import functools import itertools import json import logging import os import queue import random import sys import threading import time from types import TracebackType from typing import ( TYPE_CHECKING, Any, Callable, Dict, List, NamedTuple, Optional, Set, Tuple, Type, Union, ) if TYPE_CHECKING: from typing import TypedDict class ProcessedChunk(TypedDict): offset: int content: List[str] class ProcessedBinaryChunk(TypedDict): offset: int content: str encoding: str import requests import wandb from wandb import util from wandb.sdk.internal import internal_api from ..lib import file_stream_utils logger = logging.getLogger(__name__) class Chunk(NamedTuple): filename: str data: str class DefaultFilePolicy: def __init__(self, start_chunk_id: int = 0) -> None: self._chunk_id = start_chunk_id self.has_debug_log = False def process_chunks( self, chunks: List[Chunk] ) -> Union[bool, "ProcessedChunk", "ProcessedBinaryChunk", List["ProcessedChunk"]]: chunk_id = self._chunk_id self._chunk_id += len(chunks) return {"offset": chunk_id, "content": [c.data for c in chunks]} # TODO: this is very inefficient, this is meant for temporary debugging and will be removed in future releases def _debug_log(self, data: Any): if self.has_debug_log or not os.environ.get("WANDB_DEBUG_FILESTREAM_LOG"): return loaded = json.loads(data) if not isinstance(loaded, dict): return # get key size and convert to MB key_sizes = [(k, len(json.dumps(v))) for k, v in loaded.items()] key_msg = [f"{k}: {v / 1048576:.5f} MB" for k, v in key_sizes] wandb.termerror(f"Step: {loaded['_step']} | {key_msg}", repeat=False) self.has_debug_log = True class JsonlFilePolicy(DefaultFilePolicy): def process_chunks(self, chunks: List[Chunk]) -> "ProcessedChunk": chunk_id = self._chunk_id # TODO: chunk_id is getting reset on each request... self._chunk_id += len(chunks) chunk_data = [] for chunk in chunks: if len(chunk.data) > util.MAX_LINE_BYTES: msg = f"Metric data exceeds maximum size of {util.to_human_size(util.MAX_LINE_BYTES)} ({util.to_human_size(len(chunk.data))})" wandb.termerror(msg, repeat=False) wandb._sentry.message(msg, repeat=False) self._debug_log(chunk.data) else: chunk_data.append(chunk.data) return { "offset": chunk_id, "content": chunk_data, } class SummaryFilePolicy(DefaultFilePolicy): def process_chunks(self, chunks: List[Chunk]) -> Union[bool, "ProcessedChunk"]: data = chunks[-1].data if len(data) > util.MAX_LINE_BYTES: msg = f"Summary data exceeds maximum size of {util.to_human_size(util.MAX_LINE_BYTES)}. Dropping it." wandb.termerror(msg, repeat=False) wandb._sentry.message(msg, repeat=False) self._debug_log(data) return False return {"offset": 0, "content": [data]} class StreamCRState: r"""Stream state that tracks carriage returns. There are two streams: stdout and stderr. We create two instances for each stream. An instance holds state about: found_cr: if a carriage return has been found in this stream. cr: most recent offset (line number) where we found \r. We update this offset with every progress bar update. last_normal: most recent offset without a \r in this stream. i.e. the most recent "normal" line. """ found_cr: bool cr: Optional[int] last_normal: Optional[int] def __init__(self) -> None: self.found_cr = False self.cr = None self.last_normal = None class CRDedupeFilePolicy(DefaultFilePolicy): r"""File stream policy for removing carriage-return erased characters. This is what a terminal does. We use it for console output to reduce the amount of data we need to send over the network (eg. for progress bars), while preserving the output's appearance in the web app. CR stands for "carriage return", for the character \r. It tells the terminal to move the cursor back to the start of the current line. Progress bars (like tqdm) use \r repeatedly to overwrite a line with newer updates. This gives the illusion of the progress bar filling up in real-time. """ def __init__(self, start_chunk_id: int = 0) -> None: super().__init__(start_chunk_id=start_chunk_id) self._prev_chunk = None self.global_offset = 0 # cr refers to carriage return \r self.stderr = StreamCRState() self.stdout = StreamCRState() @staticmethod def get_consecutive_offsets(console: Dict[int, str]) -> List[List[int]]: """Compress consecutive line numbers into an interval. Args: console: Dict[int, str] which maps offsets (line numbers) to lines of text. It represents a mini version of our console dashboard on the UI. Returns: A list of intervals (we compress consecutive line numbers into an interval). Example: >>> console = {2: "", 3: "", 4: "", 5: "", 10: "", 11: "", 20: ""} >>> get_consecutive_offsets(console) [(2, 5), (10, 11), (20, 20)] """ offsets = sorted(list(console.keys())) intervals: List = [] for i, num in enumerate(offsets): if i == 0: intervals.append([num, num]) continue largest = intervals[-1][1] if num == largest + 1: intervals[-1][1] = num else: intervals.append([num, num]) return intervals @staticmethod def split_chunk(chunk: Chunk) -> Tuple[str, str]: r"""Split chunks. Args: chunk: object with two fields: filename (str) & data (str) `chunk.data` is a str containing the lines we want. It usually contains \n or \r or both. `chunk.data` has two possible formats (for the two streams - stdout and stderr): - "2020-08-25T20:38:36.895321 this is my line of text\nsecond line\n" - "ERROR 2020-08-25T20:38:36.895321 this is my line of text\nsecond line\nthird\n". Here's another example with a carriage return \r. - "ERROR 2020-08-25T20:38:36.895321 \r progress bar\n" Returns: A 2-tuple of strings. First str is prefix, either "ERROR {timestamp} " or "{timestamp} ". Second str is the rest of the string. Example: >>> chunk = Chunk( ... filename="output.log", ... data="ERROR 2020-08-25T20:38 this is my line of text\n", ... ) >>> split_chunk(chunk) ("ERROR 2020-08-25T20:38 ", "this is my line of text\n") """ prefix = "" token, rest = chunk.data.split(" ", 1) if token == "ERROR": prefix += token + " " token, rest = rest.split(" ", 1) prefix += token + " " return prefix, rest def process_chunks(self, chunks: List[Chunk]) -> List["ProcessedChunk"]: r"""Process chunks. Args: chunks: List of Chunk objects. See description of chunk above in `split_chunk(...)`. Returns: List[Dict]. Each dict in the list contains two keys: an `offset` which holds the line number and `content` which maps to a list of consecutive lines starting from that offset. `offset` here means global line number in our console on the UI. Example: >>> chunks = [ Chunk("output.log", "ERROR 2020-08-25T20:38 this is my line of text\nboom\n"), Chunk("output.log", "2020-08-25T20:38 this is test\n"), ] >>> process_chunks(chunks) [ {"offset": 0, "content": [ "ERROR 2020-08-25T20:38 this is my line of text\n", "ERROR 2020-08-25T20:38 boom\n", "2020-08-25T20:38 this is test\n" ] } ] """ # Dict[int->str], each offset (line number) mapped to a line. # Represents a mini-version of our console pane on the UI. console = {} sep = os.linesep for c in chunks: prefix, logs_str = self.split_chunk(c) logs = logs_str.split(sep) for line in logs: stream = self.stderr if prefix.startswith("ERROR ") else self.stdout if line.startswith("\r"): # line starting with \r will always overwrite a previous offset. offset: int = ( stream.cr if (stream.found_cr and stream.cr is not None) else (stream.last_normal or 0) ) stream.cr = offset stream.found_cr = True console[offset] = prefix + line[1:] + "\n" # Usually logs_str = "\r progress bar\n" for progress bar updates. # If instead logs_str = "\r progress bar\n text\n text\n", # treat this as the end of a progress bar and reset accordingly. if ( logs_str.count(sep) > 1 and logs_str.replace(sep, "").count("\r") == 1 ): stream.found_cr = False elif line: console[self.global_offset] = prefix + line + "\n" stream.last_normal = self.global_offset self.global_offset += 1 intervals = self.get_consecutive_offsets(console) ret = [] for a, b in intervals: processed_chunk: ProcessedChunk = { "offset": self._chunk_id + a, "content": [console[i] for i in range(a, b + 1)], } ret.append(processed_chunk) return ret class FileStreamApi: """Pushes chunks of files to our streaming endpoint. This class is used as a singleton. It has a thread that serializes access to the streaming endpoint and performs rate-limiting and batching. TODO: Differentiate between binary/text encoding. """ class Finish(NamedTuple): exitcode: int class Preempting(NamedTuple): pass class PushSuccess(NamedTuple): artifact_id: str save_name: str MAX_ITEMS_PER_PUSH = 10000 def __init__( self, api: "internal_api.Api", run_id: str, start_time: float, timeout: float = 0, settings: Optional[dict] = None, ) -> None: settings = settings or dict() # NOTE: exc_info is set in thread_except_body context and readable by calling threads self._exc_info: Optional[ Union[ Tuple[Type[BaseException], BaseException, TracebackType], Tuple[None, None, None], ] ] = None self._settings = settings self._api = api self._run_id = run_id self._start_time = start_time self._client = requests.Session() timeout = timeout or 0 if timeout > 0: self._client.post = functools.partial(self._client.post, timeout=timeout) # type: ignore[method-assign] self._client.auth = api.client.transport.session.auth self._client.headers.update(api.client.transport.headers or {}) self._client.cookies.update(api.client.transport.cookies or {}) # type: ignore[no-untyped-call] self._client.proxies.update(api.client.transport.session.proxies or {}) self._file_policies: Dict[str, DefaultFilePolicy] = {} self._dropped_chunks: int = 0 