FEATURE: Add contributions framework. Fix and improve daily adsb release using Github actions for map reduce.

src/adsb/Dockerfile.reducer

@@ -5,6 +5,7 @@ WORKDIR /app
 COPY requirements.reducer.txt requirements.txt
 RUN pip install --no-cache-dir -r requirements.txt
 
+COPY compress_adsb_to_aircraft_data.py .
 COPY reducer.py .
 
 CMD ["python", "-u", "reducer.py"]

src/adsb/Dockerfile.worker

@@ -6,6 +6,7 @@ COPY requirements.worker.txt requirements.txt
 RUN pip install --no-cache-dir -r requirements.txt
 
 COPY compress_adsb_to_aircraft_data.py .
+COPY download_adsb_data_to_parquet.py .
 COPY worker.py .
 
 CMD ["python", "-u", "worker.py"]

src/adsb/adsb_to_aircraft_data_daily.py

@@ -1,7 +0,0 @@
-from pathlib import Path
-from datetime import datetime, timezone,timedelta
-from adsb_to_aircraft_data_historical import load_historical_for_day
-
-
-day = datetime.now(timezone.utc) - timedelta(days=1)
-load_historical_for_day(day)

src/adsb/adsb_to_aircraft_data_historical.py

@@ -1,87 +0,0 @@
-"""
-Process historical ADS-B data by date range.
-Downloads and compresses ADS-B messages for each day in the specified range.
-"""
-import argparse
-from datetime import datetime, timedelta
-from pathlib import Path
-import pandas as pd
-from compress_adsb_to_aircraft_data import load_historical_for_day, COLUMNS
-
-def deduplicate_by_signature(df):
-    """For each icao, keep only the earliest row with each unique signature."""
-    df["_signature"] = df[COLUMNS].astype(str).agg('|'.join, axis=1)
-    # Group by icao and signature, keep first (earliest) occurrence
-    df_deduped = df.groupby(['icao', '_signature'], as_index=False).first()
-    df_deduped = df_deduped.drop(columns=['_signature'])
-    df_deduped = df_deduped.sort_values('time')
-    return df_deduped
-
-
-def main(start_date_str: str, end_date_str: str):
-    """Process historical ADS-B data for the given date range."""
-    OUT_ROOT = Path("data/planequery_aircraft")
-    OUT_ROOT.mkdir(parents=True, exist_ok=True)
-
-    # Parse dates
-    start_date = datetime.strptime(start_date_str, "%Y-%m-%d")
-    end_date = datetime.strptime(end_date_str, "%Y-%m-%d")
-
-    # Calculate total number of days
-    total_days = (end_date - start_date).days
-    print(f"Processing {total_days} days from {start_date_str} to {end_date_str}")
-
-    # Initialize accumulated dataframe
-    df_accumulated = pd.DataFrame()
-
-    # Cache directory path
-    cache_dir = Path("data/adsb")
-
-    # Iterate through each day
-    current_date = start_date
-    while current_date < end_date:
-        print(f"Processing {current_date.strftime('%Y-%m-%d')}...")
-        
-        df_compressed = load_historical_for_day(current_date)
-        
-        # Concatenate to accumulated dataframe
-        if df_accumulated.empty:
-            df_accumulated = df_compressed
-        else:
-            df_accumulated = pd.concat([df_accumulated, df_compressed], ignore_index=True)
-        
-        print(f"  Added {len(df_compressed)} records (total: {len(df_accumulated)})")
-        
-        # Save intermediate output after each day
-        current_date_str = current_date.strftime('%Y-%m-%d')
-        output_file = OUT_ROOT / f"planequery_aircraft_adsb_{start_date_str}_{current_date_str}.csv.gz"
-        df_deduped = deduplicate_by_signature(df_accumulated.copy())
-        df_deduped.to_csv(output_file, index=False, compression='gzip')
-        print(f"  Saved to {output_file.name}")
-        
-        # Delete cache after processing if processing more than 10 days
-        if total_days > 5 and cache_dir.exists():
-            import shutil
-            shutil.rmtree(cache_dir)
-            print(f"  Deleted cache directory to save space")
-        
-        # Move to next day
-        current_date += timedelta(days=1)
-
-    # Save the final accumulated data
-    output_file = OUT_ROOT / f"planequery_aircraft_adsb_{start_date_str}_{end_date_str}.csv.gz"
-    df_accumulated = deduplicate_by_signature(df_accumulated)
-    df_accumulated.to_csv(output_file, index=False, compression='gzip')
-
-    print(f"Completed processing from {start_date_str} to {end_date_str}")
-    print(f"Saved {len(df_accumulated)} total records to {output_file}")
-
-
-if __name__ == '__main__':
-    # Parse command line arguments
-    parser = argparse.ArgumentParser(description="Process historical ADS-B data from ClickHouse")
-    parser.add_argument("start_date", help="Start date (YYYY-MM-DD, inclusive)")
-    parser.add_argument("end_date", help="End date (YYYY-MM-DD, exclusive)")
-    args = parser.parse_args()
-    
-    main(args.start_date, args.end_date)

