thermograph/climate.py

"""Fetch historical + recent daily weather from Open-Meteo and cache to parquet.

The full multi-decade daily record for a grid cell is fetched once and stored as
a zstd-compressed parquet file keyed by cell id. Percentiles are computed on the
fly from this raw record (see grading.py), which keeps the cache small and lets
us handle ties (e.g. many zero-precip days) correctly.
"""
import datetime
import os
import threading
import time

import httpx
import numpy as np
import pandas as pd

import audit
import store

CACHE_DIR = os.path.join(os.path.dirname(__file__), "..", "data", "cache")
MAX_ATTEMPTS = 3                   # per upstream call, before giving up
START_DATE = "1980-01-01"          # ERA5 reaches back to 1940; 1980 = 45 yrs, fast + robust
ARCHIVE_LATENCY_DAYS = 6           # ERA5 archive lags real time by a few days
# History rarely changes, so the full multi-decade archive is fetched once and
# cached indefinitely (refetched only to add new metric columns). Only the recent
# tail is refreshed — a small incremental fetch, at most hourly.
HISTORY_TOPUP_INTERVAL = 3600      # seconds between recent-tail refresh attempts
FORECAST_TTL_HOURS = 1             # refetch the forward forecast hourly to track updates

ARCHIVE_URL = "https://archive-api.open-meteo.com/v1/archive"
FORECAST_URL = "https://api.open-meteo.com/v1/forecast"
# Backup archive when Open-Meteo is unavailable (e.g. its daily rate limit). NASA
# POWER is free + keyless, global, daily from 1981. It lacks gusts + apparent temp,
# so gusts read as unavailable and "feels like" is computed from heat index/chill.
NASA_POWER_URL = "https://power.larc.nasa.gov/api/temporal/daily/point"
NASA_START = "19810101"
NASA_FILL = -900.0   # POWER's missing-value sentinel is ~-999

DAILY_VARS = (
    "temperature_2m_max,temperature_2m_min,precipitation_sum,"
    "wind_speed_10m_max,wind_gusts_10m_max,"
    "apparent_temperature_max,apparent_temperature_min,"
    "relative_humidity_2m_mean"
)

# Columns added after the original tmax/tmin/precip schema. A cached parquet
# missing any of these predates the wind/feels/humidity features (or the split
# apparent high/low) and is refetched so the new metrics show up immediately
# instead of after the 30-day cache TTL.
NEW_COLS = ("wind", "gust", "feels", "humid", "fmax", "fmin")

# Thermoneutral baseline (°F). The combined "feels like" metric reports whichever
# apparent-temperature extreme — the heat-index-driven daily max or the
# wind-chill-driven daily min — sits further from this comfort point, so a single
# daily value captures both heat index (hot side) and wind chill (cold side).
COMFORT_F = 65.0


def _cache_path(cell_id: str) -> str:
    return os.path.join(CACHE_DIR, f"{cell_id}.parquet")


# One lock per cell so concurrent requests/refreshes don't each pull the full
# multi-decade archive (which quickly trips the upstream rate limit).
_LOCKS_GUARD = threading.Lock()
_CELL_LOCKS: dict[str, threading.Lock] = {}

# When the archive rate-limits us (429), back off globally for a bit rather than
# re-hitting it on every refresh — that only prolongs the limit.
ARCHIVE_COOLDOWN = 120  # seconds
_archive_cooldown_until = 0.0


_archive_limit_daily = False   # is the current cooldown a daily-quota exhaustion?


def _is_rate_limit(e) -> bool:
    return getattr(getattr(e, "response", None), "status_code", None) == 429


def _rate_limit_reason(e) -> str:
    """Upstream's human-readable 429 reason, if present in the response body."""
    resp = getattr(e, "response", None)
    if resp is not None:
        try:
            j = resp.json()
            if isinstance(j, dict) and j.get("reason"):
                return str(j["reason"])
        except Exception:  # noqa: BLE001
            pass
    return ""


def _seconds_to_utc_reset() -> float:
    """Seconds until just after the next UTC midnight (when daily quotas reset)."""
    now = datetime.datetime.now(datetime.timezone.utc)
    reset = (now + datetime.timedelta(days=1)).replace(hour=0, minute=10, second=0, microsecond=0)
    return max((reset - now).total_seconds(), 600)


