"""Read side: DuckDB over the live Parquet file set selected from the Postgres manifest (`lake_files`).

Adjustments are applied at query time from `corporate_actions` (explicit, reproducible):
  split      : multiply prices by Π(1/ratio) for splits with ex_date > bar session_date; divide volume by the same factor
  split_div  : additionally multiply by Π(1 - div/prev_close) for dividends with ex_date > session_date
               (prev_close = last raw close strictly before ex_date in the requested data — if unavailable the
               dividend factor is skipped and reported in `adjustment_notes`).
"""
from __future__ import annotations

import re
from datetime import date
from typing import Any

import duckdb
import pyarrow as pa
from findata_core.db import connection
from findata_schema import BARS_SCHEMA

from findata_ingest.lake import Lake, live_files


def _con(read_only: bool = True) -> duckdb.DuckDBPyConnection:
    import resource
    try:  # scans over many Parquet files need file descriptors (macOS default soft limit is 256)
        soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
        if soft < 65536:
            resource.setrlimit(resource.RLIMIT_NOFILE, (min(65536, hard), hard))
    except (ValueError, OSError):
        pass
    con = duckdb.connect(database=":memory:", config={"threads": 4, "memory_limit": "4GB", "TimeZone": "UTC"})
    return con


def _arrow(con: duckdb.DuckDBPyConnection, sql: str, params: dict[str, Any] | None) -> pa.Table:
    try:
        rel = con.sql(sql, params=params) if params else con.sql(sql)
        return rel.to_arrow_table()
    except TypeError:  # older duckdb without params= on sql()
        return con.execute(sql, params).fetch_arrow_table()


def empty_bars() -> pa.Table:
    return pa.table({f.name: pa.array([], type=f.type) for f in BARS_SCHEMA}, schema=BARS_SCHEMA)


def read_bars(dataset_id: str, symbols: list[str], start: date | None, end: date | None, version: int | None = None,
              limit: int | None = None, columns: list[str] | None = None, lake: Lake | None = None) -> tuple[pa.Table, dict[str, Any]]:
    lake = lake or Lake()
    with connection() as conn:
        files = live_files(conn, dataset_id, symbols, start, end, version)
        ver = conn.execute("SELECT current_version FROM datasets WHERE id=%s", (dataset_id,)).fetchone()
    meta = {"dataset_id": dataset_id, "version": version if version is not None else (ver["current_version"] if ver else None),
            "files_scanned": len(files)}
    if not files:
        return empty_bars(), meta
    paths = [str(lake.resolve(f["path"])) for f in files]
    con = _con()
    con.register("paths", pa.table({"p": paths}))
    cols = ", ".join(columns) if columns else "*"
    where = ["symbol_as_reported IN (SELECT UNNEST($syms))"]
    params: dict[str, Any] = {"syms": symbols}
    if start:
        where.append("session_date >= $start")
        params["start"] = start
    if end:
        where.append("session_date <= $end")
        params["end"] = end
    sql = f"SELECT {cols} FROM read_parquet($paths, union_by_name=true, hive_partitioning=false) WHERE {' AND '.join(where)} ORDER BY symbol_as_reported, timestamp_utc"
    params["paths"] = paths
    if limit:
        sql += f" LIMIT {int(limit) + 1}"
    tbl = _arrow(con, sql, params)
    meta["truncated"] = bool(limit and tbl.num_rows > limit)
    if limit and tbl.num_rows > limit:
        tbl = tbl.slice(0, limit)
    if not columns:
        tbl = tbl.cast(BARS_SCHEMA)
    return tbl, meta