src/adsb/combine_chunks_to_csv.py

@@ -0,0 +1,205 @@
+"""
+Combines chunk parquet files and compresses to final aircraft CSV.
+This is the reduce phase of the map-reduce pipeline.
+
+Memory-efficient: processes each chunk separately, compresses, then combines.
+
+Usage:
+    python -m src.adsb.combine_chunks_to_csv --chunks-dir data/output/adsb_chunks
+"""
+import gc
+import os
+import sys
+import glob
+import argparse
+from datetime import datetime, timedelta
+
+import polars as pl
+
+from src.adsb.download_adsb_data_to_parquet import OUTPUT_DIR, get_resource_usage
+from src.adsb.compress_adsb_to_aircraft_data import compress_multi_icao_df, COLUMNS
+
+
+DEFAULT_CHUNK_DIR = os.path.join(OUTPUT_DIR, "adsb_chunks")
+FINAL_OUTPUT_DIR = "./data/planequery_aircraft"
+os.makedirs(FINAL_OUTPUT_DIR, exist_ok=True)
+
+
+def get_target_day() -> datetime:
+    """Get yesterday's date (the day we're processing)."""
+    return datetime.utcnow() - timedelta(days=1)
+
+
+def process_single_chunk(chunk_path: str) -> pl.DataFrame:
+    """Load and compress a single chunk parquet file."""
+    print(f"Processing {os.path.basename(chunk_path)}... | {get_resource_usage()}")
+    
+    # Load chunk - only columns we need
+    needed_columns = ['time', 'icao'] + COLUMNS
+    df = pl.read_parquet(chunk_path, columns=needed_columns)
+    print(f"  Loaded {len(df)} rows")
+    
+    # Compress to aircraft records (one per ICAO) using shared function
+    compressed = compress_multi_icao_df(df, verbose=True)
+    print(f"  Compressed to {len(compressed)} aircraft records")
+    
+    del df
+    gc.collect()
+    
+    return compressed
+
+
+def combine_compressed_chunks(compressed_dfs: list[pl.DataFrame]) -> pl.DataFrame:
+    """Combine multiple compressed DataFrames.
+    
+    Since chunks are partitioned by ICAO hash, each ICAO only appears in one chunk.
+    No deduplication needed here - just concatenate.
+    """
+    print(f"Combining {len(compressed_dfs)} compressed chunks... | {get_resource_usage()}")
+    
+    # Concat all
+    combined = pl.concat(compressed_dfs)
+    print(f"Combined: {len(combined)} records")
+    
+    return combined
+
+
+def download_and_merge_base_release(compressed_df: pl.DataFrame) -> pl.DataFrame:
+    """Download base release and merge with new data."""
+    from src.get_latest_planequery_aircraft_release import download_latest_aircraft_adsb_csv
+    
+    print("Downloading base ADS-B release...")
+    try:
+        base_path = download_latest_aircraft_adsb_csv(
+            output_dir="./data/planequery_aircraft_base"
+        )
+        print(f"Download returned: {base_path}")
+        
+        if base_path and os.path.exists(str(base_path)):
+            print(f"Loading base release from {base_path}")
+            base_df = pl.read_csv(base_path)
+            print(f"Base release has {len(base_df)} records")
+            
+            # Ensure columns match
+            base_cols = set(base_df.columns)
+            new_cols = set(compressed_df.columns)
+            print(f"Base columns: {sorted(base_cols)}")
+            print(f"New columns: {sorted(new_cols)}")
+            
+            # Add missing columns
+            for col in new_cols - base_cols:
+                base_df = base_df.with_columns(pl.lit(None).alias(col))
+            for col in base_cols - new_cols:
+                compressed_df = compressed_df.with_columns(pl.lit(None).alias(col))
+            
+            # Reorder columns to match
+            compressed_df = compressed_df.select(base_df.columns)
+            
+            # Concat and deduplicate by icao (keep new data - it comes last)
+            combined = pl.concat([base_df, compressed_df])
+            print(f"After concat: {len(combined)} records")
+            
+            deduplicated = combined.unique(subset=["icao"], keep="last")
+            
+            print(f"Combined with base: {len(combined)} -> {len(deduplicated)} after dedup")
+            
+            del base_df, combined
+            gc.collect()
+            
+            return deduplicated
+        else:
+            print(f"No base release found at {base_path}, using only new data")
+            return compressed_df
+    except Exception as e:
+        import traceback
+        print(f"Failed to download base release: {e}")
+        traceback.print_exc()
+        return compressed_df
+
+
+def cleanup_chunks(date_str: str, chunks_dir: str):
+    """Delete chunk parquet files after successful merge."""
+    pattern = os.path.join(chunks_dir, f"chunk_*_{date_str}.parquet")
+    chunk_files = glob.glob(pattern)
+    for f in chunk_files:
+        try:
+            os.remove(f)
+            print(f"Deleted {f}")
+        except Exception as e:
+            print(f"Failed to delete {f}: {e}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Combine chunk parquets to final CSV")
+    parser.add_argument("--date", type=str, help="Date in YYYY-MM-DD format (default: yesterday)")
+    parser.add_argument("--chunks-dir", type=str, default=DEFAULT_CHUNK_DIR, help="Directory containing chunk parquet files")
+    parser.add_argument("--skip-base", action="store_true", help="Skip downloading and merging base release")
+    parser.add_argument("--keep-chunks", action="store_true", help="Keep chunk files after merging")
+    args = parser.parse_args()
+    
+    if args.date:
+        target_day = datetime.strptime(args.date, "%Y-%m-%d")
+    else:
+        target_day = get_target_day()
+    
+    date_str = target_day.strftime("%Y-%m-%d")
+    chunks_dir = args.chunks_dir
+    
+    print(f"Combining chunks for {date_str}")
+    print(f"Chunks directory: {chunks_dir}")
+    print(f"Resource usage at start: {get_resource_usage()}")
+    
+    # Find chunk files
+    pattern = os.path.join(chunks_dir, f"chunk_*_{date_str}.parquet")
+    chunk_files = sorted(glob.glob(pattern))
+    
+    if not chunk_files:
+        print(f"No chunk files found matching: {pattern}")
+        sys.exit(1)
+    
+    print(f"Found {len(chunk_files)} chunk files")
+    
+    # Process each chunk separately to save memory
+    compressed_chunks = []
+    for chunk_path in chunk_files:
+        compressed = process_single_chunk(chunk_path)
+        compressed_chunks.append(compressed)
+        gc.collect()
+    
+    # Combine all compressed chunks
+    combined = combine_compressed_chunks(compressed_chunks)
+    
+    # Free memory from individual chunks
+    del compressed_chunks
+    gc.collect()
+    print(f"After combining: {get_resource_usage()}")
+    
+    # Merge with base release
+    if not args.skip_base:
+        combined = download_and_merge_base_release(combined)
+    
+    # Convert list columns to strings for CSV compatibility
+    for col in combined.columns:
+        if combined[col].dtype == pl.List:
+            combined = combined.with_columns(
+                pl.col(col).list.join(",").alias(col)
+            )
+    
+    # Sort by time for consistent output
+    if 'time' in combined.columns:
+        combined = combined.sort('time')
+    
+    # Write final CSV
+    output_path = os.path.join(FINAL_OUTPUT_DIR, f"planequery_aircraft_adsb_{date_str}.csv")
+    combined.write_csv(output_path)
+    print(f"Wrote {len(combined)} records to {output_path}")
+    
+    # Cleanup
+    if not args.keep_chunks:
+        cleanup_chunks(date_str, chunks_dir)
+    
+    print(f"Done! | {get_resource_usage()}")
+
+
+if __name__ == "__main__":
+    main()

src/adsb/compress_adsb_to_aircraft_data.py

@@ -1,32 +1,67 @@
-# SOME KIND OF MAP REDUCE SYSTEM
+# Shared compression logic for ADS-B aircraft data
 import os
+import polars as pl
 
 COLUMNS = ['dbFlags', 'ownOp', 'year', 'desc', 'aircraft_category', 'r', 't']
-def compress_df(df):
-    icao = df.name
-    df["_signature"] = df[COLUMNS].astype(str).agg('|'.join, axis=1)
+
+
+def deduplicate_by_signature(df: pl.DataFrame) -> pl.DataFrame:
+    """For each icao, keep only the earliest row with each unique signature.
     
-    # Compute signature counts before grouping (avoid copy)
-    signature_counts = df["_signature"].value_counts()
+    This is used for deduplicating across multiple compressed chunks.
+    """
+    # Create signature column
+    df = df.with_columns(
+        pl.concat_str([pl.col(c).cast(pl.Utf8).fill_null("") for c in COLUMNS], separator="|").alias("_signature")
+    )
+    # Group by icao and signature, take first row (earliest due to time sort)
+    df = df.sort("time")
+    df_deduped = df.group_by(["icao", "_signature"]).first()
+    df_deduped = df_deduped.drop("_signature")
+    df_deduped = df_deduped.sort("time")
+    return df_deduped
+
+
+def compress_df_polars(df: pl.DataFrame, icao: str) -> pl.DataFrame:
+    """Compress a single ICAO group to its most informative row using Polars."""
+    # Create signature string
+    df = df.with_columns(
+        pl.concat_str([pl.col(c).cast(pl.Utf8) for c in COLUMNS], separator="|").alias("_signature")
+    )
     
-    df = df.groupby("_signature", as_index=False).first() # check if it works with both last and first.
-    # For each row, create a dict of non-empty column values. This is using sets and subsets...
-    def get_non_empty_dict(row):
-        return {col: row[col] for col in COLUMNS if row[col] != ''}
+    # Compute signature counts
+    signature_counts = df.group_by("_signature").len().rename({"len": "_sig_count"})
     
-    df['_non_empty_dict'] = df.apply(get_non_empty_dict, axis=1)
-    df['_non_empty_count'] = df['_non_empty_dict'].apply(len)
+    # Group by signature and take first row
+    df = df.group_by("_signature").first()
+    
+    if df.height == 1:
+        # Only one unique signature, return it
+        result = df.drop("_signature").with_columns(pl.lit(icao).alias("icao"))
+        return result
+    
+    # For each row, create dict of non-empty column values and check subsets
+    # Convert to list of dicts for subset checking (same logic as pandas version)
+    rows_data = []
+    for row in df.iter_rows(named=True):
+        non_empty = {col: row[col] for col in COLUMNS if row[col] != '' and row[col] is not None}
+        rows_data.append({
+            'signature': row['_signature'],
+            'non_empty_dict': non_empty,
+            'non_empty_count': len(non_empty),
+            'row_data': row
+        })
     