def _note_rate_limit(e) -> None:
    """Record a rate limit: back off ~2 min, or until tomorrow for a daily quota."""
    global _archive_cooldown_until, _archive_limit_daily
    reason = _rate_limit_reason(e).lower()
    _archive_limit_daily = "daily" in reason or "tomorrow" in reason
    _archive_cooldown_until = time.time() + (
        _seconds_to_utc_reset() if _archive_limit_daily else ARCHIVE_COOLDOWN)


def _cooldown_message() -> str:
    if _archive_limit_daily:
        return ("Open-Meteo's daily request limit is exhausted — new locations will work again "
                "tomorrow. Places you've already viewed still work.")
    return "The weather archive is rate-limited right now — please try again in a minute."


def _cell_lock(cell_id: str) -> threading.Lock:
    with _LOCKS_GUARD:
        lk = _CELL_LOCKS.get(cell_id)
        if lk is None:
            lk = _CELL_LOCKS[cell_id] = threading.Lock()
        return lk


def _request(url, params, timeout, *, phase, headers=None, attempts=MAX_ATTEMPTS):
    """GET with bounded retries; every retry and the final failure are logged to
    the errors folder (tagged ``retry`` / ``error``). A 429 (rate limit) fails fast
    without retrying, so we don't hammer the limit and make it worse."""
    last = None
    for attempt in range(1, attempts + 1):
        try:
            r = httpx.get(url, params=params, timeout=timeout, headers=headers)
            r.raise_for_status()
            return r
        except Exception as e:  # noqa: BLE001 - upstream/network failures are expected
            last = e
            status = getattr(getattr(e, "response", None), "status_code", None)
            rate_limited = status == 429
            final = rate_limited or attempt == attempts
            audit.log_event(
                "error" if final else "retry",
                {"phase": phase, "attempt": attempt, "max_attempts": attempts,
                 "url": url, "status": status, "error": repr(e)},
            )
            if final:
                break
            time.sleep(min(0.5 * 2 ** (attempt - 1), 4.0))
    raise last


def _derive_humidity(df: pd.DataFrame) -> pd.DataFrame:
    """Replace the raw mean *relative* humidity column with *absolute* humidity
    (grams of water vapor per m³), in place.

    Absolute humidity is derived from the day's mean RH and its mean temperature
    ((tmax+tmin)/2) via the Magnus saturation-vapor-pressure formula. It's a far
    more informative "how muggy was it" signal than relative humidity, which mostly
    tracks the day/night temperature swing (cold nights read ~100% RH regardless of
    actual moisture). Applied at the read boundary so the parquet cache keeps the
    raw RH the archive returns — no refetch needed for existing cells."""
    if "humid" not in df.columns:
        return df
    tmean_c = ((df["tmax"] + df["tmin"]) / 2.0 - 32.0) * 5.0 / 9.0
    rh = pd.to_numeric(df["humid"], errors="coerce")
    es = 6.112 * np.exp(17.67 * tmean_c / (tmean_c + 243.5))   # sat. vapor pressure, hPa
    df["humid"] = (es * rh * 2.1674 / (273.15 + tmean_c)).round(1)   # absolute humidity, g/m³
    return df


def _combined_feels(amax: pd.Series, amin: pd.Series) -> pd.Series:
    """One daily "feels like" value: the apparent-temperature extreme furthest from
    the comfort baseline — the heat-index high on warm days, the wind-chill low on
    cold ones. Falls back to whichever side is present if one is missing."""
    hot = amax - COMFORT_F
    cold = COMFORT_F - amin
    feels = amax.where(hot >= cold, amin)   # NaN comparisons pick the min side
    return feels.fillna(amax).fillna(amin)


def _to_frame(daily: dict) -> pd.DataFrame:
    n = len(daily["time"])
    # Older upstream responses (or a narrowed variable set) may omit a series; fall
    # back to an all-null column of the right length so the frame shape is stable.
    def col(key):
        vals = daily.get(key)
        return vals if vals else [None] * n

