降档时报告页顶部写明出具档位和原因(B5):出具声明和 issuance 事件另带结构化的 sql_checks(哪次查询、问题、受影响的引用、对应的缺口),证据页签在报告上方放出具横幅,SQL 检查导致的降档单列一块写在显眼处、不在其余缺口里重复;不是完整出具时「N 个数字都有出处」不再画成绿色

This commit is contained in:
ThearchyHelios
2026-10-04 02:33:15 +08:00
parent 31531925a9
commit 6263725c2a
12 changed files with 196 additions and 18 deletions
+17 -4
View File
@@ -466,6 +466,12 @@ def incomplete_gaps(metrics: list[dict[str, Any]]) -> list[str]:
def sql_check_gaps(metrics: list[dict[str, Any]], cited: dict[str, list[str]] | None = None, *,
label: Callable[[str, str | None], str | None] | None = None) -> list[str]:
"""sql_check_findings 的那几句缺口(见下)。"""
return [f["text"] for f in sql_check_findings(metrics, cited, label=label)]
def sql_check_findings(metrics: list[dict[str, Any]], cited: dict[str, list[str]] | None = None, *,
label: Callable[[str, str | None], str | None] | None = None) -> list[dict[str, Any]]:
"""没通过 SQL 检查(error 级:一对多关联后重复计算、存量跨期加总)的查询,每条一句缺口。
两处来历:契约收来的指标(口径卡标了 sql_check_failed,sql_check_sources 记着取数的查询快照),以及报告直接
@@ -475,9 +481,12 @@ def sql_check_gaps(metrics: list[dict[str, Any]], cited: dict[str, list[str]] |
按出问题的那条查询归并(合并查询顺着输入找到源查询):同一条查询算出三个指标、报告又引用了它的格,也只记一句,
句子里列出受影响的引用。label(快照 id, 节点 id) 给出「查询「取数」(Q1)」这样的称呼,给不出就写「所依据的查询」。
指标标了 sql_check_failed、快照却读不出来(被清理、被改过)的,退回口径卡记下的那句:缺口不能因为读不到快照就消失。
每条一个 dict:text(缺口那句)、query(称呼)、artifact、node_id、problems、refs。出具声明里另存一份
sql_checks:报告页据此在显眼处说「因为 SQL 检查没通过而降档」,不用从缺口的中文里认。
"""
groups: dict[str, dict[str, Any]] = {}
fallback: list[str] = []
fallback: list[dict[str, Any]] = []
def add(artifact: Any, ref: str) -> bool:
found = query_problems(artifact) if isinstance(artifact, str) and artifact else []
@@ -495,7 +504,8 @@ def sql_check_gaps(metrics: list[dict[str, Any]], cited: dict[str, list[str]] |
name = f"指标「{m.get('name') or m.get('id')}」"
hits = [add(src.get("artifact"), name) for src in m.get("sql_check_sources") or [] if isinstance(src, dict)]
if not any(hits):
fallback.append(f"{name}{m.get('sql_check_reason') or '所依据的查询未通过 SQL 检查'}")
fallback.append({"text": f"{name}{m.get('sql_check_reason') or '所依据的查询未通过 SQL 检查'}",
"query": None, "artifact": None, "node_id": None, "problems": [], "refs": [name]})
for artifact, refs in (cited or {}).items():
for ref in refs:
add(artifact, ref)
@@ -503,10 +513,13 @@ def sql_check_gaps(metrics: list[dict[str, Any]], cited: dict[str, list[str]] |
out = []
for artifact, group in groups.items():
problem = group["problem"]
who = (label(artifact, problem.node_id) if label else None) or "所依据的查询"
named = label(artifact, problem.node_id) if label else None
who = named or "所依据的查询"
refs = group["refs"]
listed = "、".join(refs[:_SQL_REFS]) + (f"等 {len(refs)} 处" if len(refs) > _SQL_REFS else "")
out.append(f"{who}{problem_clause(problem.problems)};受影响的引用:{listed}")
out.append({"text": f"{who}{problem_clause(problem.problems)};受影响的引用:{listed}", "query": named,
"artifact": artifact, "node_id": problem.node_id, "problems": problem.problems,
"refs": refs[:_SQL_REFS * 2]})
return out + fallback
+19 -4
View File
@@ -147,7 +147,7 @@ def _apply_contract(
) -> dict[str, Any]:
from datetime import datetime, timezone
from app.engine.issuance import decide_tier, incomplete_gaps, sql_check_gaps, trace_numbers
from app.engine.issuance import decide_tier, incomplete_gaps, sql_check_findings, trace_numbers
nodes = state.get("nodes") or {}
# 写了 report_from 就是引用模式:按报告文档逐段复核,不再按数值回指叙述
@@ -236,7 +236,17 @@ def _apply_contract(
return f"查询「{title}」({alias})"
return f"查询「{title}」" if title else (f"查询 {alias}" if alias else "")