     # Check if row i's non-empty values are a subset of row j's non-empty values
     def is_subset_of_any(idx):
-        row_dict = df.loc[idx, '_non_empty_dict']
-        row_count = df.loc[idx, '_non_empty_count']
+        row_dict = rows_data[idx]['non_empty_dict']
+        row_count = rows_data[idx]['non_empty_count']
         
-        for other_idx in df.index:
+        for other_idx, other_data in enumerate(rows_data):
             if idx == other_idx:
                 continue
-            other_dict = df.loc[other_idx, '_non_empty_dict']
-            other_count = df.loc[other_idx, '_non_empty_count']
+            other_dict = other_data['non_empty_dict']
+            other_count = other_data['non_empty_count']
             
             # Check if all non-empty values in current row match those in other row
             if all(row_dict.get(k) == other_dict.get(k) for k in row_dict.keys()):
@@ -36,32 +71,94 @@ def compress_df(df):
         return False
     
     # Keep rows that are not subsets of any other row
-    keep_mask = ~df.index.to_series().apply(is_subset_of_any)
-    df = df[keep_mask]
+    keep_indices = [i for i in range(len(rows_data)) if not is_subset_of_any(i)]
+    
+    if len(keep_indices) == 0:
+        keep_indices = [0]  # Fallback: keep first row
+    
+    remaining_signatures = [rows_data[i]['signature'] for i in keep_indices]
+    df = df.filter(pl.col("_signature").is_in(remaining_signatures))
+    
+    if df.height > 1:
+        # Use signature counts to pick the most frequent one
+        df = df.join(signature_counts, on="_signature", how="left")
+        max_count = df["_sig_count"].max()
+        df = df.filter(pl.col("_sig_count") == max_count).head(1)
+        df = df.drop("_sig_count")
+    
+    result = df.drop("_signature").with_columns(pl.lit(icao).alias("icao"))
+    
+    # Ensure empty strings are preserved
+    for col in COLUMNS:
+        if col in result.columns:
+            result = result.with_columns(pl.col(col).fill_null(""))
+    
+    return result
 
-    if len(df) > 1:
-        # Use pre-computed signature counts instead of original_df
-        remaining_sigs = df['_signature']
-        sig_counts = signature_counts[remaining_sigs]
-        max_signature = sig_counts.idxmax()
-        df = df[df['_signature'] == max_signature]
 
-    df['icao'] = icao
-    df = df.drop(columns=['_non_empty_dict', '_non_empty_count', '_signature'])
-    # Ensure empty strings are preserved, not NaN
-    df[COLUMNS] = df[COLUMNS].fillna('')
-    return df
+def compress_multi_icao_df(df: pl.DataFrame, verbose: bool = True) -> pl.DataFrame:
+    """Compress a DataFrame with multiple ICAOs to one row per ICAO.
+    
+    This is the main entry point for compressing ADS-B data.
+    Used by both daily GitHub Actions runs and historical AWS runs.
+    
+    Args:
+        df: DataFrame with columns ['time', 'icao'] + COLUMNS
+        verbose: Whether to print progress
+    
+    Returns:
+        Compressed DataFrame with one row per ICAO
+    """
+    if df.height == 0:
+        return df
+    
+    # Sort by icao and time
+    df = df.sort(['icao', 'time'])
+    
+    # Fill null values with empty strings for COLUMNS
+    for col in COLUMNS:
+        if col in df.columns:
+            df = df.with_columns(pl.col(col).cast(pl.Utf8).fill_null(""))
+    
+    # First pass: quick deduplication of exact duplicates
+    df = df.unique(subset=['icao'] + COLUMNS, keep='first')
+    if verbose:
+        print(f"After quick dedup: {df.height} records")
+    
+    # Second pass: sophisticated compression per ICAO
+    if verbose:
+        print("Compressing per ICAO...")
+    
+    # Process each ICAO group
+    icao_groups = df.partition_by('icao', as_dict=True, maintain_order=True)
+    compressed_dfs = []
+    
+    for icao_key, group_df in icao_groups.items():
+        # partition_by with as_dict=True returns tuple keys, extract first element
+        icao = icao_key[0] if isinstance(icao_key, tuple) else icao_key
+        compressed = compress_df_polars(group_df, str(icao))
+        compressed_dfs.append(compressed)
+    
+    if compressed_dfs:
+        df_compressed = pl.concat(compressed_dfs)
+    else:
+        df_compressed = df.head(0)  # Empty with same schema
+    
+    if verbose:
+        print(f"After compress: {df_compressed.height} records")
+    
+    # Reorder columns: time first, then icao
+    cols = df_compressed.columns
+    ordered_cols = ['time', 'icao'] + [c for c in cols if c not in ['time', 'icao']]
+    df_compressed = df_compressed.select(ordered_cols)
+    
+    return df_compressed
 
-# names of releases something like
-# planequery_aircraft_adsb_2024-06-01T00-00-00Z.csv.gz
-
-# Let's build historical first. 
 
 def load_raw_adsb_for_day(day):
     """Load raw ADS-B data for a day from parquet file."""
     from datetime import timedelta
     from pathlib import Path
-    import pandas as pd
     
     start_time = day.replace(hour=0, minute=0, second=0, microsecond=0)
     
@@ -84,67 +181,72 @@ def load_raw_adsb_for_day(day):
     
     if parquet_file.exists():
         print(f"  Loading from parquet: {parquet_file}")
-        df = pd.read_parquet(
+        df = pl.read_parquet(
             parquet_file, 
             columns=['time', 'icao', 'r', 't', 'dbFlags', 'ownOp', 'year', 'desc', 'aircraft_category']
         )
         
         # Convert to timezone-naive datetime
-        df['time'] = df['time'].dt.tz_localize(None)
+        if df["time"].dtype == pl.Datetime:
+            df = df.with_columns(pl.col("time").dt.replace_time_zone(None))
+        
         return df
     else:
         # Return empty DataFrame if parquet file doesn't exist
         print(f"  No data available for {start_time.strftime('%Y-%m-%d')}")
-        import pandas as pd
-        return pd.DataFrame(columns=['time', 'icao', 'r', 't', 'dbFlags', 'ownOp', 'year', 'desc', 'aircraft_category'])
+        return pl.DataFrame(schema={
+            'time': pl.Datetime,
+            'icao': pl.Utf8,
+            'r': pl.Utf8,
+            't': pl.Utf8,
+            'dbFlags': pl.Int64,
+            'ownOp': pl.Utf8,
+            'year': pl.Int64,
+            'desc': pl.Utf8,
+            'aircraft_category': pl.Utf8
+        })
+
 
 def load_historical_for_day(day):
-    from pathlib import Path
-    import pandas as pd
+    """Load and compress historical ADS-B data for a day."""
     df = load_raw_adsb_for_day(day)
-    if df.empty:
+    if df.height == 0:
         return df
-    print(f"Loaded {len(df)} raw records for {day.strftime('%Y-%m-%d')}")
-    df = df.sort_values(['icao', 'time'])
-    print("done sort")
-    df[COLUMNS] = df[COLUMNS].fillna('')
     
-    # First pass: quick deduplication of exact duplicates
-    df = df.drop_duplicates(subset=['icao'] + COLUMNS, keep='first')
-    print(f"After quick dedup: {len(df)} records")
+    print(f"Loaded {df.height} raw records for {day.strftime('%Y-%m-%d')}")
     