    df = pd.DataFrame(
        {
            "date": pd.to_datetime(daily["time"]),
            "tmax": daily["temperature_2m_max"],
            "tmin": daily["temperature_2m_min"],
            "precip": daily["precipitation_sum"],
            "wind": col("wind_speed_10m_max"),
            "gust": col("wind_gusts_10m_max"),
            "humid": col("relative_humidity_2m_mean"),
        }
    )
    # The apparent (felt) high and low are kept as their own columns — graded
    # independently — alongside the combined `feels` (whichever side is further
    # from the fixed comfort baseline), which the weekly/day views still use.
    amax = pd.to_numeric(pd.Series(col("apparent_temperature_max")), errors="coerce")
    amin = pd.to_numeric(pd.Series(col("apparent_temperature_min")), errors="coerce")
    df["fmax"] = amax
    df["fmin"] = amin
    df["feels"] = _combined_feels(amax, amin)
    # Need a valid high/low to be a usable climate day; the rest may be NaN and are
    # simply ungraded for that day (the percentile helpers return None).
    df = df.dropna(subset=["tmax", "tmin"]).reset_index(drop=True)
    df["doy"] = df["date"].dt.dayofyear.astype("int16")
    return df


def _fetch_history(cell: dict) -> pd.DataFrame:
    end = (datetime.date.today() - datetime.timedelta(days=ARCHIVE_LATENCY_DAYS)).isoformat()
    params = {
        "latitude": cell["center_lat"],
        "longitude": cell["center_lon"],
        "start_date": START_DATE,
        "end_date": end,
        "daily": DAILY_VARS,
        "timezone": "auto",
        "temperature_unit": "fahrenheit",
        "precipitation_unit": "inch",
        "wind_speed_unit": "mph",
    }
    r = _request(ARCHIVE_URL, params, 180, phase="history_fetch")
    return _to_frame(r.json()["daily"])


def _heat_index(t_f, rh):
    """NWS heat index (°F); below ~80°F it's just the air temperature."""
    T, R = t_f, rh
    hi = (-42.379 + 2.04901523 * T + 10.14333127 * R - 0.22475541 * T * R
          - 0.00683783 * T * T - 0.05481717 * R * R + 0.00122874 * T * T * R
          + 0.00085282 * T * R * R - 0.00000199 * T * T * R * R)
    return np.where(np.asarray(T, dtype="float64") >= 80, hi, T)


def _wind_chill(t_f, v_mph):
    """NWS wind chill (°F); applies only when cold + breezy, else the air temp."""
    T = np.asarray(t_f, dtype="float64")
    V = np.clip(np.asarray(v_mph, dtype="float64"), 0, None)
    Vp = np.power(V, 0.16)
    wc = 35.74 + 0.6215 * T - 35.75 * Vp + 0.4275 * T * Vp
    return np.where((T <= 50) & (V >= 3), wc, T)


def _nasa_to_frame(param: dict) -> pd.DataFrame:
    """Map a NASA POWER daily response to our (tmax/tmin/precip/wind/gust/humid/feels)
    schema, converting units (°C→°F, mm→in, m/s→mph) and computing feels-like."""
    dates = sorted(param.get("T2M_MAX", {}).keys())

    def col(key, transform):
        d = param.get(key, {})
        return [np.nan if (v is None or v <= NASA_FILL) else transform(v)
                for v in (d.get(k) for k in dates)]

    c2f = lambda c: c * 9.0 / 5.0 + 32.0
    df = pd.DataFrame({
        "date": pd.to_datetime(dates, format="%Y%m%d"),
        "tmax": col("T2M_MAX", c2f),
        "tmin": col("T2M_MIN", c2f),
        "precip": col("PRECTOTCORR", lambda mm: mm / 25.4),
        "wind": col("WS10M_MAX", lambda ms: ms * 2.2369362920544),
        "gust": [np.nan] * len(dates),          # POWER has no gusts
        "humid": col("RH2M", lambda v: v),
    })
    # POWER has no apparent temperature; approximate the felt high with the NWS
    # heat index and the felt low with NWS wind chill (each falls back to the air
    # temperature outside its regime), mirroring the Open-Meteo columns.
    df["fmax"] = pd.Series(_heat_index(df["tmax"], df["humid"]), index=df.index)
    df["fmin"] = pd.Series(_wind_chill(df["tmin"], df["wind"]), index=df.index)
    df["feels"] = _combined_feels(df["fmax"], df["fmin"])
    df = df.dropna(subset=["tmax", "tmin"]).reset_index(drop=True)
    df["doy"] = df["date"].dt.dayofyear.astype("int16")
    return df


