str, explicit_schema: str | None
    ) -> tuple[str | None, str]:
        parts = [part.strip() for part in identifier.split(".") if part.strip()]
        if not parts:
            return explicit_schema, identifier
        if len(parts) >= 2:
            schema_candidate = self._strip_quotes(parts[-2])
            table_candidate = self._strip_quotes(parts[-1])
            return schema_candidate or explicit_schema, table_candidate
        return explicit_schema, self._strip_quotes(parts[-1])

    def _strip_quotes(self, value: str) -> str:
        if value.startswith('"') and value.endswith('"'):
            return value[1:-1]
        return value

    def _row_width_for_tables(self, tables: Iterable[tuple[str | None, str]]) -> int:
        widths: list[int] = []
        for schema, table in tables:
            width = self._row_width_for_table(schema, table)
            if width > 0:
                widths.append(width)
        return max(widths) if widths else 0

    def _row_width_for_table(self, schema: str | None, table: str) -> int:
        key = (schema or "", table.lower())
        cached = self._table_row_width_cache.get(key)
        if cached:
            return cached

        columns = self._columns_for_table(table, schema)
        width = sum(self._estimate_column_width(col) for col in columns)
        if width <= 0:
            width = self._DEFAULT_ROW_WIDTH
        self._table_row_width_cache[key] = width
        return width

    def _columns_for_table(
        self, table: str, schema: str | None
    ) -> list[tuple[str | None, int | None, int | None, int | None]]:
        table_lower = table.lower()
        columns: list[tuple[str | None, int | None, int | None, int | None]] = []
        seen_schemas: set[str | None] = set()
        for candidate in self._schema_candidates(schema):
            if candidate in seen_schemas:
                continue
            seen_schemas.add(candidate)
            if candidate is not None:
                try:
                    rows = self.con.execute(
                        """
                        select lower(data_type) as dtype,
                               character_maximum_length,
                               numeric_precision,
                               numeric_scale
                        from information_schema.columns
                        where lower(table_name)=lower(?)
                          and lower(table_schema)=lower(?)
                        order by ordinal_position
                        """,
                        [table_lower, candidate.lower()],
                    ).fetchall()
                except Exception:
                    continue
            else:
                try:
                    rows = self.con.execute(
                        """
                        select lower(data_type) as dtype,
                               character_maximum_length,
                               numeric_precision,
                               numeric_scale
                        from information_schema.columns
                        where lower(table_name)=lower(?)
                        order by lower(table_schema), ordinal_position
                        """,
                        [table_lower],
                    ).fetchall()
                except Exception:
                    continue
            if rows:
                return rows
        return columns

    def _schema_candidates(self, schema: str | None) -> list[str | None]:
        candidates: list[str | None] = []

        def _add(value: str | None) -> None:
            normalized = self._normalize_schema(value)
            if normalized not in candidates:
                candidates.append(normalized)

        _add(schema)
        _add(self.schema)
        for alt in ("main", "temp"):
            _add(alt)
        _add(None)
        return candidates

    def _normalize_schema(self, schema: str | None) -> str | None:
        if not schema:
            return None
        stripped = schema.strip()
        return stripped or None

    def _estimate_column_width(
        self, column_info: tuple[str | None, int | None, int | None, int | None]
    ) -> int:
        dtype_raw, char_max, numeric_precision, _ = column_info
        dtype = self._normalize_data_type(dtype_raw)
        if dtype and dtype in self._FIXED_TYPE_SIZES:
            return self._FIXED_TYPE_SIZES[dtype]

        if dtype in {"character", "varchar", "char", "text", "string"}:
            if char_max and char_max > 0:
                return min(char_max, self._VARCHAR_MAX_WIDTH)
            return self._VARCHAR_DEFAULT_WIDTH

        if dtype in {"varbinary", "blob", "binary"}:
            if char_max and char_max > 0:
                return min(char_max, self._VARCHAR_MAX_WIDTH)
            return self._VARCHAR_DEFAULT_WIDTH

        if dtype in {"numeric", "decimal"} and numeric_precision and numeric_precision > 0:
            return min(max(int(numeric_precision), 16), 128)