-    # Second pass: sophisticated compression per ICAO
-    print("Compressing per ICAO...")
-    df_compressed = df.groupby('icao', group_keys=False).apply(compress_df)
-    print(f"After compress: {len(df_compressed)} records")
-    
-    cols = df_compressed.columns.tolist()
-    cols.remove('time')
-    cols.insert(0, 'time')
-    cols.remove("icao")
-    cols.insert(1, "icao")
-    df_compressed = df_compressed[cols]
-    return df_compressed
+    # Use shared compression function
+    return compress_multi_icao_df(df, verbose=True)
 
 
 def concat_compressed_dfs(df_base, df_new):
     """Concatenate base and new compressed dataframes, keeping the most informative row per ICAO."""
-    import pandas as pd
-    
     # Combine both dataframes
-    df_combined = pd.concat([df_base, df_new], ignore_index=True)
+    df_combined = pl.concat([df_base, df_new])
     
     # Sort by ICAO and time
-    df_combined = df_combined.sort_values(['icao', 'time'])
+    df_combined = df_combined.sort(['icao', 'time'])
     
-    # Fill NaN values
-    df_combined[COLUMNS] = df_combined[COLUMNS].fillna('')
+    # Fill null values
+    for col in COLUMNS:
+        if col in df_combined.columns:
+            df_combined = df_combined.with_columns(pl.col(col).fill_null(""))
     
     # Apply compression logic per ICAO to get the best row
-    df_compressed = df_combined.groupby('icao', group_keys=False).apply(compress_df)
+    icao_groups = df_combined.partition_by('icao', as_dict=True, maintain_order=True)
+    compressed_dfs = []
+    
+    for icao, group_df in icao_groups.items():
+        compressed = compress_df_polars(group_df, icao)
+        compressed_dfs.append(compressed)
+    
+    if compressed_dfs:
+        df_compressed = pl.concat(compressed_dfs)
+    else:
+        df_compressed = df_combined.head(0)
     
     # Sort by time
-    df_compressed = df_compressed.sort_values('time')
+    df_compressed = df_compressed.sort('time')
     
     return df_compressed
 
@@ -152,13 +254,15 @@ def concat_compressed_dfs(df_base, df_new):
 def get_latest_aircraft_adsb_csv_df():
     """Download and load the latest ADS-B CSV from GitHub releases."""
     from get_latest_planequery_aircraft_release import download_latest_aircraft_adsb_csv
-    
-    import pandas as pd
     import re
     
     csv_path = download_latest_aircraft_adsb_csv()
-    df = pd.read_csv(csv_path)
-    df = df.fillna("")
+    df = pl.read_csv(csv_path, null_values=[""])
+    
+    # Fill nulls with empty strings
+    for col in df.columns:
+        if df[col].dtype == pl.Utf8:
+            df = df.with_columns(pl.col(col).fill_null(""))
     
     # Extract start date from filename pattern: planequery_aircraft_adsb_{start_date}_{end_date}.csv
     match = re.search(r"planequery_aircraft_adsb_(\d{4}-\d{2}-\d{2})_", str(csv_path))

src/adsb/download_adsb_data_to_parquet.py

@@ -11,6 +11,8 @@ This file is self-contained and does not import from other project modules.
 import gc
 import glob
 import gzip
+import resource
+import shutil
 import sys
 import logging
 import time
@@ -22,10 +24,10 @@ import os
 import argparse
 import datetime as dt
 from datetime import datetime, timedelta, timezone
+import urllib.request
+import urllib.error
 
-import requests
 import orjson
-import pandas as pd
 import pyarrow as pa
 import pyarrow.parquet as pq
 
@@ -44,6 +46,24 @@ TOKEN = os.environ.get('GITHUB_TOKEN')  # Optional: for higher GitHub API rate l
 HEADERS = {"Authorization": f"token {TOKEN}"} if TOKEN else {}
 
 
+def get_resource_usage() -> str:
+    """Get current RAM and disk usage as a formatted string."""
+    # RAM usage (RSS = Resident Set Size)
+    ram_bytes = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss
+    # On macOS, ru_maxrss is in bytes; on Linux, it's in KB
+    if sys.platform == 'darwin':
+        ram_gb = ram_bytes / (1024**3)
+    else:
+        ram_gb = ram_bytes / (1024**2)  # Convert KB to GB
+    
+    # Disk usage
+    disk = shutil.disk_usage('.')
+    disk_free_gb = disk.free / (1024**3)
+    disk_total_gb = disk.total / (1024**3)
+    
+    return f"RAM: {ram_gb:.2f}GB | Disk: {disk_free_gb:.1f}GB free / {disk_total_gb:.1f}GB total"
+
+
 # ============================================================================
 # GitHub Release Fetching and Downloading
 # ============================================================================
@@ -72,17 +92,19 @@ def fetch_releases(version_date: str) -> list:
         
         for attempt in range(1, max_retries + 1):
             try:
-                response = requests.get(f"{BASE_URL}?page={page}", headers=HEADERS)
-                if response.status_code == 200:
-                    break
-                else:
-                    print(f"Failed to fetch releases (attempt {attempt}/{max_retries}): {response.status_code} {response.reason}")
-                    if attempt < max_retries:
-                        print(f"Waiting {retry_delay} seconds before retry...")
-                        time.sleep(retry_delay)
+                req = urllib.request.Request(f"{BASE_URL}?page={page}", headers=HEADERS)
+                with urllib.request.urlopen(req) as response:
+                    if response.status == 200:
+                        data = orjson.loads(response.read())
+                        break
                     else:
-                        print(f"Giving up after {max_retries} attempts")
-                        return releases
+                        print(f"Failed to fetch releases (attempt {attempt}/{max_retries}): {response.status} {response.reason}")
+                        if attempt < max_retries:
+                            print(f"Waiting {retry_delay} seconds before retry...")
+                            time.sleep(retry_delay)
+                        else:
+                            print(f"Giving up after {max_retries} attempts")
+                            return releases
             except Exception as e:
                 print(f"Request exception (attempt {attempt}/{max_retries}): {e}")
                 if attempt < max_retries:
@@ -91,8 +113,6 @@ def fetch_releases(version_date: str) -> list:
                 else:
                     print(f"Giving up after {max_retries} attempts")
                     return releases
-        
-        data = response.json()
         if not data:
             break
         for release in data:
@@ -115,18 +135,22 @@ def download_asset(asset_url: str, file_path: str) -> bool:
         signal.signal(signal.SIGALRM, timeout_handler)
         signal.alarm(40)  # 40-second timeout
         
-        response = requests.get(asset_url, headers=HEADERS, stream=True)
-        signal.alarm(0)
-        
-        if response.status_code == 200:
-            with open(file_path, "wb") as file:
-                for chunk in response.iter_content(chunk_size=8192):
-                    file.write(chunk)
-            print(f"Saved {file_path}")
-            return True
-        else:
-            print(f"Failed to download {asset_url}: {response.status_code} {response.reason}")
-            return False
+        req = urllib.request.Request(asset_url, headers=HEADERS)
+        with urllib.request.urlopen(req) as response:
+            signal.alarm(0)
+            
+            if response.status == 200:
+                with open(file_path, "wb") as file:
+                    while True:
+                        chunk = response.read(8192)
+                        if not chunk:
+                            break
+                        file.write(chunk)
+                print(f"Saved {file_path}")
+                return True
+            else:
+                print(f"Failed to download {asset_url}: {response.status} {response.msg}")
+                return False
     except DownloadTimeoutException as e:
         print(f"Download aborted for {asset_url}: {e}")
         return False
@@ -139,6 +163,7 @@ def extract_split_archive(file_paths: list, extract_dir: str) -> bool:
     """
     Extracts a split archive by concatenating the parts using 'cat'
     and then extracting with 'tar' in one pipeline.
+    Deletes the tar files immediately after extraction to save disk space.
     """
     if os.path.isdir(extract_dir):
         print(f"[SKIP] Extraction directory already exists: {extract_dir}")
@@ -176,6 +201,20 @@ def extract_split_archive(file_paths: list, extract_dir: str) -> bool:
         cat_proc.wait()
         