self._queue: queue.Queue = queue.Queue() self._thread = threading.Thread(target=self._thread_except_body) # It seems we need to make this a daemon thread to get sync.py's atexit handler to run, which # cleans this thread up. self._thread.name = "FileStreamThread" self._thread.daemon = True self._init_endpoint() def _init_endpoint(self) -> None: settings = self._api.settings() settings.update(self._settings) self._endpoint = "{base}/files/{entity}/{project}/{run}/file_stream".format( base=settings["base_url"], entity=settings["entity"], project=settings["project"], run=self._run_id, ) def start(self) -> None: self._init_endpoint() self._thread.start() def set_default_file_policy( self, filename: str, file_policy: "DefaultFilePolicy" ) -> None: """Set an upload policy for a file unless one has already been set.""" if filename not in self._file_policies: self._file_policies[filename] = file_policy def set_file_policy(self, filename: str, file_policy: "DefaultFilePolicy") -> None: self._file_policies[filename] = file_policy @property def heartbeat_seconds(self) -> Union[int, float]: # Defaults to 30 heartbeat_seconds: Union[int, float] = self._api.dynamic_settings[ "heartbeat_seconds" ] return heartbeat_seconds def rate_limit_seconds(self) -> Union[int, float]: run_time = time.time() - self._start_time if run_time < 60: return max(1.0, self.heartbeat_seconds / 15) elif run_time < 300: return max(2.5, self.heartbeat_seconds / 3) else: return max(5.0, self.heartbeat_seconds) def _read_queue(self) -> List: # called from the push thread (_thread_body), this does an initial read # that'll block for up to rate_limit_seconds. Then it tries to read # as much out of the queue as it can. We do this because the http post # to the server happens within _thread_body, and can take longer than # our rate limit. So next time we get a chance to read the queue we want # read all the stuff that queue'd up since last time. # # If we have more than MAX_ITEMS_PER_PUSH in the queue then the push thread # will get behind and data will buffer up in the queue. return util.read_many_from_queue( self._queue, self.MAX_ITEMS_PER_PUSH, self.rate_limit_seconds() ) def _thread_body(self) -> None: posted_data_time = time.time() posted_anything_time = time.time() ready_chunks = [] uploaded: Set[str] = set() finished: Optional[FileStreamApi.Finish] = None while finished is None: items = self._read_queue() for item in items: if isinstance(item, self.Finish): finished = item elif isinstance(item, self.Preempting): request_with_retry( self._client.post, self._endpoint, json={ "complete": False, "preempting": True, "dropped": self._dropped_chunks, "uploaded": list(uploaded), }, ) uploaded = set() elif isinstance(item, self.PushSuccess): uploaded.add(item.save_name) else: # item is Chunk ready_chunks.append(item) cur_time = time.time() if ready_chunks and ( finished or cur_time - posted_data_time > self.rate_limit_seconds() ): posted_data_time = cur_time posted_anything_time = cur_time success = self._send(ready_chunks, uploaded=uploaded) ready_chunks = [] if success: uploaded = set() # If there aren't ready chunks or uploaded files, we still want to # send regular heartbeats so the backend doesn't erroneously mark this # run as crashed. if cur_time - posted_anything_time > self.heartbeat_seconds: posted_anything_time = cur_time # If we encountered an error trying to publish the # list of uploaded files, don't reset the `uploaded` # list. Retry publishing the list on the next attempt. if not isinstance( request_with_retry( self._client.post, self._endpoint, json={ "complete": False, "failed": False, "dropped": self._dropped_chunks, "uploaded": list(uploaded), }, ), Exception, ): uploaded = set() # post the final close message. (item is self.Finish instance now) request_with_retry( self._client.post, self._endpoint, json={ "complete": True, "exitcode": int(finished.exitcode), "dropped": self._dropped_chunks, "uploaded": list(uploaded), }, ) def _thread_except_body(self) -> None: # TODO: Consolidate with internal_util.ExceptionThread try: self._thread_body() except Exception: exc_info = sys.exc_info() self._exc_info = exc_info logger.exception("generic exception in filestream thread") wandb._sentry.exception(exc_info) raise def _handle_response(self, response: Union[Exception, "requests.Response"]) -> None: """Log dropped chunks and updates dynamic settings.""" if isinstance(response, Exception): wandb.termerror( "Dropped streaming file chunk (see wandb/debug-internal.log)" ) logger.exception(f"dropped chunk {response}") self._dropped_chunks += 1 else: parsed: Optional[dict] = None try: parsed = response.json() except Exception: pass if isinstance(parsed, dict): limits = parsed.get("limits") if