#!/usr/bin/env python3
from pathlib import Path
import argparse, math
import numpy as np
import pandas as pd
import pyarrow as pa
import pyarrow.parquet as pq

N_STATIONS=1000; N_HOURS=100_000
SPLITS={'train':(0,80_000),'validation':(80_000,90_000),'test':(90_000,100_000)}
VALUE_FIELDS=['temperature_c','relative_humidity_pct','wind_speed_ms','rainfall_mm','pm25_ugm3','pm10_ugm3','no2_ugm3','co_mgm3','so2_ugm3','o3_ugm3']

def main():
    ap=argparse.ArgumentParser()
    ap.add_argument('--input',default='air_quality_100m.npy')
    ap.add_argument('--out',default='data')
    ap.add_argument('--stations',default='stations.csv')
    ap.add_argument('--graph',default='spatial_graph_knn10.csv')
    ap.add_argument('--stations-per-shard',type=int,default=50)
    a=ap.parse_args()
    src=np.load(a.input,mmap_mode='r'); out=Path(a.out); out.mkdir(parents=True,exist_ok=True)
    # Auxiliary station/graph metadata remain CSV files outside the dataset shards.
    # The README config explicitly points the Dataset Viewer only at data/*.parquet.
    pd.read_csv(a.stations); pd.read_csv(a.graph)  # fail fast if metadata are missing/corrupt
    n_shards=math.ceil(N_STATIONS/a.stations_per_shard)
    for split,(h0,h1) in SPLITS.items():
        sdir=out
        for shard,s0 in enumerate(range(0,N_STATIONS,a.stations_per_shard)):
            s1=min(N_STATIONS,s0+a.stations_per_shard)
            x=np.concatenate([src[sid*N_HOURS+h0:sid*N_HOURS+h1] for sid in range(s0,s1)])
            ts=pd.to_datetime(x['timestamp_utc_s'],unit='s',utc=True).tz_convert('Asia/Kolkata')
            cols={'timestamp':pa.Array.from_pandas(ts),'station_id':pa.array(x['station_id'],type=pa.uint16())}
            for f in VALUE_FIELDS: cols[f]=pa.array(x[f],type=pa.float32(),from_pandas=True)
            cols['event_code']=pa.array(x['event_code'],type=pa.uint8())
            cols['obs_mask']=pa.array(x['obs_mask'],type=pa.uint16())
            cols['split_code']=pa.array(x['split_code'],type=pa.uint8())
            table=pa.table(cols)
            path=sdir/f'{split}-{shard:05d}-of-{n_shards:05d}.parquet'
            pq.write_table(table,path,compression='zstd',compression_level=6,row_group_size=250_000,use_dictionary=['station_id','event_code','split_code'])
            print(path,len(x),flush=True)

if __name__=='__main__': main()