astrololo/services/data/app/providers/excel_provider.py

"""ExcelDataProvider — wyszukiwanie w setkach plików .xlsx z 4-poziomowym cache.

Ścieżka zapytania (od najszybszej):
  1) QueryCache (in-memory)  -> gotowy wynik
  2) InvertedIndex (SQLite)  -> które pliki w ogóle otwierać (zamiast skanu setek)
  3) FrameCache (Parquet)    -> wczytanie pliku bez parsowania .xlsx
  4) SchemaCache (SQLite)    -> bez ponownego wykrywania nagłówka/układu kolumn
  ...dopiero gdy wszystko spudłuje, czytamy .xlsx i wypełniamy cache.

Cała ta złożoność jest UKRYTA za interfejsem DataProvider.
"""
from __future__ import annotations

import time
from pathlib import Path

import pandas as pd

from app.cache.fingerprint import fingerprint
from app.cache.frame_cache import FrameCache
from app.cache.index import InvertedIndex
from app.cache.query_cache import QueryCache
from app.cache.schema_cache import SchemaCache
from app.config import Settings
from app.excel.header_detect import detect_header_row
from app.excel.layout import build_column_mapping
from app.models import HealthInfo, SearchQuery, SearchResult
from app.providers.base import DataProvider


class ExcelDataProvider(DataProvider):
    name = "excel"

    def __init__(self, settings: Settings) -> None:
        self.s = settings
        self.schema = SchemaCache(settings.cache_dir)
        self.frames = FrameCache(settings.cache_dir)
        self.index = InvertedIndex(settings.cache_dir)
        self.queries = QueryCache(settings.query_cache_size, settings.query_cache_ttl)

    # ---- ładowanie pojedynczego arkusza z pełnym cache ----
    def _load_frame(self, path: str, sheet: str | int = 0) -> pd.DataFrame:
        fp = fingerprint(path)
        sheet_key = str(sheet)

        cached = self.frames.get(fp, sheet_key)        # poziom 2: Parquet
        if cached is not None:
            return cached

        raw = pd.read_excel(path, sheet_name=sheet, header=None, dtype=object)
        meta = self.schema.get(fp, sheet_key)          # poziom 1: schemat
        if meta is None:
            header_row = detect_header_row(raw, self.s.header_scan_rows)
            header_cells = [str(c) for c in raw.iloc[header_row].tolist()]
            mapping = build_column_mapping(header_cells)
            self.schema.put(fp, sheet_key, header_row, mapping)
        else:
            header_row, mapping = meta
            header_cells = [str(c) for c in raw.iloc[header_row].tolist()]

        data = raw.iloc[header_row + 1 :].copy()
        data.columns = header_cells
        data = data.dropna(how="all")
        inverse = {orig: canon for canon, orig in mapping.items()}
        data = data.rename(columns=inverse).reset_index(drop=True)

        self.frames.put(fp, sheet_key, data)           # zapisz Parquet na przyszłość
        return data

    # ---- budowa odwróconego indeksu (warmup / po zmianie pliku) ----
    def _ensure_indexed(self, path: str) -> None:
        fp = fingerprint(path)
        if self.index.file_fingerprint(path) == fp:
            return  # aktualny
        frame = self._load_frame(path)
        rows: list[tuple[str, str, str]] = []
        for key in self.s.indexed_keys:
            if key in frame.columns:
                for v in frame[key].dropna().astype(str).unique():
                    rows.append((key, v, "0"))
        self.index.reindex_file(path, fp, rows)

    def warmup(self) -> None:
        for path in self._excel_files():
            try:
                self._ensure_indexed(path)
            except Exception as e:  # jeden uszkodzony plik nie może zablokować startu
                print(f"[data] pominięto plik przy indeksowaniu: {path} — {e}")

    def _excel_files(self) -> list[str]:
        base = Path(self.s.excel_dir)
        return [str(p) for p in sorted(base.glob("**/*.xlsx")) if not p.name.startswith("~$")]

    # ---- publiczne API ----
    def search(self, query: SearchQuery) -> SearchResult:
        t0 = time.perf_counter()
        cache_key = f"{query.key}|{query.value}|{query.exact}|{query.limit}|{query.fields}"

        hit = self.queries.get(cache_key)              # poziom 3: wynik zapytania
        if hit is not None:
            hit = hit.model_copy(update={"cache": "hit", "elapsed_ms": _ms(t0)})
            return hit

        candidates = self.index.lookup(query.key, query.value, query.exact)
        if not candidates:
            # brak w indeksie (np. klucz nieindeksowany) -> przeszukaj wszystkie pliki
            candidates = [(p, "0") for p in self._excel_files()]

        rows: list[dict] = []
        for path, _sheet in candidates:
            frame = self._load_frame(path)
            if query.key not in frame.columns:
                continue
            col = frame[query.key].astype(str)
            if query.exact:
                mask = col.str.lower() == query.value.lower()
            else:
                # regex=False: wartości sygnifikatorów zawierają znaki [ + itd.,
                # które są metaznakami regex — szukamy dosłownie.
                mask = col.str.lower().str.contains(query.value.lower(), na=False, regex=False)
            matched = frame[mask]
            if query.fields:
                keep = [c for c in query.fields if c in matched.columns]
                matched = matched[keep]
            rows.extend(matched.to_dict(orient="records"))
            if len(rows) >= query.limit:
                break

        result = SearchResult(
            rows=rows[: query.limit],
            total=len(rows),
            elapsed_ms=_ms(t0),
            cache="miss",
            provider=self.name,
        )
        self.queries.put(cache_key, result)
        return result

    def health(self) -> HealthInfo:
        return HealthInfo(
            provider=self.name,
            indexed_files=self.index.count_files(),
            details={"excel_dir": str(self.s.excel_dir), "files_on_disk": len(self._excel_files())},
        )


def _ms(t0: float) -> float:
    return round((time.perf_counter() - t0) * 1000, 2)