def _fetch_history_nasa(cell: dict) -> pd.DataFrame:
    """Backup history fetch from NASA POWER (used when Open-Meteo is unavailable)."""
    end = (datetime.date.today() - datetime.timedelta(days=ARCHIVE_LATENCY_DAYS)).strftime("%Y%m%d")
    params = {
        "parameters": "T2M_MAX,T2M_MIN,PRECTOTCORR,WS10M_MAX,RH2M",
        "community": "RE",
        "latitude": cell["center_lat"],
        "longitude": cell["center_lon"],
        "start": NASA_START,
        "end": end,
        "format": "JSON",
    }
    r = _request(NASA_POWER_URL, params, 180, phase="history_nasa")
    return _nasa_to_frame(r.json()["properties"]["parameter"])


def _fetch_history_range(cell: dict, start_date: str, end_date: str) -> pd.DataFrame:
    """Fetch just a date range of archive history (used to top up the recent tail)."""
    params = {
        "latitude": cell["center_lat"],
        "longitude": cell["center_lon"],
        "start_date": start_date,
        "end_date": end_date,
        "daily": DAILY_VARS,
        "timezone": "auto",
        "temperature_unit": "fahrenheit",
        "precipitation_unit": "inch",
        "wind_speed_unit": "mph",
    }
    r = _request(ARCHIVE_URL, params, 60, phase="history_topup")
    return _to_frame(r.json()["daily"])


def _read_history_cache(path):
    """Schema-complete cached frame (no doy) + its file age in seconds, or None.

    No age expiry — history is cached indefinitely; a pre-wind/humidity schema is
    the only reason to refetch (to add the new metric columns)."""
    if not os.path.exists(path):
        return None
    df = pd.read_parquet(path)
    if not all(c in df.columns for c in NEW_COLS):
        return None
    return df, time.time() - os.path.getmtime(path)


def _topup_tail(cell: dict, df: pd.DataFrame, path: str) -> pd.DataFrame:
    """Append newly-available archive days to a cached record. Best-effort and
    serialized per cell; a small incremental fetch, not the full multi-decade pull."""
    global _archive_cooldown_until
    with _cell_lock(cell["id"]):
        fresh = _read_history_cache(path)   # another thread may have just refreshed it
        if fresh is not None:
            df, age_s = fresh
            if age_s < HISTORY_TOPUP_INTERVAL:
                return df
        if time.time() < _archive_cooldown_until:
            return df
        expected = datetime.date.today() - datetime.timedelta(days=ARCHIVE_LATENCY_DAYS)
        cached_max = pd.Timestamp(df["date"].max()).date()
        if cached_max >= expected:
            try: os.utime(path, None)   # tail already current — reset the hourly timer
            except OSError: pass
            return df
        try:
            recent = _fetch_history_range(
                cell, (cached_max + datetime.timedelta(days=1)).isoformat(), expected.isoformat())
        except Exception as e:  # noqa: BLE001 - keep the cached record on failure
            if _is_rate_limit(e):
                _note_rate_limit(e)
            return df
        merged = (pd.concat([df, recent.drop(columns=["doy"], errors="ignore")])
                  .drop_duplicates(subset="date", keep="last")
                  .sort_values("date").reset_index(drop=True))
        merged.to_parquet(path, compression="zstd", index=False)
        return merged


def get_history(cell: dict) -> tuple[pd.DataFrame, dict]:
    """Return (daily history frame, cache metadata) for a cell, with humidity as
    absolute humidity (g/m³). Thin wrapper over the raw loader (see below)."""
    df, meta = _load_history(cell)
    _derive_humidity(df)
    return df, meta


def load_cached_history(cell: dict) -> pd.DataFrame | None:
    """History from the parquet cache ONLY — never fetches upstream and never tops
    up the tail. Powers the warm-only prefetch path (which must not spend upstream
    quota) and the offline migrate script. None when the cell has no
    (schema-complete) cached record."""
    hit = _read_history_cache(_cache_path(cell["id"]))
    if hit is None:
        return None
    df = hit[0].copy()
    df["doy"] = df["date"].dt.dayofyear.astype("int16")
    _derive_humidity(df)
    return df


def _load_history(cell: dict) -> tuple[pd.DataFrame, dict]:
    """Return (daily history frame, cache metadata) for a cell.

