bin/dangling-references.py

#!/usr/bin/env nix-shell
#!nix-shell -i python3 -p "python3.withPackages (ps: [ps.pyarrow])"
"""
Dangling References Analysis

Usage: ./bin/dangling-references.py
"""

import pyarrow.parquet as pq
import pyarrow as pa
from pathlib import Path
import time
import sys
import multiprocessing as mp

NUM_WORKERS = 7
BATCH_SIZE = 50_000
WRITE_BATCH_SIZE = 10_000  # Write to parquet every N results

def process_row_groups(args):
    """Stream through assigned row groups, send results to queue."""
    main_file, row_group_indices, deletion_set, column_names, result_queue = args

    processed = 0
    found = 0

    parquet_file = pq.ParquetFile(main_file)

    for batch in parquet_file.iter_batches(
        batch_size=BATCH_SIZE,
        row_groups=row_group_indices
    ):
        hashes = batch.column("hash").to_pylist()
        refs_list = batch.column("references").to_pylist()

        # Build indices of rows with dangling refs
        matching_indices = []
        dangling_refs_list = []
        for i, (row_hash, refs) in enumerate(zip(hashes, refs_list)):
            if row_hash in deletion_set:
                continue
            if not refs:
                continue

            dangling_refs = [ref for ref in refs if ref in deletion_set]
            if dangling_refs:
                matching_indices.append(i)
                dangling_refs_list.append(dangling_refs)

        if matching_indices:
            # Extract matching rows as dict of columns
            row_data = {}
            for col in column_names:
                values = [batch.column(col)[i].as_py() for i in matching_indices]
                # Decode binary to string for system and ca_algo
                if col in ("system", "ca_algo"):
                    values = [v.decode("utf-8") if v else None for v in values]
                row_data[col] = values
            row_data["dangling_refs"] = dangling_refs_list
            result_queue.put(row_data)
            found += len(matching_indices)

        processed += len(hashes)

    return processed, found


def writer_process(result_queue, output_file, output_schema, num_workers):
    """Receive results from queue and write to parquet incrementally."""
    writer = None
    total_written = 0
    buffer = None  # Dict of column_name -> list of values
    buffer_len = 0
    workers_done = 0

    while workers_done < num_workers:
        item = result_queue.get()

        if item is None:
            # Sentinel: one worker finished
            workers_done += 1
            continue

        # item is a dict of column_name -> list of values
        if buffer is None:
            buffer = {col: [] for col in item.keys()}

        for col, values in item.items():
            buffer[col].extend(values)
        buffer_len += len(item["hash"])

        # Write when buffer is large enough
        if buffer_len >= WRITE_BATCH_SIZE:
            table = pa.Table.from_pydict(buffer, schema=output_schema)
            if writer is None:
                writer = pq.ParquetWriter(output_file, output_schema, compression="zstd")
            writer.write_table(table)
            total_written += buffer_len
            buffer = None
            buffer_len = 0

    # Write remaining buffer
    if buffer and buffer_len > 0:
        table = pa.Table.from_pydict(buffer, schema=output_schema)
        if writer is None:
            writer = pq.ParquetWriter(output_file, output_schema, compression="zstd")
        writer.write_table(table)
        total_written += buffer_len

    if writer:
        writer.close()
    else:
        # No results - write empty file
        pq.write_table(pa.table({}, schema=output_schema), output_file)

    return total_written


def worker_wrapper(args):
    """Wrapper that sends sentinel when done."""
    main_file, row_group_indices, deletion_set, column_names, result_queue = args
    processed, found = process_row_groups(args)
    result_queue.put(None)  # Sentinel
    return processed, found


def main():
    base = Path("datasets")
    main_file = base / "narinfos-2026-01-06T01-13Z.parquet"
    deletion_file = base / "narinfos-nixos-images-2026-01-06T01-13Z.parquet"
    dangling_refs_file = base / "narinfos-nixos-images-dangling-refs-2026-01-06T01-13Z.parquet"
    output_file = base / "narinfos-nixos-images-dangling-refs-with-parents-2026-01-06T01-13Z.parquet"