        return 16

    def _normalize_data_type(self, dtype: str | None) -> str | None:
        if not dtype:
            return None
        stripped = dtype.strip().lower()
        if "(" in stripped:
            stripped = stripped.split("(", 1)[0].strip()
        if stripped.endswith("[]"):
            stripped = stripped[:-2]
        return stripped or None

    def _detect_catalog(self) -> str | None:
        try:
            rows = self._execute_sql("PRAGMA database_list").fetchall()
            if rows:
                return str(rows[0][1])
        except Exception:
            return None
        return None

    def _apply_catalog_override(self, name: str) -> bool:
        alias = name.strip()
        if not alias:
            return False
        try:
            if self.db_path != ":memory:":
                resolved = str(Path(self.db_path).resolve())
                with suppress(Exception):
                    self._execute_sql(f"detach database {_q(alias)}")
                self._execute_sql(f"attach database '{resolved}' as {_q(alias)} (READ_ONLY FALSE)")
            self._execute_sql(f"set catalog '{alias}'")
            return True
        except Exception:
            return False

    def clone(self) -> DuckExecutor:
        """
        Generates a new Executor instance with own connection for Thread-Worker.
        """
        return DuckExecutor(self.db_path, schema=self.schema, catalog=self.catalog)

    def _exec_many(self, sql: str) -> None:
        """
        Execute multiple SQL statements separated by ';' on the same connection.
        DuckDB normally accepts one statement per execute(), so we split here.
        """
        for stmt in (part.strip() for part in sql.split(";")):
            if not stmt:
                continue
            self._execute_sql(stmt)

    # ---- Frame hooks ----
    def _quote_identifier(self, ident: str) -> str:
        return _q(ident)

    def _should_include_catalog(
        self, catalog: str | None, schema: str | None, *, explicit: bool
    ) -> bool:
        """
        DuckDB includes catalog only when explicitly provided or when it matches
        the schema (mirrors previous behaviour).
        """
        if explicit:
            return bool(catalog)
        return bool(catalog and schema and catalog.lower() == schema.lower())

    def _default_catalog_for_source(self, schema: str | None) -> str | None:
        """
        For sources, fall back to DuckDB's detected catalog when:
        - schema is set and matches the catalog, or
        - neither schema nor catalog was provided (keep old fallback)
        """
        cat = self._default_catalog()
        if not cat:
            return None
        if schema is None or cat.lower() == schema.lower():
            return cat
        return None

    def _qualified(self, relation: str, *, quoted: bool = True) -> str:
        """
        Return (catalog.)schema.relation if schema is set; otherwise just relation.
        When quoted=False, emit bare identifiers for APIs like con.table().
        """
        return self._qualify_identifier(relation, quote=quoted)

    def _read_relation(self, relation: str, node: Node, deps: Iterable[str]) -> pd.DataFrame:
        try:
            target = self._qualified(relation, quoted=False)
            return self.con.table(target).df()
        except CatalogException as e:
            existing = [
                r[0]
                for r in self._execute_sql(
                    "select table_name from information_schema.tables "
                    "where table_schema in ('main','temp')"
                ).fetchall()
            ]
            raise RuntimeError(
                f"Dependency table not found: '{relation}'\n"
                f"Deps: {list(deps)}\nExisting tables: {existing}\n"
                "Note: Use same File-DB/Connection for Seeding & Run."
            ) from e

    def _materialize_relation(self, relation: str, df: pd.DataFrame, node: Node) -> None:
        tmp = "_ff_py_out"
        try:
            self.con.register(tmp, df)
            target = self._qualified(relation)
            self._execute_sql(f'create or replace table {target} as select * from "{tmp}"')
        finally:
            try:
                self.con.unregister(tmp)
            except Exception:
                # housekeeping only; stats here are not important but harmless if recorded
                self._execute_sql(f'drop view if exists "{tmp}"')

    def _create_or_replace_view_from_table(
        self, view_name: str, backing_table: str, node: Node
    ) -> None:
        view_target = self._qualified(view_name)
        backing = self._qualified(backing_table)
        self._execute_sql(f"create or replace view {view_target} as select * from {backing}")

    def _frame_name(self) -> str:
        return "pandas"