    The full archive is cached indefinitely (fetched once); only the recent tail is
    topped up, at most hourly. Concurrent callers are serialized so only one archive
    fetch happens; on an upstream failure (e.g. 429) any existing cache is served."""
    path = _cache_path(cell["id"])

    hit = _read_history_cache(path)
    if hit is not None:
        df, age_s = hit
        if age_s > HISTORY_TOPUP_INTERVAL:
            df = _topup_tail(cell, df, path)   # refresh just the recent days
        df = df.copy()
        df["doy"] = df["date"].dt.dayofyear.astype("int16")
        return df, {"cached": True, "cache_age_days": round(age_s / 86400.0, 1)}

    global _archive_cooldown_until

    with _cell_lock(cell["id"]):
        hit = _read_history_cache(path)   # another thread may have populated it while we waited
        if hit is not None:
            df, age_s = hit
            df["doy"] = df["date"].dt.dayofyear.astype("int16")
            return df, {"cached": True, "cache_age_days": round(age_s / 86400.0, 1)}

        stale = pd.read_parquet(path) if os.path.exists(path) else None

        def _serve_stale():
            stale["doy"] = stale["date"].dt.dayofyear.astype("int16")
            return stale, {"cached": True, "stale": True, "cache_age_days": None}

        # Fetch. Open-Meteo is primary (richer: gusts + apparent temp); NASA POWER is
        # the backup when Open-Meteo is unavailable (network error or its daily limit).
        # Skip Open-Meteo entirely while it's in a rate-limit cooldown.
        df = None
        source = None
        if time.time() >= _archive_cooldown_until:
            try:
                df = _fetch_history(cell)
                source = "open-meteo"
            except Exception as e:  # noqa: BLE001
                if _is_rate_limit(e):
                    _note_rate_limit(e)
        if df is None:
            try:
                df = _fetch_history_nasa(cell)
                source = "nasa-power"
            except Exception:  # noqa: BLE001 - backup unavailable too
                df = None
        if df is None:
            if stale is not None:
                return _serve_stale()
            raise RuntimeError(_cooldown_message())

        os.makedirs(CACHE_DIR, exist_ok=True)
        # Store the raw record; percentiles are derived at request time.
        df.drop(columns=["doy"]).to_parquet(path, compression="zstd", index=False)
        return df, {"cached": False, "cache_age_days": 0, "source": source}


RECENT_PAST_DAYS = 25   # recent observations window (covers the ~2-week graded view)
FORECAST_DAYS = 8       # today + 7 days ahead


def _rf_cache_path(cell_id: str) -> str:
    return os.path.join(CACHE_DIR, f"{cell_id}_rf.parquet")


def recent_stamp(cell_id: str) -> int:
    """Identity stamp (mtime, whole seconds) of the cell's cached recent+forecast
    parquet; 0 when absent. Changes exactly when the recent/forecast data does, so
    it's the freshness token for every payload that grades recent or future days."""
    try:
        return int(os.path.getmtime(_rf_cache_path(cell_id)))
    except OSError:
        return 0


def get_recent_forecast(cell: dict) -> pd.DataFrame:
    """Recent observations + forward forecast, with humidity as absolute humidity
    (g/m³). Thin wrapper over the raw loader (see below)."""
    df = _load_recent_forecast(cell)
    _derive_humidity(df)
    return df


def _load_recent_forecast(cell: dict) -> pd.DataFrame:
    """Recent observations AND the forward forecast in ONE forecast-API call.