    # Step 1: Load deletion hashes from both files
    print("Loading deletion set...")
    start = time.time()
    deletion_table = pq.read_table(deletion_file, columns=["hash"])
    deletion_hashes = set(h.as_py() for h in deletion_table.column("hash"))
    del deletion_table
    print(f"  Loaded {len(deletion_hashes):,} hashes from nixos-images")

    dangling_table = pq.read_table(dangling_refs_file, columns=["hash"])
    dangling_hashes = set(h.as_py() for h in dangling_table.column("hash"))
    del dangling_table
    print(f"  Loaded {len(dangling_hashes):,} hashes from dangling-refs")

    deletion_set = frozenset(deletion_hashes | dangling_hashes)
    del deletion_hashes, dangling_hashes
    print(f"  Total deletion set: {len(deletion_set):,} hashes in {time.time()-start:.1f}s")

    # Step 2: Get row group info and schema
    parquet_file = pq.ParquetFile(main_file)
    num_row_groups = parquet_file.metadata.num_row_groups
    total_rows = parquet_file.metadata.num_rows
    input_schema = parquet_file.schema_arrow
    column_names = [field.name for field in input_schema]
    print(f"  Total: {total_rows:,} rows in {num_row_groups} row groups")

    # Build output schema: input columns + dangling_refs
    # Convert system and ca_algo from binary to string for better compatibility
    output_fields = []
    for field in input_schema:
        if field.name in ("system", "ca_algo"):
            output_fields.append(pa.field(field.name, pa.string()))
        else:
            output_fields.append(field)
    output_fields.append(pa.field("dangling_refs", pa.list_(pa.binary())))
    output_schema = pa.schema(output_fields)

    # Distribute row groups across workers
    workers = min(NUM_WORKERS, num_row_groups)
    row_groups_per_worker = [[] for _ in range(workers)]
    for i in range(num_row_groups):
        row_groups_per_worker[i % workers].append(i)

    # Create queue for results
    manager = mp.Manager()
    result_queue = manager.Queue()

    # Start writer process
    writer_proc = mp.Process(
        target=writer_process,
        args=(result_queue, str(output_file), output_schema, workers)
    )
    writer_proc.start()

    # Prepare worker args
    worker_args = [
        (str(main_file), rgs, deletion_set, column_names, result_queue)
        for rgs in row_groups_per_worker
    ]

    # Step 3: Process in parallel
    print(f"Processing with {workers} workers (batch size {BATCH_SIZE:,})...")
    print(f"  Streaming results to {output_file}")
    sys.stdout.flush()
    start = time.time()

    with mp.Pool(workers) as pool:
        worker_results = pool.map(worker_wrapper, worker_args)

    # Wait for writer to finish
    writer_proc.join()

    # Aggregate stats
    total_processed = sum(r[0] for r in worker_results)
    total_found = sum(r[1] for r in worker_results)

    elapsed = time.time() - start
    print(f"  Completed in {elapsed:.1f}s ({elapsed/60:.1f} min)")
    print(f"  Rate: {total_processed/elapsed:,.0f} rows/sec")

    # Summary
    print("\n" + "="*50)
    print("SUMMARY")
    print("="*50)
    print(f"Entries with dangling refs: {total_found:,}")

    if total_found > 0:
        # Read back sample for display
        result_table = pq.read_table(output_file)

        # Sample up to 50 entries evenly across the output
        num_rows = len(result_table)
        sample_count = min(50, num_rows)
        step = max(1, num_rows // sample_count)
        sample_indices = [i * step for i in range(sample_count) if i * step < num_rows]

        print(f"\nSample pnames with dangling refs ({len(sample_indices)} of {num_rows:,}):")
        for i in sample_indices:
            pname = result_table.column("pname")[i].as_py()
            num_refs = len(result_table.column("dangling_refs")[i].as_py())
            print(f"  {pname}: {num_refs} dangling ref(s)")
    else:
        print("No dangling references found")


if __name__ == "__main__":
    main()