import os import time import importlib from typing import List import torch import torch.distributed as dist import json import multiprocessing as mp from types import ModuleType import pyarrow.fs as pf from concurrent.futures import ThreadPoolExecutor, as_completed from dataloader.hdfs_io import hisdir, hlist_files def is_bitwise_ckpt_enable(): return os.getenv("SAHARA_ENABLE_BITWISE_CKPT", '1') == '1' _HADOOP_COMMAND_TEMPLATE = 'hadoop fs {command}' NATIVE_LIBHDFS_FOLDER = "/opt/tiger/native_libhdfs" if os.path.isdir(NATIVE_LIBHDFS_FOLDER): NATIVE_HDFS_FOLDER = NATIVE_LIBHDFS_FOLDER else: NATIVE_HDFS_FOLDER = os.path.realpath(os.path.join(os.path.dirname(os.path.realpath(__file__)), "../3rdparty/native_dfs_client")) try: with open('/etc/os-release', 'r') as f: os_release = f.read() if 'VERSION_ID="11' in os_release: NATIVE_HDFS_FOLDER = os.path.realpath(os.path.join(os.path.dirname(os.path.realpath(__file__)), "../3rdparty/native_dfs_client_debian11")) except Exception as e: print(f"unable to update NATIVE_HDFS_FOLDER with exception: {e}. use default.") NATIVE_HDFS_PATH = os.getenv("NATIVE_HDFS_PATH", NATIVE_HDFS_FOLDER) if os.getenv("CRUISE_LOCAL_CACHE_DIR", None): CRUISE_LOADER_WS = os.getenv("CRUISE_LOCAL_CACHE_DIR") elif os.getenv("ARNOLD_TRIAL_ID", None): CRUISE_LOADER_WS = "/opt/tiger/cruise_loader_ws" else: CRUISE_LOADER_WS = None def get_fname_from_url(url): if CRUISE_LOADER_WS is None: return None # try to create work space if not os.path.exists(CRUISE_LOADER_WS): try: os.makedirs(CRUISE_LOADER_WS) except: pass return "{}/{}".format(CRUISE_LOADER_WS, url.split(":")[-1].replace("/", "_")) def acquire_file_lock(url): lock_file = url + '.lock_file' try: # Open a file fd = os.open(lock_file, os.O_CREAT | os.O_EXCL) # Close opened file os.close(fd) return True except: return False def get_parquet_row_group_info_from_meta(parquet_file): meta = parquet_file.metadata rg = meta.num_row_groups group_sizes = [meta.row_group(i).num_rows for i in range(rg)] return group_sizes def use_native_hdfs(): return os.path.exists(NATIVE_HDFS_PATH) and int(os.getenv("USE_NATIVE_HDFS_CLIENT", "1")) > 0 def set_native_hdfs_security_permission(): os.environ["INFSEC_HADOOP_ENABLED"] = "1" os.environ["INFSEC_HADDOP_ENABLED"] = "1" def native_hdfs_check(): if use_native_hdfs(): set_native_hdfs_security_permission() def _get_hdfs_command(command): """Return hadoop fs command""" return _HADOOP_COMMAND_TEMPLATE.format(command=command) def get_hdfs_host(): arnold_base_dir = os.environ.get('ARNOLD_BASE_DIR', '') if arnold_base_dir.startswith('hdfs://harunava'): return 'hdfs://harunava' elif arnold_base_dir.startswith('hdfs://harunaoci'): return 'hdfs://harunaoci' elif arnold_base_dir.startswith('hdfs://harunacompass'): return 'hdfs://harunacompass' elif arnold_base_dir.startswith('hdfs://haruna'): return 'hdfs://haruna' elif os.environ.get('ARNOLD_WORKSPACE_CLUSTER_NAME') == 'candy-maliva': return 'hdfs://harunava' try: import xml.etree.ElementTree as ET tree = ET.parse('/opt/tiger/yarn_deploy/hadoop/conf/core-site.xml') root = tree.getroot() for child in root: if child.tag == 'property' and child[0].text == 'fs.defaultFS': return child[1].text except: return 'hdfs://haruna' def get_hdfs_block_size(): try: import xml.etree.ElementTree as ET tree = ET.parse('/opt/tiger/yarn_deploy/hadoop/conf/hdfs-site.xml') root = tree.getroot() for child in root: if child.tag == 'property' and child[0].text == 'dfs.block.size': return int(child[1].text) except: pass return 134217728 def get_hdfs_extra_conf(): hdfs_celer = os.environ.get("ARNOLD_HDFS_CELER", "false") hdfs_celer = hdfs_celer.lower() in ('y', 'yes', 't', 'true', 'on', '1') if hdfs_celer: try: import xml.etree.ElementTree as ET tree = ET.parse("/opt/tiger/arnold/hdfs_client/conf/celer.xml") conf = {} for prop in tree.getroot(): if prop.tag != "property": continue key, val = None, None for elem in prop: if elem.tag == "name": key = elem.text elif elem.tag == "value": val = elem.text if key is None or val is None: continue