"""sender.""" from __future__ import annotations import contextlib import glob import gzip import json import logging import os import queue import threading import time import traceback from collections import defaultdict from collections.abc import Generator from datetime import datetime from queue import Queue from typing import TYPE_CHECKING, Any, Literal import requests import wandb from wandb import util from wandb.analytics import get_sentry from wandb.errors import CommError, UsageError from wandb.errors.util import ProtobufErrorHandler from wandb.filesync.dir_watcher import DirWatcher from wandb.proto import wandb_internal_pb2 from wandb.sdk.artifacts.artifact_saver import ArtifactSaver from wandb.sdk.interface import interface from wandb.sdk.interface.interface_queue import InterfaceQueue from wandb.sdk.internal import ( context, datastore, file_stream, internal_api, sender_config, ) from wandb.sdk.internal.file_pusher import FilePusher from wandb.sdk.internal.job_builder import JobBuilder from wandb.sdk.internal.settings_static import SettingsStatic from wandb.sdk.lib import ( config_util, filenames, filesystem, proto_util, redirect, retry, telemetry, ) from wandb.sdk.lib.proto_util import message_to_dict if TYPE_CHECKING: from wandb.proto.wandb_internal_pb2 import ( ArtifactManifest, ArtifactManifestEntry, ArtifactRecord, EnvironmentRecord, HttpResponse, LocalInfo, Record, Result, RunExitResult, RunRecord, SummaryRecord, ) StreamLiterals = Literal["stdout", "stderr"] logger = logging.getLogger(__name__) _OUTPUT_MIN_CALLBACK_INTERVAL = 2 # seconds def _framework_priority() -> Generator[tuple[str, str], None, None]: yield from [ ("lightgbm", "lightgbm"), ("catboost", "catboost"), ("xgboost", "xgboost"), ("transformers_huggingface", "huggingface"), # backwards compatibility ("transformers", "huggingface"), ("pytorch_ignite", "ignite"), # backwards compatibility ("ignite", "ignite"), ("pytorch_lightning", "lightning"), ("fastai", "fastai"), ("torch", "torch"), ("keras", "keras"), ("tensorflow", "tensorflow"), ("sklearn", "sklearn"), ] def _manifest_json_from_proto(manifest: ArtifactManifest) -> dict: if manifest.version == 1: if manifest.manifest_file_path: contents = {} with gzip.open(manifest.manifest_file_path, "rt") as f: for line in f: entry_json = json.loads(line) path = entry_json.pop("path") contents[path] = entry_json else: contents = { content.path: _manifest_entry_from_proto(content) for content in manifest.contents } else: raise ValueError(f"unknown artifact manifest version: {manifest.version}") return { "version": manifest.version, "storagePolicy": manifest.storage_policy, "storagePolicyConfig": { config.key: json.loads(config.value_json) for config in manifest.storage_policy_config }, "contents": contents, } def _manifest_entry_from_proto(entry: ArtifactManifestEntry) -> dict: birth_artifact_id = entry.birth_artifact_id if entry.birth_artifact_id else None return { "digest": entry.digest, "birthArtifactID": birth_artifact_id, "ref": entry.ref if entry.ref else None, "size": entry.size if entry.size is not None else None, "local_path": entry.local_path if entry.local_path else None, "skip_cache": entry.skip_cache, "extra": {extra.key: json.loads(extra.value_json) for extra in entry.extra}, } class ResumeState: resumed: bool step: int history: int events: int output: int runtime: float wandb_runtime: int | None summary: dict[str, Any] | None config: dict[str, Any] | None tags: list[str] | None def __init__(self) -> None: self.resumed = False self.step = 0 self.history = 0 self.events = 0 self.output = 0 self.runtime = 0 # wandb_runtime is the canonical runtime (stored in summary._wandb.runtime) self.wandb_runtime = None self.summary = None self.config = None self.tags = None def __str__(self) -> str: obj = ",".join(map(lambda it: f"{it[0]}={it[1]}", vars(self).items())) return f"ResumeState({obj})" class _OutputRawStream: _stopped: threading.Event _queue: queue.Queue _emulator: redirect.TerminalEmulator _writer_thr: threading.Thread _reader_thr: threading.Thread def __init__(self, stream: str, sm: SendManager): self._stopped = threading.Event() self._queue = queue.Queue() self._emulator = redirect.TerminalEmulator() self._writer_thr = threading.Thread( target=sm._output_raw_writer_thread, kwargs=dict(stream=stream), daemon=True, name=f"OutRawWr-{stream}", ) self._reader_thr = threading.Thread( target=sm._output_raw_reader_thread, kwargs=dict(stream=stream), daemon=True, name=f"OutRawRd-{stream}", ) def start(self) -> None: self._writer_thr.start() self._reader_thr.start() class SendManager: UPDATE_CONFIG_TIME: int = 30 UPDATE_STATUS_TIME: int = 5 _settings: SettingsStatic _record_q: Queue[Record] _result_q: Queue[Result] _interface: InterfaceQueue _api_settings: dict[str, str] _partial_output: dict[str, str] _context_keeper: context.ContextKeeper _telemetry_obj: telemetry.TelemetryRecord _environment_obj: EnvironmentRecord _fs: file_stream.FileStreamApi | None _run: RunRecord | None _entity: str | None _project: str | None _dir_watcher: DirWatcher | None _pusher: FilePusher | None _record_exit: Record | None _exit_result: RunExitResult | None _resume_state: ResumeState _rewind_response: dict[str, Any] | None _cached_server_info: dict[str, Any] _cached_viewer: dict[str, Any] _server_messages: list[dict[str, Any]] _ds: datastore.DataStore | None _output_raw_streams: dict[StreamLiterals, _OutputRawStream] _output_raw_file: filesystem.CRDedupedFile | None _send_record_num: int _send_end_offset: int _debounce_config_time: float _debounce_status_time: float def __init__( self, settings: SettingsStatic, record_q: Queue[Record], result_q: Queue[Result], interface: InterfaceQueue, context_keeper: context.ContextKeeper, ) -> None: self._settings = settings self._record_q = record_q self._result_q = result_q self._interface = interface self._context_keeper = context_keeper self._ds = None self._send_record_num = 0 self._send_end_offset = 0 self._fs = None self._pusher = None self._dir_watcher = None # State updated by login self._entity = None self._flags = None # State updated by wandb.init self._run = None self._project = None # keep track of config from key/val updates self._consolidated_config = sender_config.ConfigState() self._start_time: int = 0 self._telemetry_obj = telemetry.TelemetryRecord() self._environment_obj = wandb_internal_pb2.EnvironmentRecord() self._config_metric_pbdict_list: list[dict[int, Any]] = [] self._metadata_summary: dict[str, Any] = defaultdict() self._cached_summary: dict[str, Any] = dict() self._config_metric_index_dict: dict[str, int] = {} self._config_metric_dict: dict[str, wandb_internal_pb2.MetricRecord] = {} self._consolidated_summary: dict[str, Any] = dict() self._cached_server_info = dict() self._cached_viewer = dict() self._server_messages = [] # State updated by resuming self._resume_state = ResumeState() self._rewind_response = None # State added when run_exit is initiated and complete self._record_exit = None self._exit_result = None self._api = internal_api.Api( default_settings=settings, retry_callback=self.retry_callback ) self._api_settings = dict() # queue filled by retry_callback self._retry_q: Queue[HttpResponse] = queue.Queue() # do we need to debounce? self._config_needs_debounce: bool = False # TODO(jhr): do something better, why do we need to send full lines? self._partial_output = dict() self._exit_code = 0 # internal vars for handing raw console output self._output_raw_streams = dict() self._output_raw_file = None # job builder self._job_builder = JobBuilder( settings, files_dir=settings.files_dir, ) time_now = time.monotonic() self._debounce_config_time = time_now self._debounce_status_time = time_now @classmethod def setup( cls, root_dir: str, resume: None | bool | str, ) -> SendManager: """Set up a standalone SendManager. Exclusively used in `sync.py`. """ files_dir = os.path.join(root_dir, "files") settings = wandb.Settings( x_files_dir=files_dir, root_dir=root_dir, # _start_time=0, resume=resume, # ignore_globs=(), x_sync=True, disable_job_creation=False, x_file_stream_timeout_seconds=0, ) record_q: Queue[Record] = queue.Queue() result_q: Queue[Result] = queue.Queue() publish_interface = InterfaceQueue(record_q=record_q) context_keeper = context.ContextKeeper() return SendManager( settings=SettingsStatic(dict(settings)), record_q=record_q, result_q=result_q, interface=publish_interface, context_keeper=context_keeper, ) def __len__(self) -> int: return self._record_q.qsize() def __enter__(self) -> SendManager: return self def __exit__( self, exc_type: type[BaseException] | None, exc_value: BaseException | None, exc_traceback: traceback.TracebackException | None, ) -> Literal[False]: while self: data = next(self) self.send(data) self.finish() return