common_dagster.checks
checks.py — the ready-made check factories (Dagster asset checks underneath).
A factory targets the step’s LANDED table (it requires produces=) and queries it through the flow’s DBClient at check time. Custom logic = @check methods on the flow class; these factories cover the generic floor.
All SQL is COMPOSED (identifiers via psycopg.sql, values as params) — the house law; a factory never interpolates. row_delta’s previous count is DURABLE through the store’s stash/retrieve (a fresh pod per run means process memory is amnesia by construction — the old per-process dict could never fire).
Classes
CheckFactory
A declared generic check: named, blocking-optional, table-addressed.
__init__
def __init__(kind: str, params: dict, blocking: bool = False) -> NoneFields and properties declared here; inherited members belong to the bases above.
blocking
blocking: bool = Falsekind
kind: strname
Read-only property.
name: strparams
params: dictevaluate
def evaluate(db: Any, table: str) -> tuple[bool, dict]Run against the landed table → (passed, metadata).
Functions
fresh
def fresh(days: int, column: str = 'date', *, blocking: bool = False) -> CheckFactorymax(column) is within the last N days — the staleness tripwire.
nonempty
def nonempty(*, blocking: bool = False) -> CheckFactoryThe table has rows at all.
row_delta
def row_delta(max_drop_pct: float, *, blocking: bool = True) -> CheckFactoryRow count never DROPS more than max_drop_pct vs the last DURABLY seen count (a shrinking raw table = something destructive happened). The previous count survives process boundaries via the store table.
unique_key
def unique_key(*columns: str, blocking: bool = False) -> CheckFactoryNo duplicate rows on the given key columns (the census-PK guard).