    # ---- SQL hooks ----
    def _create_or_replace_view(self, target_sql: str, select_body: str, node: Node) -> None:
        self._execute_sql(f"create or replace view {target_sql} as {select_body}")

    def _create_or_replace_table(self, target_sql: str, select_body: str, node: Node) -> None:
        self._execute_sql(f"create or replace table {target_sql} as {select_body}")

    # ---- Meta hook ----
    def on_node_built(self, node: Node, relation: str, fingerprint: str) -> None:
        """
        After successful materialization, ensure the meta table exists and upsert the row.
        """
        ensure_meta_table(self)
        upsert_meta(self, node.name, relation, fingerprint, "duckdb")

    # ── Incremental API ────────────────────────────────────────────────────
    def exists_relation(self, relation: str) -> bool:
        where_tables: list[str] = ["lower(table_name) = lower(?)"]
        params: list[str] = [relation]
        if self.catalog:
            where_tables.append("lower(table_catalog) = lower(?)")
            params.append(self.catalog)
        if self.schema:
            where_tables.append("lower(table_schema) = lower(?)")
            params.append(self.schema)
        else:
            where_tables.append("table_schema in ('main','temp')")
        where = " AND ".join(where_tables)
        sql_tables = f"select 1 from information_schema.tables where {where} limit 1"
        if self._execute_sql(sql_tables, params).fetchone():
            return True
        sql_views = f"select 1 from information_schema.views where {where} limit 1"
        return bool(self._execute_sql(sql_views, params).fetchone())

    def create_table_as(self, relation: str, select_sql: str) -> None:
        # Use only the SELECT body and strip trailing semicolons for safety.
        body = self._selectable_body(select_sql).strip().rstrip(";\n\t ")
        self._execute_sql(f"create table {self._qualified(relation)} as {body}")

    def incremental_insert(self, relation: str, select_sql: str) -> None:
        # Ensure the inner SELECT is clean (no trailing semicolon; SELECT body only).
        body = self._selectable_body(select_sql).strip().rstrip(";\n\t ")
        self._execute_sql(f"insert into {self._qualified(relation)} {body}")

    def incremental_merge(self, relation: str, select_sql: str, unique_key: list[str]) -> None:
        """
        Fallback strategy for DuckDB:
        - DELETE collisions via DELETE ... USING (<select>) s
        - INSERT all rows via INSERT ... SELECT * FROM (<select>)
        """
        # 1) clean inner SELECT
        body = self._selectable_body(select_sql).strip().rstrip(";\n\t ")

        # 2) predicate for DELETE
        keys_pred = " AND ".join([f"t.{k}=s.{k}" for k in unique_key]) or "FALSE"

        # 3) first: delete collisions
        delete_sql = f"delete from {self._qualified(relation)} t using ({body}) s where {keys_pred}"
        self._execute_sql(delete_sql)

        # 4) then: insert fresh rows
        insert_sql = f"insert into {self._qualified(relation)} select * from ({body}) src"
        self._execute_sql(insert_sql)

    def alter_table_sync_schema(
        self, relation: str, select_sql: str, *, mode: str = "append_new_columns"
    ) -> None:
        """
        Best-effort: add new columns with inferred type.
        """
        # Probe: empty projection from the SELECT (cleaned to avoid parser issues).
        body = self._first_select_body(select_sql).strip().rstrip(";\n\t ")
        probe = self._execute_sql(f"select * from ({body}) as q limit 0")
        cols = [c[0] for c in probe.description or []]
        existing = {
            r[0]
            for r in self._execute_sql(
                "select column_name from information_schema.columns "
                + "where lower(table_name)=lower(?)"
                + (" and lower(table_schema)=lower(?)" if self.schema else ""),
                ([relation, self.schema] if self.schema else [relation]),
            ).fetchall()
        }
        add = [c for c in cols if c not in existing]
        for c in add:
            col = _q(c)
            target = self._qualified(relation)
            try:
                self._execute_sql(f"alter table {target} add column {col} varchar")
            except Exception:
                self._execute_sql(f"alter table {target} add column {col} varchar")

    def execute_hook_sql(self, sql: str) -> None:
        """
        Execute one or multiple SQL statements for pre/post/on_run hooks.