def apply_adjustment(table: pa.Table, adjustment: str, lake: Lake | None = None) -> tuple[pa.Table, list[str]]:
    """Return an adjusted copy of raw bars (`adjustment_type` column updated). Never mutates the lake."""
    if adjustment in ("raw", None) or table.num_rows == 0:
        return table, []
    import polars as pl
    df = pl.from_arrow(table)
    syms = df.get_column("symbol_as_reported").unique().to_list()
    iids = [f"firstrate:{df.filter(pl.col('symbol_as_reported') == s).get_column('asset_class')[0]}:{s}" for s in syms]
    with connection() as conn:
        acts = conn.execute("SELECT instrument_id, action_type, ex_date, value::float8 AS value FROM corporate_actions WHERE instrument_id = ANY(%s) ORDER BY ex_date",
                            (iids,)).fetchall()
    notes: list[str] = []
    if not acts:
        notes.append("no corporate actions on file for requested symbols")
        return df.with_columns(pl.lit(adjustment).alias("adjustment_type")).to_arrow().cast(BARS_SCHEMA), notes
    out_frames = []
    for s in syms:
        sub = df.filter(pl.col("symbol_as_reported") == s).sort("timestamp_utc")
        iid = f"firstrate:{sub.get_column('asset_class')[0]}:{s}"
        factor = pl.Series("f", [1.0] * sub.height)
        sd = sub.get_column("session_date")
        for a in acts:
            if a["instrument_id"] != iid:
                continue
            if a["action_type"] == "split":
                f = 1.0 / float(a["value"])
                factor = factor * pl.Series([f if d < a["ex_date"] else 1.0 for d in sd])
            elif a["action_type"] == "dividend" and adjustment == "split_div":
                prev = sub.filter(pl.col("session_date") < a["ex_date"]).get_column("close")
                if prev.len() == 0:
                    notes.append(f"{s}: dividend {a['ex_date']} skipped (no prior close in range)")
                    continue
                f = 1.0 - float(a["value"]) / float(prev[-1])
                factor = factor * pl.Series([f if d < a["ex_date"] else 1.0 for d in sd])
        sub = sub.with_columns(factor.alias("_f")).with_columns(
            (pl.col("open") * pl.col("_f")).alias("open"), (pl.col("high") * pl.col("_f")).alias("high"),
            (pl.col("low") * pl.col("_f")).alias("low"), (pl.col("close") * pl.col("_f")).alias("close"),
            (pl.col("volume") / pl.col("_f")).alias("volume"), pl.lit(adjustment).alias("adjustment_type")).drop("_f")
        out_frames.append(sub)
    res = pl.concat(out_frames).sort(["symbol_as_reported", "timestamp_utc"])
    return res.to_arrow().cast(BARS_SCHEMA), notes


def run_sql(sql: str, max_rows: int, timeout_s: int, views: dict[str, Any] | None = None) -> tuple[pa.Table, dict[str, Any]]:
    """Sandboxed read-only SQL: DuckDB in-memory, external access limited to the registered dataset views."""
    import re
    import threading