    Both the recent (past) view and the forecast (future) view slice from this, so
    a cell needs just one upstream forecast request per hour (plus the ~monthly
    archive fetch). Cached per cell for one hour so it still follows model updates.
    """
    path = _rf_cache_path(cell["id"])
    if os.path.exists(path):
        age_h = (time.time() - os.path.getmtime(path)) / 3600.0
        if age_h < FORECAST_TTL_HOURS:
            df = pd.read_parquet(path)
            df["doy"] = df["date"].dt.dayofyear.astype("int16")
            return df

    params = {
        "latitude": cell["center_lat"],
        "longitude": cell["center_lon"],
        "daily": DAILY_VARS,
        "timezone": "auto",
        "temperature_unit": "fahrenheit",
        "precipitation_unit": "inch",
        "wind_speed_unit": "mph",
        "past_days": RECENT_PAST_DAYS,
        "forecast_days": FORECAST_DAYS,
    }
    r = _request(FORECAST_URL, params, 60, phase="recent_forecast_fetch")
    df = _to_frame(r.json()["daily"])
    os.makedirs(CACHE_DIR, exist_ok=True)
    df.drop(columns=["doy"]).to_parquet(path, compression="zstd", index=False)
    return df


_REVGEO_CACHE: dict[str, str | None] = {}


def reverse_geocode_cached(lat: float, lon: float) -> tuple[bool, str | None]:
    """(found, label) from the in-memory + SQLite revgeo caches only — never calls
    Nominatim. Used by paths that must not add upstream traffic (prefetch)."""
    key = store.revgeo_key(lat, lon)
    if key in _REVGEO_CACHE:
        return (True, _REVGEO_CACHE[key])
    found, label = store.get_revgeo(key)
    if found:
        _REVGEO_CACHE[key] = label
    return (found, label)


def reverse_geocode(lat: float, lon: float) -> str | None:
    """Best-effort neighbourhood/city label for a point (OpenStreetMap Nominatim).

    Cached per ~cell — in memory for the process and in SQLite across restarts —
    so panning around (or redeploying) doesn't re-hit the service, and failures
    return None so the caller can fall back to bare coordinates.
    """
    key = store.revgeo_key(lat, lon)
    found, label = reverse_geocode_cached(lat, lon)
    if found:
        return label
    label = None
    try:
        r = _request(
            "https://nominatim.openstreetmap.org/reverse",
            # zoom 14 resolves to the suburb/neighbourhood level so we can lead
            # with it when OSM has one (zoom 10 only ever returns the city).
            {"lat": lat, "lon": lon, "format": "jsonv2", "zoom": 14,
             "addressdetails": 1},
            15,
            phase="reverse_geocode",
            headers={"User-Agent": "Thermograph/0.1 (local weather grading app)"},
        )
        a = r.json().get("address", {}) or {}
        # Lead with the neighbourhood when available, then the city, then region.
        neighborhood = (a.get("neighbourhood") or a.get("suburb")
                        or a.get("quarter") or a.get("city_district")
                        or a.get("borough"))
        city = (a.get("city") or a.get("town") or a.get("village")
                or a.get("hamlet") or a.get("county"))
        region = a.get("state") or a.get("province") or a.get("region")
        # dict.fromkeys drops duplicates (e.g. neighbourhood == city) in order.
        parts = dict.fromkeys(p for p in (neighborhood, city, region) if p)
        label = ", ".join(parts) or None
    except Exception:  # noqa: BLE001 - reverse geocoding is a nicety, never fatal
        label = None
    _REVGEO_CACHE[key] = label
    store.put_revgeo(key, label)   # survive restarts (a None label is retried daily)
    return label


def geocode(name: str, count: int = 5) -> list[dict]:
    """Look up places by name (US/Canada biased) via Open-Meteo's geocoder."""
    r = _request(
        "https://geocoding-api.open-meteo.com/v1/search",
        {"name": name, "count": count, "language": "en", "format": "json"},
        30,
        phase="geocode",
    )
    results = r.json().get("results", []) or []
    return [
        {
            "name": g.get("name"),
            "admin1": g.get("admin1"),
            "country": g.get("country"),
            "country_code": g.get("country_code"),
            "lat": g.get("latitude"),
            "lon": g.get("longitude"),
        }
        for g in results
    ]