         print(f"Successfully extracted archive to {extract_dir}")
+        
+        # Delete tar files immediately after extraction
+        for tar_file in file_paths:
+            try:
+                os.remove(tar_file)
+                print(f"Deleted tar file: {tar_file}")
+            except Exception as e:
+                print(f"Failed to delete {tar_file}: {e}")
+        
+        # Check disk usage after deletion
+        disk = shutil.disk_usage('.')
+        free_gb = disk.free / (1024**3)
+        print(f"Disk space after tar deletion: {free_gb:.1f}GB free")
+        
         return True
     except subprocess.CalledProcessError as e:
         print(f"Failed to extract split archive: {e}")
@@ -309,7 +348,7 @@ def process_file(filepath: str) -> list:
             insert_rows.append(inserted_row)
     
     if insert_rows:
-        print(f"Got {len(insert_rows)} rows from {filepath}")
+        # print(f"Got {len(insert_rows)} rows from {filepath}")
         return insert_rows
     else:
         return []
@@ -342,8 +381,8 @@ COLUMNS = [
 
 OS_CPU_COUNT = os.cpu_count() or 1
 MAX_WORKERS = OS_CPU_COUNT if OS_CPU_COUNT > 4 else 1
-CHUNK_SIZE = MAX_WORKERS * 1000
-BATCH_SIZE = (os.cpu_count() or 1) * 100000
+CHUNK_SIZE = MAX_WORKERS * 500  # Reduced for lower RAM usage
+BATCH_SIZE = 250_000  # Fixed size for predictable memory usage (~500MB per batch)
 
 # PyArrow schema for efficient Parquet writing
 PARQUET_SCHEMA = pa.schema([
@@ -448,10 +487,18 @@ def safe_process(fp):
         return []
 
 
-def rows_to_dataframe(rows: list) -> pd.DataFrame:
-    """Convert list of rows to a pandas DataFrame."""
-    df = pd.DataFrame(rows, columns=COLUMNS)
-    return df
+def rows_to_arrow_table(rows: list) -> pa.Table:
+    """Convert list of rows to a PyArrow Table directly (no pandas)."""
+    # Transpose rows into columns
+    columns = list(zip(*rows))
+    
+    # Build arrays for each column according to schema
+    arrays = []
+    for i, field in enumerate(PARQUET_SCHEMA):
+        col_data = list(columns[i]) if i < len(columns) else [None] * len(rows)
+        arrays.append(pa.array(col_data, type=field.type))
+    
+    return pa.Table.from_arrays(arrays, schema=PARQUET_SCHEMA)
 
 
 def write_batch_to_parquet(rows: list, version_date: str, batch_idx: int):
@@ -459,23 +506,17 @@ def write_batch_to_parquet(rows: list, version_date: str, batch_idx: int):
     if not rows:
         return
     
-    df = rows_to_dataframe(rows)
-    
-    # Ensure datetime column is timezone-aware
-    if not df['time'].dt.tz:
-        df['time'] = df['time'].dt.tz_localize('UTC')
+    table = rows_to_arrow_table(rows)
     
     parquet_path = os.path.join(PARQUET_DIR, f"{version_date}_batch_{batch_idx:04d}.parquet")
     
-    # Convert to PyArrow table and write
-    table = pa.Table.from_pandas(df, schema=PARQUET_SCHEMA, preserve_index=False)
     pq.write_table(table, parquet_path, compression='snappy')
     
-    print(f"Written parquet batch {batch_idx} ({len(rows)} rows) to {parquet_path}")
+    print(f"Written parquet batch {batch_idx} ({len(rows)} rows) | {get_resource_usage()}")
 
 
 def merge_parquet_files(version_date: str, delete_batches: bool = True):
-    """Merge all batch parquet files for a version_date into a single file."""
+    """Merge all batch parquet files for a version_date into a single file using streaming."""
     pattern = os.path.join(PARQUET_DIR, f"{version_date}_batch_*.parquet")
     batch_files = sorted(glob.glob(pattern))
     
@@ -483,28 +524,42 @@ def merge_parquet_files(version_date: str, delete_batches: bool = True):
         print(f"No batch files found for {version_date}")
         return None
     
-    print(f"Merging {len(batch_files)} batch files for {version_date}...")
+    print(f"Merging {len(batch_files)} batch files for {version_date} (streaming)...")
     
-    # Read all batch files
-    tables = []
-    for f in batch_files:
-        tables.append(pq.read_table(f))
-    
-    # Concatenate all tables
-    merged_table = pa.concat_tables(tables)
-    
-    # Write merged file
     merged_path = os.path.join(PARQUET_DIR, f"{version_date}.parquet")
-    pq.write_table(merged_table, merged_path, compression='snappy')
+    total_rows = 0
     
-    print(f"Merged parquet file written to {merged_path} ({merged_table.num_rows} total rows)")
+    # Stream write: read one batch at a time to minimize RAM usage
+    writer = None
+    try:
+        for i, f in enumerate(batch_files):
+            table = pq.read_table(f)
+            total_rows += table.num_rows
+            
+            if writer is None:
+                writer = pq.ParquetWriter(merged_path, table.schema, compression='snappy')
+            
+            writer.write_table(table)
+            
+            # Delete batch file immediately after reading to free disk space
+            if delete_batches:
+                os.remove(f)
+            
+            # Free memory
+            del table
+            if (i + 1) % 10 == 0:
+                gc.collect()
+                print(f"  Merged {i + 1}/{len(batch_files)} batches... | {get_resource_usage()}")
+    finally:
+        if writer is not None:
+            writer.close()
+    
+    print(f"Merged parquet file written to {merged_path} ({total_rows} total rows) | {get_resource_usage()}")
     
-    # Optionally delete batch files
     if delete_batches:
-        for f in batch_files:
-            os.remove(f)
-        print(f"Deleted {len(batch_files)} batch files")
+        print(f"Deleted {len(batch_files)} batch files during merge")
     
+    gc.collect()
     return merged_path
 
 
@@ -608,15 +663,15 @@ def process_version_date(version_date: str, keep_folders: bool = False):
     
     print(f"Total rows processed for version_date {version_date}: {total_num_rows}")
     
+    # Clean up extracted directory immediately after processing (before merging parquet files)
+    if not keep_folders and os.path.isdir(extract_dir):
+        print(f"Deleting extraction directory with 100,000+ files: {extract_dir}")
+        shutil.rmtree(extract_dir)
+        print(f"Successfully deleted extraction directory: {extract_dir} | {get_resource_usage()}")
+    
     # Merge batch files into a single parquet file
     merge_parquet_files(version_date, delete_batches=True)
     
-    # Clean up extracted directory if not keeping
-    if not keep_folders and os.path.isdir(extract_dir):
-        import shutil
-        shutil.rmtree(extract_dir)
-        print(f"Cleaned up extraction directory: {extract_dir}")
-    
     return total_num_rows
 
