xiaobai-review/xiaobai-datahub/datahub/pipeline.py

from __future__ import annotations

import json
import time
from collections.abc import Callable
from typing import Any

from datahub.adapters.base import AdapterError
from datahub.adapters.tushare import DEFAULT_INDEX_CODES, TushareAdapter
from datahub.db import DATASET_TABLES, HubDB
from datahub.governance.circuit import CircuitBreaker
from datahub.governance.ratelimit import TokenBucket
from datahub.governance.retry import RetryError, retry_call
from datahub.logutil import get_logger
from datahub.normalize import finite_number, normalize_daily
from datahub.settings import Settings
from datahub.timeutil import add_days, isoformat, now_shanghai, yyyymmdd

LOGGER = get_logger()

HARD_DATASETS = {"daily", "valuation", "index_daily"}
SOFT_DATASETS = {"moneyflow", "auction"}

STAGING_INSERT = {
    "daily": (
        "INSERT INTO staging_bars(ts_code,trade_date,batch_id,open,high,low,close,pct_chg,volume,amount,adj_factor) "
        "VALUES (?,?,?,?,?,?,?,?,?,?,?)",
        lambda r, b: (
            r["ts_code"], r["trade_date"], b, r.get("open"), r.get("high"), r.get("low"),
            r.get("close"), r.get("pct_chg"), r.get("volume"), r.get("amount"), r.get("adj_factor"),
        ),
    ),
    "valuation": (
        "INSERT INTO staging_valuation(ts_code,trade_date,batch_id,turnover_rate,volume_ratio,total_mv,circ_mv,pe_ttm,pb,ps_ttm,dv_ttm) "
        "VALUES (?,?,?,?,?,?,?,?,?,?,?)",
        lambda r, b: (
            r["ts_code"], r["trade_date"], b, r.get("turnover_rate"), r.get("volume_ratio"),
            r.get("total_mv"), r.get("circ_mv"), r.get("pe_ttm"), r.get("pb"), r.get("ps_ttm"), r.get("dv_ttm"),
        ),
    ),
    "moneyflow": (
        "INSERT INTO staging_moneyflow(ts_code,trade_date,batch_id,buy_sm_amount,sell_sm_amount,buy_md_amount,sell_md_amount,buy_lg_amount,sell_lg_amount,buy_elg_amount,sell_elg_amount,net_mf_amount) "
        "VALUES (?,?,?,?,?,?,?,?,?,?,?,?)",
        lambda r, b: (
            r["ts_code"], r["trade_date"], b,
            r.get("buy_sm_amount"), r.get("sell_sm_amount"), r.get("buy_md_amount"), r.get("sell_md_amount"),
            r.get("buy_lg_amount"), r.get("sell_lg_amount"), r.get("buy_elg_amount"), r.get("sell_elg_amount"),
            r.get("net_mf_amount"),
        ),
    ),
    "auction": (
        "INSERT INTO staging_auction(ts_code,trade_date,batch_id,volume,price,amount,pre_close,turnover_rate,volume_ratio,float_share) "
        "VALUES (?,?,?,?,?,?,?,?,?,?)",
        lambda r, b: (
            r["ts_code"], r["trade_date"], b, r.get("volume"), r.get("price"), r.get("amount"),
            r.get("pre_close"), r.get("turnover_rate"), r.get("volume_ratio"), r.get("float_share"),
        ),
    ),
    "index_daily": (
        "INSERT INTO staging_index_bars(ts_code,trade_date,batch_id,open,high,low,close,pct_chg,volume,amount) "
        "VALUES (?,?,?,?,?,?,?,?,?,?)",
        lambda r, b: (
            r["ts_code"], r["trade_date"], b, r.get("open"), r.get("high"), r.get("low"),
            r.get("close"), r.get("pct_chg"), r.get("volume"), r.get("amount"),
        ),
    ),
}