        Accepts a string that may contain ';'-separated statements.
        """
        self._exec_many(sql)

    # ---- Snapshot mixin hooks ----
    def _snapshot_target_identifier(self, rel_name: str) -> str:
        return self._qualified(rel_name)

    def _snapshot_current_timestamp(self) -> str:
        return "current_timestamp"

    def _snapshot_null_timestamp(self) -> str:
        return "cast(null as timestamp)"

    def _snapshot_null_hash(self) -> str:
        return "cast(null as varchar)"

    def _snapshot_hash_expr(self, check_cols: list[str], src_alias: str) -> str:
        concat_expr = self._snapshot_concat_expr(check_cols, src_alias)
        return f"cast(md5({concat_expr}) as varchar)"

    def _snapshot_cast_as_string(self, expr: str) -> str:
        return f"cast({expr} as varchar)"

    def _snapshot_source_ref(
        self, rel_name: str, select_body: str
    ) -> tuple[str, Callable[[], None]]:
        src_view_name = f"__ff_snapshot_src_{rel_name}".replace(".", "_")
        src_quoted = _q(src_view_name)
        self._execute_sql(f"create or replace temp view {src_quoted} as {select_body}")

        def _cleanup() -> None:
            self._execute_sql(f"drop view if exists {src_quoted}")

        return src_quoted, _cleanup

        # ---- Unit-test helpers -------------------------------------------------

    def utest_load_relation_from_rows(self, relation: str, rows: list[dict]) -> None:
        """
        Load rows into a DuckDB table for unit tests, fully qualified to
        this executor's schema/catalog.
        """
        df = pd.DataFrame(rows)
        tmp = f"_ff_utest_tmp_{uuid.uuid4().hex[:12]}"
        self.con.register(tmp, df)
        try:
            target = self._qualified(relation)
            self._execute_sql(f"create or replace table {target} as select * from {tmp}")
        finally:
            with suppress(Exception):
                self.con.unregister(tmp)
            # Fallback for older DuckDB where unregister might not exist
            with suppress(Exception):
                self._execute_sql(f'drop view if exists "{tmp}"')

    def utest_read_relation(self, relation: str) -> pd.DataFrame:
        """
        Read a relation as a DataFrame for unit-test assertions.
        """
        target = self._qualified(relation, quoted=False)
        return self.con.table(target).df()

    def utest_clean_target(self, relation: str) -> None:
        """
        Drop any table/view with the given name in this schema/catalog.
        Safe because utest uses its own DB/path.
        """
        target = self._qualified(relation)
        # best-effort; ignore failures
        with suppress(Exception):
            self._execute_sql(f"drop view if exists {target}")
        with suppress(Exception):
            self._execute_sql(f"drop table if exists {target}")

clone

clone()

Generates a new Executor instance with own connection for Thread-Worker.

Source code in src/fastflowtransform/executors/duckdb.py

def clone(self) -> DuckExecutor:
    """
    Generates a new Executor instance with own connection for Thread-Worker.
    """
    return DuckExecutor(self.db_path, schema=self.schema, catalog=self.catalog)

on_node_built

on_node_built(node, relation, fingerprint)

After successful materialization, ensure the meta table exists and upsert the row.

Source code in src/fastflowtransform/executors/duckdb.py

def on_node_built(self, node: Node, relation: str, fingerprint: str) -> None:
    """
    After successful materialization, ensure the meta table exists and upsert the row.
    """
    ensure_meta_table(self)
    upsert_meta(self, node.name, relation, fingerprint, "duckdb")

incremental_merge

incremental_merge(relation, select_sql, unique_key)

Fallback strategy for DuckDB: - DELETE collisions via DELETE ... USING () s - INSERT all rows via INSERT ... SELECT * FROM ()

Source code in src/fastflowtransform/executors/duckdb.py

def incremental_merge(self, relation: str, select_sql: str, unique_key: list[str]) -> None:
    """
    Fallback strategy for DuckDB:
    - DELETE collisions via DELETE ... USING (<select>) s
    - INSERT all rows via INSERT ... SELECT * FROM (<select>)
    """
    # 1) clean inner SELECT
    body = self._selectable_body(select_sql).strip().rstrip(";\n\t ")

    # 2) predicate for DELETE
    keys_pred = " AND ".join([f"t.{k}=s.{k}" for k in unique_key]) or "FALSE"

    # 3) first: delete collisions
    delete_sql = f"delete from {self._qualified(relation)} t using ({body}) s where {keys_pred}"
    self._execute_sql(delete_sql)

    # 4) then: insert fresh rows
    insert_sql = f"insert into {self._qualified(relation)} select * from ({body}) src"
    self._execute_sql(insert_sql)

alter_table_sync_schema

alter_table_sync_schema(relation, select_sql, *, mode='append_new_columns')

Best-effort: add new columns with inferred type.

