"""从查询快照建口径草稿图:同一份输入连着建两次必须一模一样。 build_graph 抽自 freeze。草稿图要落成工作流版本、算 graph_hash,节点顺序、位置、报告配置、指标 expression 只要有一点 抖动,同一份快照就会算出另一个版本、比对也认不出等价;所以确定性单独盯一条。 """ from __future__ import annotations import json from app.ledger.certify import build_graph, shape_of SQL = ("SELECT SUM(amount) AS gmv FROM orders WHERE status = 'paid' " "AND paid_at >= '2026-09-01' AND paid_at < '2026-09-08'") KINDS = {"shop": "sqlite"} def snapshot() -> dict: return {"source": "shop", "sql": SQL, "columns": ["gmv"], "rows": [[4231.5]]} async def _labels(_snap: dict, _source: str): return lambda _key: None async def test_build_graph_twice_is_identical(): shape = shape_of(snapshot(), KINDS) assert shape is not None and shape.problem is None # 别的运行里两个日期都变了、常量 'paid' 没变:两次建图应当都认出一样的 p1 / p2 shifted = [tuple(lit.replace("2026-09-01", start).replace("2026-09-08", end) for lit in shape.canonical) for start, end in (("2026-09-02", "2026-09-09"), ("2026-09-03", "2026-09-10"))] call = dict(shapes=[shape], snaps=[snapshot()], kinds=KINDS, labels=_labels, name="九月第一周销售额", question="九月第一周卖了多少", peers={(shape.source, shape.key): shifted}) first = await build_graph(**call) second = await build_graph(**call) assert [f["name"] for f in first[1]] == ["p1", "p2"] assert first == second assert json.dumps(first[0], ensure_ascii=False) == json.dumps(second[0], ensure_ascii=False)