 

src/adsb/download_and_list_icaos.py

@@ -0,0 +1,148 @@
+"""
+Downloads and extracts adsb.lol tar files, then lists all ICAO folders.
+This is the first step of the map-reduce pipeline.
+
+Outputs:
+- Extracted trace files in data/output/{version_date}-planes-readsb-prod-0.tar_0/
+- ICAO manifest at data/output/icao_manifest_{date}.txt
+"""
+import os
+import sys
+import argparse
+import glob
+import subprocess
+from datetime import datetime, timedelta
+
+# Re-use download/extract functions from download_adsb_data_to_parquet
+from src.adsb.download_adsb_data_to_parquet import (
+    OUTPUT_DIR,
+    fetch_releases,
+    download_asset,
+    extract_split_archive,
+    collect_trace_files_with_find,
+)
+
+
+def get_target_day() -> datetime:
+    """Get yesterday's date (the day we're processing)."""
+    # return datetime.utcnow() - timedelta(days=1)
+    return datetime.utcnow() - timedelta(days=1)
+
+
+def download_and_extract(version_date: str) -> str | None:
+    """Download and extract tar files, return extract directory path."""
+    extract_dir = os.path.join(OUTPUT_DIR, f"{version_date}-planes-readsb-prod-0.tar_0")
+    
+    # Check if already extracted
+    if os.path.isdir(extract_dir):
+        print(f"[SKIP] Already extracted: {extract_dir}")
+        return extract_dir
+    
+    # Check for existing tar files
+    pattern = os.path.join(OUTPUT_DIR, f"{version_date}-planes-readsb-prod-0*")
+    matches = [p for p in glob.glob(pattern) if os.path.isfile(p)]
+    
+    if matches:
+        print(f"Found existing tar files for {version_date}")
+        normal_matches = [
+            p for p in matches
+            if "-planes-readsb-prod-0." in os.path.basename(p)
+            and "tmp" not in os.path.basename(p)
+        ]
+        downloaded_files = normal_matches if normal_matches else matches
+    else:
+        # Download from GitHub
+        print(f"Downloading releases for {version_date}...")
+        releases = fetch_releases(version_date)
+        if not releases:
+            print(f"No releases found for {version_date}")
+            return None
+        
+        downloaded_files = []
+        for release in releases:
+            tag_name = release["tag_name"]
+            print(f"Processing release: {tag_name}")
+            
+            assets = release.get("assets", [])
+            normal_assets = [
+                a for a in assets
+                if "planes-readsb-prod-0." in a["name"] and "tmp" not in a["name"]
+            ]
+            tmp_assets = [
+                a for a in assets
+                if "planes-readsb-prod-0tmp" in a["name"]
+            ]
+            use_assets = normal_assets if normal_assets else tmp_assets
+            
+            for asset in use_assets:
+                asset_name = asset["name"]
+                asset_url = asset["browser_download_url"]
+                file_path = os.path.join(OUTPUT_DIR, asset_name)
+                if download_asset(asset_url, file_path):
+                    downloaded_files.append(file_path)
+    
+    if not downloaded_files:
+        print(f"No files downloaded for {version_date}")
+        return None
+    
+    # Extract
+    if extract_split_archive(downloaded_files, extract_dir):
+        return extract_dir
+    return None
+
+
+def list_icao_folders(extract_dir: str) -> list[str]:
+    """List all ICAO folder names from extracted directory."""
+    trace_files = collect_trace_files_with_find(extract_dir)
+    icaos = sorted(trace_files.keys())
+    print(f"Found {len(icaos)} unique ICAOs")
+    return icaos
+
+
+def write_manifest(icaos: list[str], date_str: str) -> str:
+    """Write ICAO list to manifest file."""
+    manifest_path = os.path.join(OUTPUT_DIR, f"icao_manifest_{date_str}.txt")
+    with open(manifest_path, "w") as f:
+        for icao in icaos:
+            f.write(f"{icao}\n")
+    print(f"Wrote manifest with {len(icaos)} ICAOs to {manifest_path}")
+    return manifest_path
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Download and list ICAOs from adsb.lol data")
+    parser.add_argument("--date", type=str, help="Date in YYYY-MM-DD format (default: yesterday)")
+    args = parser.parse_args()
+    
+    if args.date:
+        target_day = datetime.strptime(args.date, "%Y-%m-%d")
+    else:
+        target_day = get_target_day()
+    
+    date_str = target_day.strftime("%Y-%m-%d")
+    version_date = f"v{target_day.strftime('%Y.%m.%d')}"
+    
+    print(f"Processing date: {date_str} (version: {version_date})")
+    
+    # Download and extract
+    extract_dir = download_and_extract(version_date)
+    if not extract_dir:
+        print("Failed to download/extract data")
+        sys.exit(1)
+    
+    # List ICAOs
+    icaos = list_icao_folders(extract_dir)
+    if not icaos:
+        print("No ICAOs found")
+        sys.exit(1)
+    
+    # Write manifest
+    manifest_path = write_manifest(icaos, date_str)
+    
+    print(f"\nDone! Extract dir: {extract_dir}")
+    print(f"Manifest: {manifest_path}")
+    print(f"Total ICAOs: {len(icaos)}")
+
+
+if __name__ == "__main__":
+    main()