Source code in src/fastflowtransform/executors/duckdb.py

def alter_table_sync_schema(
    self, relation: str, select_sql: str, *, mode: str = "append_new_columns"
) -> None:
    """
    Best-effort: add new columns with inferred type.
    """
    # Probe: empty projection from the SELECT (cleaned to avoid parser issues).
    body = self._first_select_body(select_sql).strip().rstrip(";\n\t ")
    probe = self._execute_sql(f"select * from ({body}) as q limit 0")
    cols = [c[0] for c in probe.description or []]
    existing = {
        r[0]
        for r in self._execute_sql(
            "select column_name from information_schema.columns "
            + "where lower(table_name)=lower(?)"
            + (" and lower(table_schema)=lower(?)" if self.schema else ""),
            ([relation, self.schema] if self.schema else [relation]),
        ).fetchall()
    }
    add = [c for c in cols if c not in existing]
    for c in add:
        col = _q(c)
        target = self._qualified(relation)
        try:
            self._execute_sql(f"alter table {target} add column {col} varchar")
        except Exception:
            self._execute_sql(f"alter table {target} add column {col} varchar")

execute_hook_sql

execute_hook_sql(sql)

Execute one or multiple SQL statements for pre/post/on_run hooks.

Accepts a string that may contain ';'-separated statements.

Source code in src/fastflowtransform/executors/duckdb.py

def execute_hook_sql(self, sql: str) -> None:
    """
    Execute one or multiple SQL statements for pre/post/on_run hooks.

    Accepts a string that may contain ';'-separated statements.
    """
    self._exec_many(sql)

utest_load_relation_from_rows

utest_load_relation_from_rows(relation, rows)

Load rows into a DuckDB table for unit tests, fully qualified to this executor's schema/catalog.

Source code in src/fastflowtransform/executors/duckdb.py

def utest_load_relation_from_rows(self, relation: str, rows: list[dict]) -> None:
    """
    Load rows into a DuckDB table for unit tests, fully qualified to
    this executor's schema/catalog.
    """
    df = pd.DataFrame(rows)
    tmp = f"_ff_utest_tmp_{uuid.uuid4().hex[:12]}"
    self.con.register(tmp, df)
    try:
        target = self._qualified(relation)
        self._execute_sql(f"create or replace table {target} as select * from {tmp}")
    finally:
        with suppress(Exception):
            self.con.unregister(tmp)
        # Fallback for older DuckDB where unregister might not exist
        with suppress(Exception):
            self._execute_sql(f'drop view if exists "{tmp}"')

utest_read_relation

utest_read_relation(relation)

Read a relation as a DataFrame for unit-test assertions.

Source code in src/fastflowtransform/executors/duckdb.py

def utest_read_relation(self, relation: str) -> pd.DataFrame:
    """
    Read a relation as a DataFrame for unit-test assertions.
    """
    target = self._qualified(relation, quoted=False)
    return self.con.table(target).df()

utest_clean_target

utest_clean_target(relation)

Drop any table/view with the given name in this schema/catalog. Safe because utest uses its own DB/path.

Source code in src/fastflowtransform/executors/duckdb.py

def utest_clean_target(self, relation: str) -> None:
    """
    Drop any table/view with the given name in this schema/catalog.
    Safe because utest uses its own DB/path.
    """
    target = self._qualified(relation)
    # best-effort; ignore failures
    with suppress(Exception):
        self._execute_sql(f"drop view if exists {target}")
    with suppress(Exception):
        self._execute_sql(f"drop table if exists {target}")

run_sql

run_sql(node, env)

Orchestrate SQL models

1) Render Jinja (ref/source/this) and strip leading {{ config(...) }}. 2) If the SQL is full DDL (CREATE …), execute it verbatim (passthrough). 3) Otherwise, normalize to CREATE OR REPLACE {TABLE|VIEW} AS . The body is CTE-aware (keeps WITH … SELECT … intact).