EOD_COPY = {
    "daily": (
        "INSERT OR REPLACE INTO eod_bars "
        "SELECT ts_code,trade_date,open,high,low,close,pct_chg,volume,amount,adj_factor,batch_id "
        "FROM staging_bars WHERE batch_id = ?"
    ),
    "valuation": (
        "INSERT OR REPLACE INTO eod_valuation "
        "SELECT ts_code,trade_date,turnover_rate,volume_ratio,total_mv,circ_mv,pe_ttm,pb,ps_ttm,dv_ttm,batch_id "
        "FROM staging_valuation WHERE batch_id = ?"
    ),
    "moneyflow": (
        "INSERT OR REPLACE INTO eod_moneyflow "
        "SELECT ts_code,trade_date,buy_sm_amount,sell_sm_amount,buy_md_amount,sell_md_amount,"
        "buy_lg_amount,sell_lg_amount,buy_elg_amount,sell_elg_amount,net_mf_amount,batch_id "
        "FROM staging_moneyflow WHERE batch_id = ?"
    ),
    "auction": (
        "INSERT OR REPLACE INTO eod_auction "
        "SELECT ts_code,trade_date,volume,price,amount,pre_close,turnover_rate,volume_ratio,float_share,batch_id "
        "FROM staging_auction WHERE batch_id = ?"
    ),
    "index_daily": (
        "INSERT OR REPLACE INTO eod_index_bars "
        "SELECT ts_code,trade_date,open,high,low,close,pct_chg,volume,amount,batch_id "
        "FROM staging_index_bars WHERE batch_id = ?"
    ),
}


class QualityError(RuntimeError):
    def __init__(self, message: str, report: dict[str, Any]) -> None:
        super().__init__(message)
        self.report = report


class Pipeline:
    def __init__(
        self,
        db: HubDB,
        adapter: TushareAdapter,
        settings: Settings,
        bucket: TokenBucket | None = None,
        breaker: CircuitBreaker | None = None,
        before_commit: Callable[[], None] | None = None,
        clock=None,
    ) -> None:
        self.db = db
        self.adapter = adapter
        self.settings = settings
        self.bucket = bucket or TokenBucket(settings.tushare_rate_per_minute)
        self.breaker = breaker or CircuitBreaker()
        self.before_commit = before_commit
        self.clock = clock or now_shanghai

    def next_batch_id(self, dataset: str, trade_date: str) -> str:
        row = self.db.fetchone(
            "SELECT COUNT(*) AS n FROM batches WHERE dataset = ? AND trade_date = ?",
            (dataset, trade_date),
        )
        seq = int((row or {}).get("n") or 0) + 1
        return f"{trade_date}-{dataset}-{seq:03d}"

    def ingest_reference(self, trade_date: str | None = None) -> dict[str, Any]:
        """Refresh trade calendar (window) and stock master. Not versioned by batch."""
        day = yyyymmdd(trade_date or self.clock())
        start = add_days(day, -400)
        end = add_days(day, 30)
        calendar = self.adapter.normalize(
            "calendar",
            self._guarded_fetch("calendar", {"exchange": "SSE", "start_date": start, "end_date": end}),
        )
        stocks = self.adapter.normalize("stocks", self._guarded_fetch("stocks", {"list_status": "L"}))
        fetched_at = isoformat(self.clock())
        with self.db.write() as connection:
            for row in calendar:
                connection.execute(
                    """
                    INSERT INTO trade_calendar(exchange, cal_date, is_open, pretrade_date, fetched_at)
                    VALUES (?, ?, ?, ?, ?)
                    ON CONFLICT(exchange, cal_date) DO UPDATE SET
                      is_open=excluded.is_open, pretrade_date=excluded.pretrade_date, fetched_at=excluded.fetched_at
                    """,
                    (row["exchange"], row["cal_date"], row["is_open"], row.get("pretrade_date"), fetched_at),
                )
            for row in stocks:
                connection.execute(
                    """
                    INSERT INTO stock_master(ts_code,symbol,name,area,industry,market,list_status,list_date,updated_at)
                    VALUES (?,?,?,?,?,?,?,?,?)
                    ON CONFLICT(ts_code) DO UPDATE SET
                      symbol=excluded.symbol, name=excluded.name, area=excluded.area,
                      industry=excluded.industry, market=excluded.market,
                      list_status=excluded.list_status, list_date=excluded.list_date,
                      updated_at=excluded.updated_at
                    """,
                    (
                        row["ts_code"], row.get("symbol"), row.get("name"), row.get("area"),
                        row.get("industry"), row.get("market"), row.get("list_status"),
                        row.get("list_date"), fetched_at,
                    ),
                )
        return {"calendar": len(calendar), "stocks": len(stocks), "trade_date": day}