src/adsb/process_icao_chunk.py

@@ -0,0 +1,270 @@
+"""
+Processes a chunk of ICAOs from pre-extracted trace files.
+This is the map phase of the map-reduce pipeline.
+
+Expects extract_dir to already exist with trace files.
+Reads ICAO manifest to determine which ICAOs to process based on chunk-id.
+
+Usage:
+    python -m src.adsb.process_icao_chunk --chunk-id 0 --total-chunks 4
+"""
+import gc
+import os
+import sys
+import argparse
+import time
+import concurrent.futures
+from datetime import datetime, timedelta
+
+import pyarrow as pa
+import pyarrow.parquet as pq
+
+from src.adsb.download_adsb_data_to_parquet import (
+    OUTPUT_DIR,
+    PARQUET_DIR,
+    PARQUET_SCHEMA,
+    COLUMNS,
+    MAX_WORKERS,
+    process_file,
+    get_resource_usage,
+    collect_trace_files_with_find,
+)
+
+
+CHUNK_OUTPUT_DIR = os.path.join(OUTPUT_DIR, "adsb_chunks")
+os.makedirs(CHUNK_OUTPUT_DIR, exist_ok=True)
+
+# Smaller batch size for memory efficiency
+BATCH_SIZE = 100_000
+
+
+def get_target_day() -> datetime:
+    """Get yesterday's date (the day we're processing)."""
+    return datetime.utcnow() - timedelta(days=1)
+
+
+def read_manifest(date_str: str) -> list[str]:
+    """Read ICAO manifest file."""
+    manifest_path = os.path.join(OUTPUT_DIR, f"icao_manifest_{date_str}.txt")
+    if not os.path.exists(manifest_path):
+        raise FileNotFoundError(f"Manifest not found: {manifest_path}")
+    
+    with open(manifest_path, "r") as f:
+        icaos = [line.strip() for line in f if line.strip()]
+    return icaos
+
+
+def deterministic_hash(s: str) -> int:
+    """Return a deterministic hash for a string (unlike Python's hash() which is randomized)."""
+    # Use sum of byte values - simple but deterministic
+    return sum(ord(c) for c in s)
+
+
+def get_chunk_icaos(icaos: list[str], chunk_id: int, total_chunks: int) -> list[str]:
+    """Get the subset of ICAOs for this chunk based on deterministic hash partitioning."""
+    return [icao for icao in icaos if deterministic_hash(icao) % total_chunks == chunk_id]
+
+
+def build_trace_file_map(extract_dir: str) -> dict[str, str]:
+    """Build a map of ICAO -> trace file path using find command."""
+    print(f"Building trace file map from {extract_dir}...")
+    
+    # Debug: check what's in extract_dir
+    if os.path.isdir(extract_dir):
+        items = os.listdir(extract_dir)[:10]
+        print(f"First 10 items in extract_dir: {items}")
+        # Check if there are subdirectories
+        for item in items[:3]:
+            subpath = os.path.join(extract_dir, item)
+            if os.path.isdir(subpath):
+                subitems = os.listdir(subpath)[:5]
+                print(f"  Contents of {item}/: {subitems}")
+    
+    trace_map = collect_trace_files_with_find(extract_dir)
+    print(f"Found {len(trace_map)} trace files")
+    
+    if len(trace_map) == 0:
+        # Debug: try manual find
+        import subprocess
+        result = subprocess.run(
+            ['find', extract_dir, '-type', 'f', '-name', 'trace_full_*'],
+            capture_output=True, text=True
+        )
+        print(f"Manual find output (first 500 chars): {result.stdout[:500]}")
+        print(f"Manual find stderr: {result.stderr[:200]}")
+    
+    return trace_map
+
+
+def safe_process(filepath: str) -> list:
+    """Safely process a file, returning empty list on error."""
+    try:
+        return process_file(filepath)
+    except Exception as e:
+        print(f"Error processing {filepath}: {e}")
+        return []
+
+
+def rows_to_table(rows: list) -> pa.Table:
+    """Convert list of rows to PyArrow table."""
+    import pandas as pd
+    df = pd.DataFrame(rows, columns=COLUMNS)
+    if not df['time'].dt.tz:
+        df['time'] = df['time'].dt.tz_localize('UTC')
+    return pa.Table.from_pandas(df, schema=PARQUET_SCHEMA, preserve_index=False)
+
+
+def process_chunk(
+    chunk_id: int,
+    total_chunks: int,
+    trace_map: dict[str, str],
+    icaos: list[str],
+    date_str: str,
+) -> str | None:
+    """Process a chunk of ICAOs and write to parquet."""
+    chunk_icaos = get_chunk_icaos(icaos, chunk_id, total_chunks)
+    print(f"Chunk {chunk_id}/{total_chunks}: Processing {len(chunk_icaos)} ICAOs")
+    
+    if not chunk_icaos:
+        print(f"Chunk {chunk_id}: No ICAOs to process")
+        return None
+    
+    # Get trace file paths from the map
+    trace_files = []
+    for icao in chunk_icaos:
+        if icao in trace_map:
+            trace_files.append(trace_map[icao])
+    
+    print(f"Chunk {chunk_id}: Found {len(trace_files)} trace files")
+    
+    if not trace_files:
+        print(f"Chunk {chunk_id}: No trace files found")
+        return None
+    
+    # Process files and write parquet in batches
+    output_path = os.path.join(CHUNK_OUTPUT_DIR, f"chunk_{chunk_id}_{date_str}.parquet")
+    
+    start_time = time.perf_counter()
+    total_rows = 0
+    batch_rows = []
+    writer = None
+    
+    try:
+        # Process in parallel batches
+        files_per_batch = MAX_WORKERS * 100
+        for offset in range(0, len(trace_files), files_per_batch):
+            batch_files = trace_files[offset:offset + files_per_batch]
+            
+            with concurrent.futures.ProcessPoolExecutor(max_workers=MAX_WORKERS) as executor:
+                for rows in executor.map(safe_process, batch_files):
+                    if rows:
+                        batch_rows.extend(rows)
+                        
+                        # Write when batch is full
+                        if len(batch_rows) >= BATCH_SIZE:
+                            table = rows_to_table(batch_rows)
+                            total_rows += len(batch_rows)
+                            
+                            if writer is None:
+                                writer = pq.ParquetWriter(output_path, PARQUET_SCHEMA, compression='snappy')
+                            writer.write_table(table)
+                            
+                            batch_rows = []
+                            del table
+                            gc.collect()
+                            
+                            elapsed = time.perf_counter() - start_time
+                            print(f"Chunk {chunk_id}: {total_rows} rows, {elapsed:.1f}s | {get_resource_usage()}")
+            
+            gc.collect()
+        
+        # Write remaining rows
+        if batch_rows:
+            table = rows_to_table(batch_rows)
+            total_rows += len(batch_rows)
+            
+            if writer is None:
+                writer = pq.ParquetWriter(output_path, PARQUET_SCHEMA, compression='snappy')
+            writer.write_table(table)
+            del table
+    
+    finally:
+        if writer:
+            writer.close()
+    
+    elapsed = time.perf_counter() - start_time
+    print(f"Chunk {chunk_id}: Done! {total_rows} rows in {elapsed:.1f}s | {get_resource_usage()}")
+    
+    if total_rows > 0:
+        return output_path
+    return None
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Process a chunk of ICAOs")
+    parser.add_argument("--chunk-id", type=int, required=True, help="Chunk ID (0-indexed)")
+    parser.add_argument("--total-chunks", type=int, required=True, help="Total number of chunks")
+    parser.add_argument("--date", type=str, help="Date in YYYY-MM-DD format (default: yesterday)")
+    args = parser.parse_args()
+    
+    if args.date:
+        target_day = datetime.strptime(args.date, "%Y-%m-%d")
+    else:
+        target_day = get_target_day()
+    
+    date_str = target_day.strftime("%Y-%m-%d")
+    version_date = f"v{target_day.strftime('%Y.%m.%d')}"
+    
+    print(f"Processing chunk {args.chunk_id}/{args.total_chunks} for {date_str}")
+    print(f"OUTPUT_DIR: {OUTPUT_DIR}")
+    print(f"CHUNK_OUTPUT_DIR: {CHUNK_OUTPUT_DIR}")
+    print(f"Resource usage at start: {get_resource_usage()}")
+    
+    # Debug: List what's in OUTPUT_DIR
+    print(f"\nContents of {OUTPUT_DIR}:")
+    if os.path.isdir(OUTPUT_DIR):
+        for item in os.listdir(OUTPUT_DIR)[:20]:
+            print(f"  - {item}")
+    else:
+        print(f"  Directory does not exist!")
+    
+    # Find extract directory
+    extract_dir = os.path.join(OUTPUT_DIR, f"{version_date}-planes-readsb-prod-0.tar_0")
+    print(f"\nLooking for extract_dir: {extract_dir}")
+    if not os.path.isdir(extract_dir):
+        print(f"Extract directory not found: {extract_dir}")
+        # Try to find any extracted directory
+        import glob
+        pattern = os.path.join(OUTPUT_DIR, "*-planes-readsb-prod-0*")
+        matches = glob.glob(pattern)
+        print(f"Searching for pattern: {pattern}")
+        print(f"Found matches: {matches}")
+        sys.exit(1)
+    
+    # Build trace file map using find
+    trace_map = build_trace_file_map(extract_dir)
+    if not trace_map:
+        print("No trace files found in extract directory")
+        sys.exit(1)
+    
+    # Read manifest
+    icaos = read_manifest(date_str)
+    print(f"Total ICAOs in manifest: {len(icaos)}")
+    
+    # Process chunk
+    output_path = process_chunk(
+        args.chunk_id,
+        args.total_chunks,
+        trace_map,
+        icaos,
+        date_str,
+    )
+    
+    if output_path:
+        print(f"Output: {output_path}")
+    else:
+        print("No output generated")
+
+
+if __name__ == "__main__":
+    main()

src/adsb/reducer.py

@@ -8,23 +8,15 @@ Environment variables:
   GLOBAL_START_DATE — overall start date for output filename
   GLOBAL_END_DATE   — overall end date for output filename
 """
+import gzip
 import os
+import shutil
 from pathlib import Path
 
 import boto3
-import pandas as pd
+import polars as pl
 
-
-COLUMNS = ["dbFlags", "ownOp", "year", "desc", "aircraft_category", "r", "t"]
-
-
-def deduplicate_by_signature(df: pd.DataFrame) -> pd.DataFrame:
-    """For each icao, keep only the earliest row with each unique signature."""
-    df["_signature"] = df[COLUMNS].astype(str).agg("|".join, axis=1)
-    df_deduped = df.groupby(["icao", "_signature"], as_index=False).first()
-    df_deduped = df_deduped.drop(columns=["_signature"])
-    df_deduped = df_deduped.sort_values("time")
-    return df_deduped
+from compress_adsb_to_aircraft_data import COLUMNS, deduplicate_by_signature
 
 
 def main():
@@ -55,42 +47,50 @@ def main():
     download_dir = Path("/tmp/chunks")
     download_dir.mkdir(parents=True, exist_ok=True)
 