    forbidden = re.compile(r"\b(attach|copy|export|import|install|load|create|insert|update|delete|drop|alter|pragma|set|call|read_csv|read_json|read_text|glob|httpfs)\b", re.IGNORECASE)
    if forbidden.search(sql) or ";" in sql.strip().rstrip(";"):
        raise ValueError("only a single read-only SELECT statement is allowed")
    if not re.match(r"^\s*(select|with|describe|show|summarize|explain)\b", sql, re.IGNORECASE):
        raise ValueError("statement must start with SELECT / WITH")
    is_explain = bool(re.match(r"^\s*explain\b", sql, re.I))
    con = _con()
    scoped_used: list[str] = []
    dirs = []
    for r in Lake().roots():
        rr = str(r.resolve()).replace("'", "''")
        dirs += [f"'{rr}/curated'", f"'{rr}/_retired'"]
    con.execute(f"SET allowed_directories = [{', '.join(dirs)}]")
    for name, spec in (views or {}).items():
        if not re.fullmatch(r"[a-z][a-z0-9_]*", name):
            raise ValueError(f"bad view name {name}")
        if name not in _referenced_views(sql, {name: spec}):
            continue  # only materialize views the statement uses (directory globs are not free)
        if isinstance(spec, dict) and spec.get("sql"):
            fast = None
            if spec.get("dataset_id"):
                try:
                    fast = scoped_view_sql(spec["dataset_id"], spec.get("kind", ""), extract_scope(sql), Lake())
                except Exception:
                    fast = None
            con.execute(f"CREATE VIEW {name} AS {fast or spec['sql']}")
            scoped_used.append(name if fast else f"{name} (full scan)")
        elif isinstance(spec, list) and spec:  # legacy explicit file list
            lst = ", ".join("'" + g.replace("'", "''") + "'" for g in spec)
            con.execute(f"CREATE VIEW {name} AS SELECT * FROM read_parquet([{lst}], union_by_name=true, hive_partitioning=false)")
        else:
            con.execute(f"CREATE VIEW {name} AS SELECT * FROM (VALUES (NULL)) t(empty) WHERE false")
    con.execute("SET enable_external_access = false")
    timer = threading.Timer(timeout_s, con.interrupt)
    timer.start()
    try:
        body = sql.strip().rstrip(";")
        tbl = _arrow(con, body if is_explain else f"SELECT * FROM ({body}) q LIMIT {int(max_rows) + 1}", None)
    except duckdb.InterruptException as e:
        raise TimeoutError(f"query exceeded {timeout_s}s") from e
    finally:
        timer.cancel()
    truncated = tbl.num_rows > max_rows
    return tbl.slice(0, max_rows), {"truncated": truncated, "views": scoped_used}


_views_cache: dict[str, Any] = {"at": 0.0, "value": None}


def dataset_views() -> dict[str, dict[str, Any]]:
    """SQL views for the playground/assistant: one per dataset with data, built from Hive-style directory globs
    (`curated/<dataset>/{base,delta}/**`) so that DuckDB prunes partitions (`symbol=`, `year=`, `session_date=`,
    `underlying=`, `period=`) from the path instead of receiving millions of file names. The directory tree is the
    live set (retired files are moved out); files of an in-progress publication may be visible for a few minutes.
    Cached 5 minutes."""
    import time
    if _views_cache["value"] is not None and time.monotonic() - _views_cache["at"] < 300:
        return _views_cache["value"]
    lake = Lake()
    out: dict[str, dict[str, Any]] = {}
    with connection() as conn:
        ds = conn.execute("SELECT id, kind FROM datasets WHERE kind IN ('market_bars','derived','options') AND current_version > 0").fetchall()
        for d in ds:
            root = lake.root_for(d["id"])
            layouts = {}
            for layout in ("base", "delta"):
                ldir = root / "curated" / d["id"] / layout
                if ldir.exists() and next(ldir.rglob("*.parquet"), None) is not None:  # skip emptied layouts
                    layouts[layout] = str(ldir / "**" / "*.parquet").replace("'", "''")
            if not layouts:
                continue
            parts = []
            if d["kind"] == "options":
                parts.append(f"SELECT * FROM read_parquet('{layouts['base']}', hive_partitioning=true)")
            else:
                if "base" in layouts:   # partition columns symbol=, year= come from the path (pruned before any file is opened)
                    parts.append(f"SELECT * FROM read_parquet('{layouts['base']}', hive_partitioning=true)")
                if "delta" in layouts:  # session files: derive the same columns so UNION ALL BY NAME lines up
                    parts.append(f"SELECT *, symbol_as_reported AS symbol, year(session_date)::BIGINT AS year "
                                 f"FROM read_parquet('{layouts['delta']}', hive_partitioning=false)")
            out[d["id"].replace(".", "_")] = {"sql": " UNION ALL BY NAME ".join(parts), "layouts": sorted(layouts), "root": str(root),
                                              "kind": d["kind"], "dataset_id": d["id"]}
    _views_cache.update(at=time.monotonic(), value=out)
    return out