gaps[sql_at:sql_at] = sql_check_gaps(metrics, cited.get("queries"), label=query_label)
findings = sql_check_findings(metrics, cited.get("queries"), label=query_label)
gaps[sql_at:sql_at] = [f["text"] for f in findings]
# 出具声明另存一份结构化的:报告页在显眼处说「因为 SQL 检查没通过而降档」,不从缺口的中文里认。没有就不加这个键
from app.engine.sql_problems import node_of
for f in findings:
if not f.get("node_id") and f.get("artifact"):
f["node_id"] = node_of(f["artifact"], state.get("evidence"))
# gap 是 gaps 里对应的那一句:界面单列这几条时,据它不在「其余缺口」里重复
sql_checks = [{**{k: f[k] for k in ("query", "node_id", "problems", "refs") if f.get(k)}, "gap": f["text"]}
for f in findings]
# 声明了 metrics_from 却一个指标都没收到,同样是"没查成"
if sources and not metrics:
@@ -265,7 +275,7 @@ def _apply_contract(
return _declare_citations(tier, cited, calibers=calibers, metrics=metrics, gaps=gaps,
missing_required=missing_required, missing_expected=missing_expected,
report_from=str(report_from), declared_at=datetime.now(timezone.utc),
ctx=ctx)
ctx=ctx, sql_checks=sql_checks)
# 逐个数字的出处。以前只有一个计数,界面标不出「这个数来自哪个指标」
matched = [
@@ -284,6 +294,7 @@ def _apply_contract(
# 校验没跑全的原因照实印出来——读的人要能分辨"查过都对"和"根本没查"
"gaps": gaps,
"declared_at": datetime.now(timezone.utc).isoformat(),
**({"sql_checks": sql_checks} if sql_checks else {}),
}
ctx.emit(
EventType.ISSUANCE,
@@ -298,6 +309,7 @@ def _apply_contract(
metrics_checked=len(metrics),
matched_numbers=len(trace.matched) if trace else 0,
matched=matched,
**({"sql_checks": sql_checks} if sql_checks else {}),
)
return issuance
@@ -584,7 +596,7 @@ def _cited_queries(doc: dict[str, Any], catalog: dict[str, Any]) -> tuple[dict[s
def _declare_citations(
tier: str, cited: dict[str, Any], *, calibers: list[dict[str, str]], metrics: list[dict[str, Any]],
gaps: list[str], missing_required: list[str], missing_expected: list[str], report_from: str,
declared_at: Any, ctx: NodeContext,
declared_at: Any, ctx: NodeContext, sql_checks: list[dict[str, Any]] | None = None,
) -> dict[str, Any]:
"""引用模式的 _issuance。旧字段一个不少(老前端照常显示),另加 mode、report、unresolved。"""
report = {"node_id": report_from, "doc_artifact": cited["doc_artifact"]}
@@ -606,6 +618,8 @@ def _declare_citations(
# 结论句策略和可疑实体:只在用上时才有这两个键,升级前的运行和没用上的契约形状不变
**({"claims": cited["claims"]} if cited.get("claims") else {}),
**({"entities": cited["entities"]} if cited.get("entities") else {}),
# 没通过 SQL 检查的查询(结构化):报告页据此在显眼处说明降档原因。没有就不加这个键
**({"sql_checks": sql_checks} if sql_checks else {}),
}
entities = cited.get("entities")
ctx.emit(
@@ -619,6 +633,7 @@ def _declare_citations(
unresolved=len(cited["unresolved"]),
calibers=calibers,
gaps=gaps,
**({"sql_checks": sql_checks} if sql_checks else {}),
metrics_checked=len(metrics),
matched_numbers=len(cited["matched"]),
matched=cited["matched"][:50],
@@ -97,6 +97,12 @@ async def test_one_failing_query_is_one_gap_and_one_timeline_row(scenic_source,
[gap] = [g for g in issuance["gaps"] if "SQL 检查" in g]
assert gap.count(PROBLEM) == 1
assert "指标「订单金额」" in gap and "指标「订单金额(整形后)」" in gap and "Q1 第 1 行「gmv」" in gap
# 出具声明和事件另带结构化的 sql_checks:报告页在显眼处说明降档原因,不从缺口的中文里认(B5)
[found] = issuance["sql_checks"]
assert found == {"query": "查询「fetch」(Q1)", "node_id": "fetch", "problems": [PROBLEM],
"refs": ["指标「订单金额」", "指标「订单金额(整形后)」", "Q1 第 1 行「gmv」"], "gap": gap}, found
[declared] = [e.data for e in await events(row.id, "issuance")]
assert declared["sql_checks"] == issuance["sql_checks"]
# 运行时间线:同一条查询的问题一行,列出受影响的指标;带着结构化字段,界面不用解析中文
[log] = [e.data for e in await events(row.id, "log", "card") if e.data.get("code") == "metric_sql_check"]