-    df_accumulated = pd.DataFrame()
+    dfs = []
 
     for key in chunk_keys:
-        local_path = download_dir / Path(key).name
+        gz_path = download_dir / Path(key).name
+        csv_path = gz_path.with_suffix("")  # Remove .gz
         print(f"Downloading {key}...")
-        s3.download_file(s3_bucket, key, str(local_path))
+        s3.download_file(s3_bucket, key, str(gz_path))
 
-        df_chunk = pd.read_csv(local_path, compression="gzip", keep_default_na=False)
-        print(f"  Loaded {len(df_chunk)} rows from {local_path.name}")
+        # Decompress
+        with gzip.open(gz_path, 'rb') as f_in:
+            with open(csv_path, 'wb') as f_out:
+                shutil.copyfileobj(f_in, f_out)
+        gz_path.unlink()
 
-        if df_accumulated.empty:
-            df_accumulated = df_chunk
-        else:
-            df_accumulated = pd.concat(
-                [df_accumulated, df_chunk], ignore_index=True
-            )
+        df_chunk = pl.read_csv(csv_path)
+        print(f"  Loaded {df_chunk.height} rows from {csv_path.name}")
+        dfs.append(df_chunk)
 
         # Free disk space after loading
-        local_path.unlink()
+        csv_path.unlink()
 
-    print(f"Combined: {len(df_accumulated)} rows before dedup")
+    df_accumulated = pl.concat(dfs) if dfs else pl.DataFrame()
+    print(f"Combined: {df_accumulated.height} rows before dedup")
 
     # Final global deduplication
     df_accumulated = deduplicate_by_signature(df_accumulated)
-    print(f"After dedup: {len(df_accumulated)} rows")
+    print(f"After dedup: {df_accumulated.height} rows")
 
     # Write and upload final result
     output_name = f"planequery_aircraft_adsb_{global_start}_{global_end}.csv.gz"
-    local_output = Path(f"/tmp/{output_name}")
-    df_accumulated.to_csv(local_output, index=False, compression="gzip")
+    csv_output = Path(f"/tmp/planequery_aircraft_adsb_{global_start}_{global_end}.csv")
+    gz_output = Path(f"/tmp/{output_name}")
+    
+    df_accumulated.write_csv(csv_output)
+    with open(csv_output, 'rb') as f_in:
+        with gzip.open(gz_output, 'wb') as f_out:
+            shutil.copyfileobj(f_in, f_out)
+    csv_output.unlink()
 
     final_key = f"final/{output_name}"
     print(f"Uploading to s3://{s3_bucket}/{final_key}")
-    s3.upload_file(str(local_output), s3_bucket, final_key)
+    s3.upload_file(str(gz_output), s3_bucket, final_key)
 
-    print(f"Final output: {len(df_accumulated)} records -> {final_key}")
+    print(f"Final output: {df_accumulated.height} records -> {final_key}")
 
 
 if __name__ == "__main__":

src/adsb/requirements.reducer.txt

@@ -1,2 +1,2 @@
-pandas>=2.0
+polars>=1.0
 boto3>=1.34

src/adsb/requirements.worker.txt

@@ -1,4 +1,5 @@
-pandas>=2.0
-clickhouse-connect>=0.7
+polars>=1.0
+pyarrow>=14.0
+orjson>=3.9
 boto3>=1.34
 zstandard>=0.22

src/adsb/worker.py

@@ -13,18 +13,13 @@ from datetime import datetime, timedelta
 from pathlib import Path
 
 import boto3
-import pandas as pd
+import polars as pl
 
-from compress_adsb_to_aircraft_data import load_historical_for_day, COLUMNS
-
-
-def deduplicate_by_signature(df: pd.DataFrame) -> pd.DataFrame:
-    """For each icao, keep only the earliest row with each unique signature."""
-    df["_signature"] = df[COLUMNS].astype(str).agg("|".join, axis=1)
-    df_deduped = df.groupby(["icao", "_signature"], as_index=False).first()
-    df_deduped = df_deduped.drop(columns=["_signature"])
-    df_deduped = df_deduped.sort_values("time")
-    return df_deduped
+from compress_adsb_to_aircraft_data import (
+    load_historical_for_day,
+    deduplicate_by_signature,
+    COLUMNS,
+)
 
 
 def main():
@@ -39,28 +34,20 @@ def main():
     total_days = (end_date - start_date).days
     print(f"Worker: processing {total_days} days [{start_date_str}, {end_date_str})")
 
-    df_accumulated = pd.DataFrame()
+    dfs = []
     current_date = start_date
 
     while current_date < end_date:
         day_str = current_date.strftime("%Y-%m-%d")
         print(f"  Loading {day_str}...")
 
-        try:
-            df_compressed = load_historical_for_day(current_date)
-        except Exception as e:
-            print(f"  WARNING: Failed to load {day_str}: {e}")
-            current_date += timedelta(days=1)
-            continue
+        df_compressed = load_historical_for_day(current_date)
+        if df_compressed.height == 0:
+            raise RuntimeError(f"No data found for {day_str}")
 
-        if df_accumulated.empty:
-            df_accumulated = df_compressed
-        else:
-            df_accumulated = pd.concat(
-                [df_accumulated, df_compressed], ignore_index=True
-            )
-
-        print(f"  +{len(df_compressed)} rows (total: {len(df_accumulated)})")
+        dfs.append(df_compressed)
+        total_rows = sum(df.height for df in dfs)
+        print(f"  +{df_compressed.height} rows (total: {total_rows})")
 
         # Delete local cache after each day to save disk in container
         cache_dir = Path("data/adsb")
@@ -70,23 +57,31 @@ def main():
 
         current_date += timedelta(days=1)
 
-    if df_accumulated.empty:
-        print("No data collected — exiting.")
-        return
+    # Concatenate all days
+    df_accumulated = pl.concat(dfs) if dfs else pl.DataFrame()
 
     # Deduplicate within this chunk
     df_accumulated = deduplicate_by_signature(df_accumulated)
-    print(f"After dedup: {len(df_accumulated)} rows")
+    print(f"After dedup: {df_accumulated.height} rows")
 
     # Write to local file then upload to S3
-    local_path = Path(f"/tmp/chunk_{start_date_str}_{end_date_str}.csv.gz")
-    df_accumulated.to_csv(local_path, index=False, compression="gzip")
+    local_path = Path(f"/tmp/chunk_{start_date_str}_{end_date_str}.csv")
+    df_accumulated.write_csv(local_path)
+    
+    # Compress with gzip
+    import gzip
+    import shutil
+    gz_path = Path(f"/tmp/chunk_{start_date_str}_{end_date_str}.csv.gz")
+    with open(local_path, 'rb') as f_in:
+        with gzip.open(gz_path, 'wb') as f_out:
+            shutil.copyfileobj(f_in, f_out)
+    local_path.unlink()  # Remove uncompressed file
 
     s3_key = f"intermediate/{run_id}/chunk_{start_date_str}_{end_date_str}.csv.gz"
     print(f"Uploading to s3://{s3_bucket}/{s3_key}")
 
     s3 = boto3.client("s3")
-    s3.upload_file(str(local_path), s3_bucket, s3_key)
+    s3.upload_file(str(gz_path), s3_bucket, s3_key)
     print("Done.")