Airflow Python 脚本退出 Task 退出并返回代码 -9，如何解决？

Question

I don't know what this error means, some people say it's memory error, I'm not sure, because the error is not explicit, but the table I load is large, 1 million of lines.

Here is the part of my script where the error happens:

# snapshot_profiles
  df_snapshot_profiles = load_table('snapshot_profiles', conn)

  def return_key(x, key):
    try:
      return (x[key])
    except:
      return (None)

  df_snapshot_profiles['is_manager'] = df_snapshot_profiles["payload"].apply(
      lambda x: return_key(x, 'is_manager'))
  df_snapshot_profiles_actual = df_snapshot_profiles.loc[:,
                                                         ['profile_id', 'date']]
  df_snapshot_profiles_actual.sort_values(['profile_id', 'date'], inplace=True)
  df_snapshot_profiles_actual = df_snapshot_profiles_actual.groupby(
      'profile_id').max().reset_index()
  df_snapshot_profiles.drop(
      ['id', 'payload', 'version', 'company_id', 'inserted_at', 'updated_at'],
      axis=1,
      inplace=True)
  df_snapshot_profiles_actual = df_snapshot_profiles_actual.merge(
      df_snapshot_profiles, on=['date', 'profile_id'], how='left')
  df_snapshot_profiles_actual.drop('date', axis=1, inplace=True)

  df = df.merge(df_snapshot_profiles_actual, on='profile_id', how='left')
  del df_snapshot_profiles

  # Excluir do banco empresas com menos de dois usuários (Empresas de testes)
  df_companies = df.groupby('company_name').count()
  df_companies.reset_index(inplace=True)
  df_companies = df_companies[df_companies['user_id'] > 2]
  df_companies.sort_values('user_id', ascending=False)

  companies = list(df_companies.company_name)

  df['check_company'] = df['company_name'].apply(lambda x: 'T'
                                                 if x in companies else 'F')
  df = df[df['check_company'] == 'T']
  df.drop('check_company', axis=1, inplace=True)

And here is the script to load the tables and print the memory usage:

def usage():
  process = psutil.Process(os.getpid())
  return process.memory_info()[0] / float(2**20)


def load_table(table, conn):
    print_x(f'{usage()} Mb')
    print_x(f'loading table {table}')
    cursor = conn.cursor()
    cursor.execute(f'''select * from {ORIGIN_SCHEMA}.{table};''')
    df = cursor.fetchall()
    cursor.execute(f'''
        select column_name from information_schema.columns where table_name = '{table}';
    ''')
    labels = cursor.fetchall()
    label_list = []
    for label in labels:
        label_list.append(label[0])
    df = pd.DataFrame.from_records(df, columns=label_list)
    return (df)

Is there a way to avoid the error, by reducing the memory usage or other way??

Answer 1

Well. It should be out of memory issue. You can expand you memory or switch part of work out of core(load work in batch mode)

• If you has budget, expand memory. 1 million line * terrible string length(1000) per column =1M*1K = 1G memory for data load. when merge dataframe, or transform dataframe, you need extra memory, so 16G should be OK.

• if you are expert, try out of core mode, that mean work on the hard disk.

  • dask is one of pandas out of core module. computer on batch mode. slow but still work.
  • use database for some feature work. I found most of database can do similar work like pandas though complicated SQL code needed.

Good luck. If you like my answer, plse vote it up.

Answer 2

I solved this issue by implementing a server side cursor, and getting info in chunks, like so:

  serverCursor = conn.cursor("serverCursor")
  serverCursor.execute(f'''select * from {ORIGIN_SCHEMA}.{table};''')

  df = []
  while True:
    records = serverCursor.fetchmany(size=50000)
    df = df + records
    if not records:
      break
  serverCursor.close()