    def run_dataset(self, dataset: str, trade_date: str, attempts: int | None = None) -> dict[str, Any]:
        trade_date = yyyymmdd(trade_date)
        batch_id = self.next_batch_id(dataset, trade_date)
        max_attempts = attempts or self.settings.max_publish_attempts
        self._set_batch(batch_id, dataset, trade_date, "scheduled", 0)
        try:
            self._set_batch(batch_id, dataset, trade_date, "fetching", 1)
            rows = retry_call(
                lambda: self._fetch_dataset(dataset, trade_date),
                attempts=max_attempts,
                base_delay=0.05,
                sleeper=lambda _d: None if attempts == 1 else time.sleep(_d),
            )
            self._stage(dataset, batch_id, rows)
            self._set_batch(batch_id, dataset, trade_date, "staged", 1, rows_in=len(rows), rows_out=len(rows))
            self._set_batch(batch_id, dataset, trade_date, "validating", 1)
            report = self.validate(dataset, batch_id, trade_date, rows)
            if report["hard_fail"]:
                self._set_batch(
                    batch_id, dataset, trade_date, "staged", 1,
                    rows_in=len(rows), rows_out=len(rows),
                    quality=report, error="; ".join(report["errors"]),
                )
                raise QualityError("integrity gate failed", report)
            self._set_batch(batch_id, dataset, trade_date, "deriving", 1, rows_in=len(rows), rows_out=len(rows), quality=report)
            self._set_batch(batch_id, dataset, trade_date, "publishing", 1, rows_in=len(rows), rows_out=len(rows), quality=report)
            state = "degraded" if report["soft_fail"] else "published"
            self.publish(dataset, trade_date, batch_id, state=state)
            self._set_batch(
                batch_id, dataset, trade_date, "published", 1,
                rows_in=len(rows), rows_out=len(rows), quality=report, finished=True,
            )
            return {"batch_id": batch_id, "dataset": dataset, "trade_date": trade_date, "rows": len(rows), "state": state, "quality": report}
        except RetryError as exc:
            self._set_batch(batch_id, dataset, trade_date, "failed", max_attempts, error=str(exc), finished=True)
            raise
        except QualityError:
            raise
        except Exception as exc:
            self._set_batch(batch_id, dataset, trade_date, "failed", 1, error=str(exc), finished=True)
            raise

    def run_eod_batch_a(self, trade_date: str) -> dict[str, Any]:
        results = {}
        for dataset in ("daily", "valuation", "moneyflow", "auction"):
            results[dataset] = self.run_dataset(dataset, trade_date)
        return results

    def run_eod_batch_b(self, trade_date: str) -> dict[str, Any]:
        return {"index_daily": self.run_dataset("index_daily", trade_date)}

    def validate(self, dataset: str, batch_id: str, trade_date: str, rows: list[dict[str, Any]]) -> dict[str, Any]:
        quality = self.settings.quality
        errors: list[str] = []
        warnings: list[str] = []
        listed = self.db.fetchone(
            "SELECT COUNT(*) AS n FROM stock_master WHERE list_status = 'L'",
        )
        listed_n = int((listed or {}).get("n") or 0)
        row_n = len(rows)
        keys = [(row.get("ts_code"), row.get("trade_date")) for row in rows]
        dup = row_n - len(set(keys))
        if dup:
            errors.append(f"duplicate keys: {dup}")
        bad_date = sum(1 for row in rows if str(row.get("trade_date")) != trade_date)
        if bad_date:
            errors.append(f"date mismatch rows: {bad_date}")
        ratio = (row_n / listed_n) if listed_n else 1.0
        if dataset == "daily" and listed_n and ratio < float(quality.get("daily_row_ratio") or 0.98):
            errors.append(f"row ratio {ratio:.4f} < {quality.get('daily_row_ratio')}")
        null_fields = ("open", "high", "low", "close", "amount") if dataset in {"daily", "index_daily"} else ()
        if null_fields and rows:
            nulls = sum(1 for row in rows if any(row.get(field) is None for field in null_fields))
            null_rate = nulls / row_n
            if null_rate >= float(quality.get("null_rate_max") or 0.01):
                errors.append(f"null rate {null_rate:.4f}")
        if dataset in SOFT_DATASETS and row_n == 0:
            warnings.append("empty soft dataset")
        hard_fail = bool(errors) and dataset in HARD_DATASETS.union({"daily", "valuation", "index_daily"})
        if dataset in SOFT_DATASETS:
            hard_fail = bool(dup or bad_date)
        return {
            "rows": row_n,
            "listed": listed_n,
            "ratio": round(ratio, 4),
            "errors": errors,
            "warnings": warnings,
            "hard_fail": hard_fail,
            "soft_fail": bool(warnings) and not hard_fail,
            "batch_id": batch_id,
        }