_SYM_RE = re.compile(r"\b(?:symbol|symbol_as_reported|underlying)\s*=\s*'([A-Za-z0-9.\-_]+)'", re.I)
_SYM_IN_RE = re.compile(r"\b(?:symbol|symbol_as_reported|underlying)\s+IN\s*\(([^)]+)\)", re.I)
_YEAR_RE = re.compile(r"\byear\s*=\s*(\d{4})\b", re.I)
_YEAR_IN_RE = re.compile(r"\byear\s+IN\s*\(([^)]+)\)", re.I)
_DATE_RE = re.compile(r"\b(?:session_date|trade_date)\s*(>=|<=|=|>|<)\s*(?:DATE\s*)?'(\d{4}-\d{2}-\d{2})'", re.I)
_BETWEEN_RE = re.compile(r"\b(?:session_date|trade_date)\s+BETWEEN\s+(?:DATE\s*)?'(\d{4}-\d{2}-\d{2})'\s+AND\s+(?:DATE\s*)?'(\d{4}-\d{2}-\d{2})'", re.I)


def extract_scope(sql: str) -> dict[str, Any]:
    """Pull simple equality/IN/range predicates on partition columns out of a statement so that the file list can
    be resolved from the catalog (indexed) instead of globbing hundreds of thousands of files."""
    from datetime import date as _date
    syms: set[str] = {m.group(1).upper() for m in _SYM_RE.finditer(sql)}
    for m in _SYM_IN_RE.finditer(sql):
        syms |= {x.strip().strip("'\"").upper() for x in m.group(1).split(",") if x.strip()}
    years: set[int] = {int(m.group(1)) for m in _YEAR_RE.finditer(sql)}
    for m in _YEAR_IN_RE.finditer(sql):
        years |= {int(x) for x in re.findall(r"\d{4}", m.group(1))}
    start = end = None
    for op, d in _DATE_RE.findall(sql):
        dd = _date.fromisoformat(d)
        if op in (">=", ">", "="):
            start = dd if start is None else max(start, dd)
        if op in ("<=", "<", "="):
            end = dd if end is None else min(end, dd)
    for a, b in _BETWEEN_RE.findall(sql):
        start = max(start, _date.fromisoformat(a)) if start else _date.fromisoformat(a)
        end = min(end, _date.fromisoformat(b)) if end else _date.fromisoformat(b)
    if years:
        ys, ye = _date(min(years), 1, 1), _date(max(years), 12, 31)
        start = ys if start is None else max(start, ys)
        end = ye if end is None else min(end, ye)
    return {"symbols": sorted(syms), "start": start, "end": end}


def scoped_view_sql(dataset_id: str, kind: str, scope: dict[str, Any], lake: Lake) -> str | None:
    """Explicit file list from the catalog for symbol-scoped statements (fast path). None → use the glob view."""
    if not scope["symbols"] and not (scope["start"] or scope["end"]):
        return None
    with connection() as conn:
        files = live_files(conn, dataset_id, scope["symbols"] or None, scope["start"], scope["end"])
    if len(files) > 20000:
        return None  # too broad for an explicit list: directory scan instead
    if not files:
        return "SELECT * FROM (VALUES (NULL)) t(empty) WHERE false"
    lst = ", ".join("'" + str(lake.resolve(f["path"])).replace("'", "''") + "'" for f in files)
    if kind == "options":
        return (f"SELECT * EXCLUDE (filename), regexp_extract(filename, 'period=([^/]+)', 1) AS period "
                f"FROM read_parquet([{lst}], union_by_name=true, hive_partitioning=false, filename=true)")
    return (f"SELECT *, symbol_as_reported AS symbol, year(session_date)::BIGINT AS year "
            f"FROM read_parquet([{lst}], union_by_name=true, hive_partitioning=false)")


def _referenced_views(sql: str, views: dict[str, Any]) -> list[str]:
    import re as _re
    low = sql.lower()
    return [name for name in views if _re.search(r"\b" + _re.escape(name) + r"\b", low)]