conf[key] = val return conf except Exception as e: print(f"fail to parse celer conf: {e}. nothing changed.") pass return None def init_arrow_hdfs_fs(): return pf.HadoopFileSystem( host=get_hdfs_host(), port=0, buffer_size=get_hdfs_block_size(), extra_conf=get_hdfs_extra_conf(), ) def get_parquet_file_handle(url): pq = LazyLoader('pq', globals(), 'pyarrow.parquet') pf = LazyLoader('pf', globals(), 'pyarrow.fs') if url.startswith("hdfs"): fs = init_arrow_hdfs_fs() f = fs.open_input_file(url) else: f = open(url, 'rb') fs = pf.LocalFileSystem() parquet_file = pq.ParquetFile(f) return parquet_file, fs, f # process_file: read parquet metadata def process_file(file_path: str): pq = LazyLoader('pq', globals(), 'pyarrow.parquet') try: num_rows = pq.read_metadata(file_path).num_rows return (file_path, num_rows) except Exception as e: print(f"Error processing {file_path}: {e}") return None def build_dataset(total_files: List[str], num_worker: int = 1): data_map = {} data_list = [] if dist.get_rank() != 0: print("Not rank 0, skipping query to prevent overwhelming HDFS hit.") data_list = [None for i in total_files] else: print(f"Total files num: {len(total_files)}") with ThreadPoolExecutor(max_workers=num_worker) as executor: futures = {executor.submit(process_file, file): file for file in total_files} for future in as_completed(futures): result = future.result() if result: file_name_i, num_rows = result file_name = file_name_i.split('/')[-1] data_map[file_name] = num_rows print("Rank 0 finished query file length ..") data_list = [v for k, v in data_map.items()] dist.broadcast_object_list(data_list, src=0) return data_list def get_single_parquet_length(url): if url.startswith('hdfs') and mp.current_process().name == 'MainProcess': print('use pyarrow fs hdfs api in main process may have fork issue!') parquet_file, _, handle = get_parquet_file_handle(url) rows = parquet_file.metadata.num_rows # close file handle handle.close() return rows def get_worker_info(): local_worker_id = 0 local_num_workers = 1 worker_info = torch.utils.data.get_worker_info() if worker_info is not None: local_worker_id = worker_info.id local_num_workers = worker_info.num_workers return local_worker_id, local_num_workers class LazyLoader(ModuleType): """Lazily import a module, mainly to avoid pulling in large dependencies. `contrib`, and `ffmpeg` are examples of modules that are large and not always needed, and this allows them to only be loaded when they are used. """ # The lint error here is incorrect. def __init__(self, local_name, parent_module_globals, name): # pylint: disable=super-on-old-class self._local_name = local_name self._parent_module_globals = parent_module_globals super(LazyLoader, self).__init__(name) def _load(self): # Import the target module and insert it into the parent's namespace module = importlib.import_module(self.__name__) self._parent_module_globals[self._local_name] = module # Update this object's dict so that if someone keeps a reference to the # LazyLoader, lookups are efficient (__getattr__ is only called on lookups # that fail). self.__dict__.update(module.__dict__) return module def __getattr__(self, item): module = self._load() return getattr(module, item) def __dir__(self): module = self._load() return dir(module) class PerfTimer: """Perf timer for a region. If the duration is longer than the threshold, [PERF WARN] will be printed Args: name (str): the name of the current region threshold (float): the warning threshold in seconds verbose (bool): whether to log when the duration is under the threshold """ def __init__(self, name: str, threshold: float = 10, verbose: bool = True): self.t0 = 0 self.t1 = 0 self.name = name self.threshold = threshold self.verbose = verbose self.logged = False self.above_threshold = False def __enter__(self): self.t0 = time.time() def __exit__(self, exc_type, exc_value, exc_tb): self.t1 = time.time() duration = self.t1 - self.t0 msg = '' if duration > self.threshold: msg += '[PERF WARN] ' self.above_threshold = True