isinstance(limits, dict): self._api.dynamic_settings.update(limits) def _send(self, chunks: List[Chunk], uploaded: Optional[Set[str]] = None) -> bool: uploaded_list = list(uploaded or []) # create files dict. dict of pairs where chunks are a list of # [chunk_id, chunk_data] tuples (as lists since this will be json). files = {} # Groupby needs group keys to be consecutive, so sort first. chunks.sort(key=lambda c: c.filename) for filename, file_chunks in itertools.groupby(chunks, lambda c: c.filename): file_chunks_list = list(file_chunks) # groupby returns iterator # Specific file policies are set by internal/sender.py self.set_default_file_policy(filename, DefaultFilePolicy()) files[filename] = self._file_policies[filename].process_chunks( file_chunks_list ) if not files[filename]: del files[filename] for fs in file_stream_utils.split_files(files, max_bytes=util.MAX_LINE_BYTES): self._handle_response( request_with_retry( self._client.post, self._endpoint, json={"files": fs, "dropped": self._dropped_chunks}, retry_callback=self._api.retry_callback, ) ) if uploaded_list: if isinstance( request_with_retry( self._client.post, self._endpoint, json={ "complete": False, "failed": False, "dropped": self._dropped_chunks, "uploaded": uploaded_list, }, ), Exception, ): return False return True def stream_file(self, path: str) -> None: name = path.split("/")[-1] with open(path) as f: self._send([Chunk(name, line) for line in f]) def enqueue_preempting(self) -> None: self._queue.put(self.Preempting()) def push(self, filename: str, data: str) -> None: """Push a chunk of a file to the streaming endpoint. Args: filename: Name of file to append to. data: Text to append to the file. """ self._queue.put(Chunk(filename, data)) def push_success(self, artifact_id: str, save_name: str) -> None: """Notification that a file upload has been successfully completed. Args: artifact_id: ID of artifact save_name: saved name of the uploaded file """ self._queue.put(self.PushSuccess(artifact_id, save_name)) def finish(self, exitcode: int) -> None: """Clean up. Anything pushed after finish will be dropped. Args: exitcode: The exitcode of the watched process. """ logger.info("file stream finish called") self._queue.put(self.Finish(exitcode)) # TODO(jhr): join on a thread which exited with an exception is a noop, clean up this path self._thread.join() logger.info("file stream finish is done") if self._exc_info: logger.error("FileStream exception", exc_info=self._exc_info) # re-raising the original exception, will get re-caught in internal.py for the sender thread if self._exc_info[1] is not None: raise self._exc_info[1].with_traceback(self._exc_info[2]) MAX_SLEEP_SECONDS = 60 * 5 def request_with_retry( func: Callable, *args: Any, **kwargs: Any, ) -> Union["requests.Response", "requests.RequestException"]: """Perform a requests http call, retrying with exponential backoff. Args: func: An http-requesting function to call, like requests.post max_retries: Maximum retries before giving up. By default, we retry 30 times in ~2 hours before dropping the chunk *args: passed through to func **kwargs: passed through to func """ max_retries: int = kwargs.pop("max_retries", 30) retry_callback: Optional[Callable] = kwargs.pop("retry_callback", None) sleep = 2 retry_count = 0 while True: try: response: requests.Response = func(*args, **kwargs) response.raise_for_status() return response except ( requests.exceptions.ConnectionError, requests.exceptions.HTTPError, requests.exceptions.Timeout, ) as e: if isinstance(e, requests.exceptions.HTTPError): # Non-retriable HTTP errors. # # We retry 500s just to be cautious, and because the back end # returns them when there are infrastructure issues. If retrying # some request winds up being problematic, we'll change the # back end to indicate that it shouldn't be retried. if e.response is not None and e.response.status_code in { 400, 403, 404, 409, }: return e if retry_count == max_retries: return e retry_count += 1 delay = sleep + random.random() * 0.25 * sleep if isinstance(e, requests.exceptions.HTTPError) and ( e.response is not None and e.response.status_code == 429 ): err_str = ( f"Filestream rate limit exceeded, retrying in {delay:.1f} seconds. " ) if retry_callback: retry_callback(e.response.status_code, err_str) logger.info(err_str) else: logger.warning( "requests_with_retry encountered retryable exception: %s. func: %s, args: %s, kwargs: %s", e, func, args, kwargs, ) time.sleep(delay) sleep *= 2 if sleep > MAX_SLEEP_SECONDS: sleep = MAX_SLEEP_SECONDS except requests.exceptions.RequestException as e: error_message = "unknown error" try: error_message = response.json()["error"] # todo: clean this up except Exception: pass logger.exception(f"requests_with_retry error: {error_message}") return e