    def publish(self, dataset: str, trade_date: str, batch_id: str, state: str = "published") -> None:
        copy_sql = EOD_COPY[dataset]
        published_at = isoformat(self.clock())
        with self.db.write() as connection:
            current = connection.execute(
                "SELECT active_batch FROM publications WHERE dataset = ? AND trade_date = ?",
                (dataset, trade_date),
            ).fetchone()
            prev = str(current["active_batch"]) if current else None
            connection.execute(copy_sql, (batch_id,))
            if self.before_commit:
                self.before_commit()
            connection.execute(
                """
                INSERT INTO publications(dataset, trade_date, active_batch, prev_batch, state, published_at)
                VALUES (?, ?, ?, ?, ?, ?)
                ON CONFLICT(dataset, trade_date) DO UPDATE SET
                  prev_batch=excluded.prev_batch,
                  active_batch=excluded.active_batch,
                  state=excluded.state,
                  published_at=excluded.published_at
                """,
                (dataset, trade_date, batch_id, prev, state, published_at),
            )
            max_gen = connection.execute(
                "SELECT COALESCE(MAX(generation), 0) AS g FROM publication_history WHERE dataset = ? AND trade_date = ?",
                (dataset, trade_date),
            ).fetchone()
            generation = int(max_gen["g"]) + 1
            connection.execute(
                "INSERT OR REPLACE INTO publication_history(dataset, trade_date, batch_id, published_at, generation) VALUES (?,?,?,?,?)",
                (dataset, trade_date, batch_id, published_at, generation),
            )
            keep = int(self.settings.quality.get("publication_generations") or 3)
            stale = connection.execute(
                """
                SELECT batch_id FROM publication_history
                WHERE dataset = ? AND trade_date = ?
                ORDER BY generation DESC
                """,
                (dataset, trade_date),
            ).fetchall()
            for row in stale[keep:]:
                connection.execute(
                    "DELETE FROM publication_history WHERE dataset = ? AND trade_date = ? AND batch_id = ?",
                    (dataset, trade_date, row["batch_id"]),
                )

    def rollback(self, dataset: str, trade_date: str, actor: str = "admin") -> dict[str, Any]:
        trade_date = yyyymmdd(trade_date)
        pub = self.db.fetchone(
            "SELECT * FROM publications WHERE dataset = ? AND trade_date = ?",
            (dataset, trade_date),
        )
        if not pub or not pub.get("prev_batch"):
            raise ValueError("没有可回滚的上一批次")
        target = pub["prev_batch"]
        published_at = isoformat(self.clock())
        with self.db.write() as connection:
            connection.execute(
                """
                UPDATE publications
                SET prev_batch = active_batch, active_batch = ?, published_at = ?, state = 'published'
                WHERE dataset = ? AND trade_date = ?
                """,
                (target, published_at, dataset, trade_date),
            )
        self.audit(actor, "rollback", f"{dataset}:{trade_date}", json.dumps({"to": target, "from": pub["active_batch"]}))
        return {"dataset": dataset, "trade_date": trade_date, "active_batch": target, "prev_batch": pub["active_batch"]}

    def active_batch(self, dataset: str, trade_date: str) -> str | None:
        row = self.db.fetchone(
            "SELECT active_batch FROM publications WHERE dataset = ? AND trade_date = ?",
            (dataset, trade_date),
        )
        return str(row["active_batch"]) if row else None

    def cleanup(self) -> dict[str, int]:
        staging_days = int(self.settings.quality.get("staging_retain_days") or 14)
        job_days = int(self.settings.quality.get("job_run_retain_days") or 90)
        cutoff_staging = add_days(yyyymmdd(self.clock()), -staging_days)
        cutoff_jobs = add_days(yyyymmdd(self.clock()), -job_days)
        deleted = 0
        with self.db.write() as connection:
            for dataset, (_eod, staging) in DATASET_TABLES.items():
                cur = connection.execute(
                    f"DELETE FROM {staging} WHERE trade_date < ?",
                    (cutoff_staging,),
                )
                deleted += cur.rowcount
            connection.execute("DELETE FROM job_runs WHERE started_at < ?", (cutoff_jobs,))
            connection.execute("DELETE FROM src_calls WHERE created_at < ?", (cutoff_jobs,))
        return {"staging_deleted": deleted}

    def audit(self, actor: str, action: str, target: str = "", detail: str = "") -> None:
        self.db.execute(
            "INSERT INTO audit_log(actor, action, target, detail, created_at) VALUES (?,?,?,?,?)",
            (actor, action, target, detail, isoformat(self.clock())),
        )