On failure, raise ModelExecutionError with a helpful snippet.

Source code in src/fastflowtransform/executors/base.py

def run_sql(self, node: Node, env: Environment) -> None:
    """
    Orchestrate SQL models:
      1) Render Jinja (ref/source/this) and strip leading {{ config(...) }}.
      2) If the SQL is full DDL (CREATE …), execute it verbatim (passthrough).
      3) Otherwise, normalize to CREATE OR REPLACE {TABLE|VIEW} AS <body>.
         The body is CTE-aware (keeps WITH … SELECT … intact).
    On failure, raise ModelExecutionError with a helpful snippet.
    """
    meta = getattr(node, "meta", {}) or {}
    if self._meta_is_incremental(meta):
        # Delegates to incremental engine: render, schema sync, merge/insert, etc.
        return _ff_incremental.run_or_dispatch(self, node, env)

    if self._meta_is_snapshot(meta):
        # Snapshots are executed via the dedicated CLI: `fft snapshot run`.
        raise ModelExecutionError(
            node_name=node.name,
            relation=relation_for(node.name),
            message=(
                "Snapshot models cannot be executed via 'fft run'. "
                "Use 'fft snapshot run' instead."
            ),
            sql_snippet="",
        )

    sql_rendered = self.render_sql(
        node,
        env,
        ref_resolver=lambda name: self._resolve_ref(name, env),
        source_resolver=self._resolve_source,
    )
    sql = self._strip_leading_config(sql_rendered).strip()

    materialization = (node.meta or {}).get("materialized", "table")
    if materialization == "ephemeral":
        return

    # 1) Direct DDL passthrough (CREATE [OR REPLACE] {TABLE|VIEW} …)
    if self._looks_like_direct_ddl(sql):
        try:
            self._execute_sql_direct(sql, node)
            return
        except NotImplementedError:
            # Engine doesn't implement direct DDL → fall back to normalized materialization.
            pass
        except Exception as e:
            raise ModelExecutionError(
                node_name=node.name,
                relation=relation_for(node.name),
                message=str(e),
                sql_snippet=sql,
            ) from e

    # 2) Normalized materialization path (CTE-safe body)
    body = self._selectable_body(sql).rstrip(" ;\n\t")
    target_sql = self._format_relation_for_ref(node.name)

    # Centralized SQL preview logging (applies to ALL engines)
    preview = (
        f"=== MATERIALIZE ===\n"
        f"-- model: {node.name}\n"
        f"-- materialized: {materialization}\n"
        f"-- target: {target_sql}\n"
        f"{body}\n"
    )
    echo_debug(preview)

    try:
        self._apply_sql_materialization(node, target_sql, body, materialization)
    except Exception as e:
        preview = f"-- materialized={materialization}\n-- target={target_sql}\n{body}"
        raise ModelExecutionError(
            node_name=node.name,
            relation=relation_for(node.name),
            message=str(e),
            sql_snippet=preview,
        ) from e

configure_query_budget_limit

configure_query_budget_limit(limit)

Inject a configured per-query byte limit (e.g. from budgets.yml).

Source code in src/fastflowtransform/executors/base.py

def configure_query_budget_limit(self, limit: int | None) -> None:
    """
    Inject a configured per-query byte limit (e.g. from budgets.yml).
    """
    if limit is None:
        self._ff_configured_query_limit = None
        return
    try:
        iv = int(limit)
    except Exception:
        self._ff_configured_query_limit = None
        return
    self._ff_configured_query_limit = iv if iv > 0 else None

reset_node_stats

reset_node_stats()

Reset per-node statistics buffer.

The run engine calls this before executing a model so that all stats recorded via _record_query_stats(...) belong to that node.

Source code in src/fastflowtransform/executors/base.py

def reset_node_stats(self) -> None:
    """
    Reset per-node statistics buffer.

    The run engine calls this before executing a model so that all
    stats recorded via `_record_query_stats(...)` belong to that node.
    """
    # just clear the buffer; next recording will re-create it
    self._ff_query_stats_buffer = []

get_node_stats

get_node_stats()

Aggregate buffered QueryStats into a simple dict:

{
  "bytes_scanned": <sum>,
  "rows": <sum>,
  "query_duration_ms": <sum>,
}

Called by the run engine after a node finishes.

Source code in src/fastflowtransform/executors/base.py

def get_node_stats(self) -> dict[str, int]:
    """
    Aggregate buffered QueryStats into a simple dict:

        {
          "bytes_scanned": <sum>,
          "rows": <sum>,
          "query_duration_ms": <sum>,
        }

    Called by the run engine after a node finishes.
    """
    stats_list = self._drain_query_stats()
    if not stats_list:
        return {}

    total_bytes = 0
    total_rows = 0
    total_duration = 0

    for s in stats_list:
        if s.bytes_processed is not None:
            total_bytes += int(s.bytes_processed)
        if s.rows is not None:
            total_rows += int(s.rows)
        if s.duration_ms is not None:
            total_duration += int(s.duration_ms)

    return {
        "bytes_scanned": total_bytes,
        "rows": total_rows,
        "query_duration_ms": total_duration,
    }

run_python

run_python(node)

Execute the Python model for a given node and materialize its result.

Source code in src/fastflowtransform/executors/base.py

def run_python(self, node: Node) -> None:
    """Execute the Python model for a given node and materialize its result."""
    func = REGISTRY.py_funcs[node.name]
    deps = REGISTRY.nodes[node.name].deps or []

    self._reset_http_ctx(node)

    # arg = self._build_python_args(node, deps)
    args, argmap = self._build_python_inputs(node, deps)
    requires = REGISTRY.py_requires.get(node.name, {})
    # if deps:
    #     self._validate_required(node.name, arg, requires)
    if deps:
        # Required-columns check works against the mapping
        self._validate_required(node.name, argmap, requires)

    # out = self._execute_python_func(func, arg, node)
    out = self._execute_python_func(func, args, node)

    target = relation_for(node.name)
    meta = getattr(node, "meta", {}) or {}
    mat = self._resolve_materialization_strategy(meta)

    if mat == "incremental":
        self._materialize_incremental(target, out, node, meta)
    elif mat == "view":
        self._materialize_view(target, out, node)
    else:
        self._materialize_relation(target, out, node)

    self._snapshot_http_ctx(node)

snapshot_prune

snapshot_prune(relation, unique_key, keep_last, *, dry_run=False)

Delete older snapshot versions while keeping the most recent keep_last rows per business key (including the current row).

Source code in src/fastflowtransform/executors/_snapshot_sql_mixin.py

    def snapshot_prune(
        self,
        relation: str,
        unique_key: list[str],
        keep_last: int,
        *,
        dry_run: bool = False,
    ) -> None:
        """
        Delete older snapshot versions while keeping the most recent `keep_last`
        rows per business key (including the current row).
        """
        ex = cast("BaseExecutor[Any]", self)

        if keep_last <= 0:
            return

        keys = [k for k in unique_key if k]
        if not keys:
            return

        target = self._snapshot_target_identifier(relation)
        vf = self.SNAPSHOT_VALID_FROM_COL  # type: ignore[attr-defined]

        key_select = ", ".join(keys)
        part_by = ", ".join(keys)

        ranked_sql = f"""
SELECT
  {key_select},
  {vf},
  ROW_NUMBER() OVER (
    PARTITION BY {part_by}
    ORDER BY {vf} DESC
  ) AS rn
FROM {target}
"""

        if dry_run:
            sql = f"""
WITH ranked AS (
  {ranked_sql}
)
SELECT COUNT(*) AS rows_to_delete
FROM ranked
WHERE rn > {int(keep_last)}
"""
            res = ex._execute_sql(sql)
            count = self._snapshot_fetch_count(res)
            echo(
                f"[DRY-RUN] snapshot_prune({relation}): would delete {count} row(s) "
                f"(keep_last={keep_last})"
            )
            return

        join_pred = " AND ".join([f"t.{k} = r.{k}" for k in keys])
        delete_sql = f"""
DELETE FROM {target} t
USING (
  {ranked_sql}
) r
WHERE
  r.rn > {int(keep_last)}
  AND {join_pred}
  AND t.{vf} = r.{vf}
"""
        ex._execute_sql(delete_sql)