    def _fetch_dataset(self, dataset: str, trade_date: str) -> list[dict[str, Any]]:
        if dataset == "daily":
            raw = self._guarded_fetch("daily", {"trade_date": trade_date})
            factors = {
                (row["ts_code"], row["trade_date"]): finite_number(row.get("adj_factor"))
                for row in self._guarded_fetch("adj_factor", {"trade_date": trade_date})
            }
            return [
                normalize_daily(row, adj_factor=factors.get((str(row.get("ts_code") or "").upper(), str(row.get("trade_date") or ""))))
                for row in raw
            ]
        if dataset == "index_daily":
            rows: list[dict[str, Any]] = []
            for ts_code in DEFAULT_INDEX_CODES:
                raw = self._guarded_fetch("index_daily", {"ts_code": ts_code, "trade_date": trade_date})
                rows.extend(self.adapter.normalize("index_daily", raw))
            return rows
        api_dataset = dataset
        raw = self._guarded_fetch(api_dataset, {"trade_date": trade_date})
        return self.adapter.normalize(api_dataset, raw)

    def _guarded_fetch(self, dataset: str, params: dict[str, Any]) -> list[dict[str, Any]]:
        if not self.breaker.allow():
            raise AdapterError("Tushare circuit open")
        self.bucket.acquire()
        started = time.perf_counter()
        try:
            # For daily we want RAW tushare rows so adj_factor can be merged later.
            rows = self.adapter.fetch(dataset, params)
            latency = round((time.perf_counter() - started) * 1000)
            self.breaker.record_success()
            self._log_call(dataset, True, latency, "")
            self._persist_health("ok")
            return rows
        except Exception as exc:
            latency = round((time.perf_counter() - started) * 1000)
            self.breaker.record_failure(str(exc))
            self._log_call(dataset, False, latency, str(exc))
            self._persist_health("error", str(exc))
            raise

    def _stage(self, dataset: str, batch_id: str, rows: list[dict[str, Any]]) -> None:
        sql, mapper = STAGING_INSERT[dataset]
        with self.db.write() as connection:
            connection.execute(
                f"DELETE FROM {DATASET_TABLES[dataset][1]} WHERE batch_id = ?",
                (batch_id,),
            )
            connection.executemany(sql, [mapper(row, batch_id) for row in rows])

    def _set_batch(
        self,
        batch_id: str,
        dataset: str,
        trade_date: str,
        state: str,
        attempt: int,
        rows_in: int | None = None,
        rows_out: int | None = None,
        quality: dict[str, Any] | None = None,
        error: str | None = None,
        finished: bool = False,
    ) -> None:
        now = isoformat(self.clock())
        existing = self.db.fetchone("SELECT batch_id FROM batches WHERE batch_id = ?", (batch_id,))
        payload = json.dumps(quality, ensure_ascii=False) if quality else None
        with self.db.write() as connection:
            if existing is None:
                connection.execute(
                    """
                    INSERT INTO batches(batch_id, dataset, trade_date, state, attempt, rows_in, rows_out, quality_json, started_at, finished_at, error)
                    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
                    """,
                    (batch_id, dataset, trade_date, state, attempt, rows_in, rows_out, payload, now, now if finished else None, error),
                )
            else:
                connection.execute(
                    """
                    UPDATE batches SET state=?, attempt=?,
                      rows_in=COALESCE(?, rows_in), rows_out=COALESCE(?, rows_out),
                      quality_json=COALESCE(?, quality_json),
                      finished_at=CASE WHEN ? THEN ? ELSE finished_at END,
                      error=COALESCE(?, error)
                    WHERE batch_id = ?
                    """,
                    (state, attempt, rows_in, rows_out, payload, 1 if finished else 0, now, error, batch_id),
                )

    def _log_call(self, endpoint: str, ok: bool, latency_ms: int, error: str) -> None:
        self.db.execute(
            "INSERT INTO src_calls(provider, endpoint, ok, latency_ms, error, created_at) VALUES (?,?,?,?,?,?)",
            ("tushare", endpoint, 1 if ok else 0, latency_ms, error, isoformat(self.clock())),
        )

    def _persist_health(self, state: str, error: str = "") -> None:
        snap = self.breaker.snapshot()
        self.db.execute(
            """
            INSERT INTO src_health(provider, endpoint_class, state, last_ok_at, last_error, consec_failures, opened_at, cooldown_until)
            VALUES (?, ?, ?, ?, ?, ?, ?, ?)
            ON CONFLICT(provider, endpoint_class) DO UPDATE SET
              state=excluded.state, last_ok_at=excluded.last_ok_at, last_error=excluded.last_error,
              consec_failures=excluded.consec_failures, opened_at=excluded.opened_at, cooldown_until=excluded.cooldown_until
            """,
            (
                "tushare", "pro",
                snap.state,
                isoformat(self.clock()) if state == "ok" else None,
                error or snap.last_error,
                snap.consec_failures,
                isoformat(self.clock()) if snap.state == "open" else None,
                None,
            ),
        )