False def retry_callback(self, status: int, response_text: str) -> None: response = wandb_internal_pb2.HttpResponse() response.http_status_code = status response.http_response_text = response_text self._retry_q.put(response) def send(self, record: Record) -> None: self._update_record_num(record.num) self._update_end_offset(record.control.end_offset) record_type = record.WhichOneof("record_type") assert record_type handler_str = "send_" + record_type send_handler = getattr(self, handler_str, None) # Don't log output to reduce log noise if record_type not in {"output", "request", "output_raw"}: logger.debug(f"send: {record_type}") assert send_handler, f"unknown send handler: {handler_str}" context_id = context.context_id_from_record(record) api_context = self._context_keeper.get(context_id) try: self._api.set_local_context(api_context) send_handler(record) except retry.RetryCancelledError: logger.debug(f"Record cancelled: {record_type}") self._context_keeper.release(context_id) finally: self._api.clear_local_context() def send_preempting(self, _: Record) -> None: if self._fs: self._fs.enqueue_preempting() def send_request_sender_mark(self, _: Record) -> None: self._maybe_report_status(always=True) def send_request(self, record: Record) -> None: request_type = record.request.WhichOneof("request_type") assert request_type handler_str = "send_request_" + request_type send_handler = getattr(self, handler_str, None) if request_type != "network_status": logger.debug(f"send_request: {request_type}") assert send_handler, f"unknown handle: {handler_str}" send_handler(record) def _respond_result(self, result: Result) -> None: context_id = context.context_id_from_result(result) self._context_keeper.release(context_id) self._result_q.put(result) def _flatten(self, dictionary: dict) -> None: if isinstance(dictionary, dict): for k, v in list(dictionary.items()): if isinstance(v, dict): self._flatten(v) dictionary.pop(k) for k2, v2 in v.items(): dictionary[k + "." + k2] = v2 def _update_record_num(self, record_num: int) -> None: if not record_num: return # Currently how we handle offline mode and syncing is not # compatible with this assertion due to how the exit record # is (mis)handled: # - using "always_send" in offline mode to trigger defer # state machine # - skipping the exit record in `wandb sync` mode so that # it is always executed as the last record if not self._settings._offline and not self._settings.x_sync: assert record_num == self._send_record_num + 1 self._send_record_num = record_num def _update_end_offset(self, end_offset: int) -> None: if not end_offset: return self._send_end_offset = end_offset def send_request_sender_read(self, record: Record) -> None: if self._ds is None: self._ds = datastore.DataStore() self._ds.open_for_scan(self._settings.sync_file) # TODO(cancel_paused): implement cancel_set logic # The idea is that there is an active request to cancel a # message that is being read from the transaction log below start_offset = record.request.sender_read.start_offset final_offset = record.request.sender_read.final_offset self._ds.seek(start_offset) current_end_offset = 0 while current_end_offset < final_offset: data = self._ds.scan_data() assert data current_end_offset = self._ds.get_offset() send_record = wandb_internal_pb2.Record() send_record.ParseFromString(data) self._update_end_offset(current_end_offset) self.send(send_record) # make sure we perform deferred operations self.debounce() # make sure that we always update writer for every sended read request self._maybe_report_status(always=True) def send_request_stop_status(self, record: Record) -> None: result = proto_util._result_from_record(record) status_resp = result.response.stop_status_response status_resp.run_should_stop = False if self._entity and self._project and self._run and self._run.run_id: try: status_resp.run_should_stop = self._api.check_stop_requested( self._project, self._entity, self._run.run_id ) except Exception as e: logger.warning("Failed to check stop requested status: %s", e) self._respond_result(result) def _maybe_update_config(self, always: bool = False) -> None: time_now = time.monotonic() if ( not always and time_now < self._debounce_config_time + self.UPDATE_CONFIG_TIME ): return if self._config_needs_debounce: self._debounce_config() self._debounce_config_time = time_now def _maybe_report_status(self, always: bool = False) -> None: time_now = time.monotonic() if ( not always and time_now < self._debounce_status_time + self.UPDATE_STATUS_TIME ): return self._debounce_status_time = time_now status_report = wandb_internal_pb2.StatusReportRequest( record_num=self._send_record_num, sent_offset=self._send_end_offset, ) status_time = time.time() status_report.sync_time.FromMicroseconds(int(status_time * 1e6)) record = self._interface._make_request(status_report=status_report) self._interface._publish(record) def debounce(self, final: bool = False) -> None: self._maybe_report_status(always=final) self._maybe_update_config(always=final) def _debounce_config(self) -> None: config_value_dict = self._config_backend_dict() # TODO(jhr): check result of upsert_run? if self._run: self._api.upsert_run( name=self._run.run_id, config=config_value_dict, **self._api_settings, # type: ignore ) self._config_save(config_value_dict) self._config_needs_debounce = False def send_request_network_status(self, record: Record) -> None: result = proto_util._result_from_record(record) status_resp = result.response.network_status_response while True: try: status_resp.network_responses.append(self._retry_q.get_nowait()) except queue.Empty: break except Exception as e: logger.warning(f"Error emptying retry queue: {e}") self._respond_result(result) def send_exit(self, record: Record) -> None: # track where the exit came from self._record_exit = record run_exit = record.exit self._exit_code = run_exit.exit_code logger.info("handling exit code: %s", run_exit.exit_code) runtime = run_exit.runtime logger.info("handling runtime: %s", run_exit.runtime) self._metadata_summary["runtime"] = runtime self._update_summary() # We need to give the request queue a chance to empty between states # so use handle_request_defer as a state machine. logger.info("send defer") self._interface.publish_defer() def send_final(self, record: Record) -> None: pass def _flush_run(self) -> None: pass def send_request_status_report(self, record: Record) -> None: # todo? this is just a noop to please wandb sync pass def send_request_defer(self, record: Record) -> None: # noqa: C901 defer = record.request.defer state = defer.state logger.info(f"handle sender defer: {state}") def transition_state() -> None: state = defer.state + 1 logger.info(f"send defer: {state}") self._interface.publish_defer(state) done = False if state == defer.BEGIN: transition_state() elif state == defer.FLUSH_RUN: self._flush_run() transition_state() elif state == defer.FLUSH_STATS: # NOTE: this is handled in handler.py:handle_request_defer() transition_state() elif state == defer.FLUSH_PARTIAL_HISTORY: # NOTE: this is handled in handler.py:handle_request_defer() transition_state() elif state == defer.FLUSH_TB: # NOTE: this is handled in handler.py:handle_request_defer() transition_state() elif state == defer.FLUSH_SUM: # NOTE: this is handled in handler.py:handle_request_defer() transition_state() elif state == defer.FLUSH_DEBOUNCER: self.debounce(final=True) transition_state() elif state == defer.FLUSH_OUTPUT: self._output_raw_finish() transition_state() elif state == defer.FLUSH_JOB: self._flush_job() transition_state() elif state == defer.FLUSH_DIR: if self._dir_watcher: self._dir_watcher.finish() self._dir_watcher = None transition_state() elif state == defer.FLUSH_FP: if self._pusher: # FilePusher generates some events for FileStreamApi, so we # need to wait for pusher to finish before going to the next # state to ensure that filestream gets all the events that we # want before telling it to finish up self._pusher.finish(transition_state) else: transition_state() elif state == defer.JOIN_FP: if self._pusher: self._pusher.join() transition_state() elif state == defer.FLUSH_FS: if self._fs: # TODO(jhr): now is a good time to output pending output lines self._fs.finish(self._exit_code) self._fs = None transition_state() elif state == defer.FLUSH_FINAL: self._interface.publish_final() self._interface.publish_footer() transition_state() elif state == defer.END: done = True else: raise AssertionError("unknown state") if not done: return exit_result = wandb_internal_pb2.RunExitResult() # mark exit done in case we are polling on exit self._exit_result = exit_result # Report response to mailbox if self._record_exit and self._record_exit.control.mailbox_slot: result = proto_util._result_from_record(self._record_exit) result.exit_result.CopyFrom(exit_result) self._respond_result(result) def send_request_poll_exit(self, record: Record) -> None: if not record.control.req_resp and not record.control.mailbox_slot: return result = proto_util._result_from_record(record) if self._pusher: _alive, status = self._pusher.get_status() file_counts = self._pusher.file_counts_by_category() resp = result.response.poll_exit_response resp.pusher_stats.uploaded_bytes = status.uploaded_bytes resp.pusher_stats.total_bytes = status.total_bytes resp.pusher_stats.deduped_bytes = status.deduped_bytes resp.file_counts.wandb_count = file_counts.wandb resp.file_counts.media_count = file_counts.media resp.file_counts.artifact_count = file_counts.artifact resp.file_counts.other_count = file_counts.other if self._exit_result: result.response.poll_exit_response.done = True result.response.poll_exit_response.exit_result.CopyFrom(self._exit_result) self._respond_result(result) def _setup_resume(self, run: RunRecord) -> wandb_internal_pb2.ErrorInfo | None: """Queries the backend for a run; fail if the settings are incompatible.""" if not self._settings.resume: return None # TODO: This causes a race, we need to make the upsert atomically # only create or update depending on the resume config # we use the runs entity if set, otherwise fallback to users entity # todo: ensure entity is not None as self._entity is Optional[str] entity = run.entity or self._entity logger.info( "checking resume status for %s/%s/%s", entity, run.project, run.run_id ) resume_status = self._api.run_resume_status( entity=entity, # type: ignore project_name=run.project, name=run.run_id, ) # No resume status = run does not exist; No t key in wandbConfig = run exists but hasn't been inited if not resume_status or '"t":' not in resume_status.get("wandbConfig", ""): if self._settings.resume == "must": error = wandb_internal_pb2.ErrorInfo() error.code = wandb_internal_pb2.ErrorInfo.ErrorCode.USAGE error.message = ( "You provided an invalid value for the `resume` argument." f" The value 'must' is not a valid option for resuming a run ({run.run_id}) that has not been initialized." " Please check your inputs and try again with a valid run ID." " If you are trying to start a new run, please omit the `resume` argument or use `resume='allow'`." ) return error return None # # handle cases where we have resume_status # if self._settings.resume == "never": error = wandb_internal_pb2.ErrorInfo() error.code = wandb_internal_pb2.ErrorInfo.ErrorCode.USAGE error.message = ( "You provided an invalid value for the `resume` argument." f" The value 'never' is not a valid option for resuming a run ({run.run_id}) that already exists." " Please check your inputs and try again with a valid value for the `resume` argument." ) return error history = {} events = {} config = {} summary = {} try: events_rt = 0 history_rt = 0 history = json.loads(resume_status["historyTail"]) if history: history = json.loads(history[-1]) history_rt = history.get("_runtime", 0) events = json.loads(resume_status["eventsTail"]) if events: events = json.loads(events[-1]) events_rt = events.get("_runtime", 0) config = json.loads(resume_status["config"] or "{}") summary = json.loads(resume_status["summaryMetrics"] or "{}") new_runtime = summary.get("_wandb", {}).get("runtime", None) if new_runtime is not None: self._resume_state.wandb_runtime = new_runtime tags = resume_status.get("tags") or [] except (IndexError, ValueError): logger.exception("unable to load resume tails") if self._settings.resume == "must": error = wandb_internal_pb2.ErrorInfo() error.code = wandb_internal_pb2.ErrorInfo.ErrorCode.USAGE error.message = f"resume='must' but could not resume ({run.run_id}) " return error # TODO: Do we need to restore config / summary? # System metrics runtime is usually greater than history self._resume_state.runtime = max(events_rt, history_rt) last_step = history.get("_step", 0) history_line_count = resume_status["historyLineCount"] self._resume_state.step = last_step + 1 if history_line_count > 0 else last_step self._resume_state.history = history_line_count self._resume_state.events = resume_status["eventsLineCount"] self._resume_state.output = resume_status["logLineCount"] self._resume_state.config = config self._resume_state.summary = summary self._resume_state.tags = tags self._resume_state.resumed = True logger.info(f"configured resuming with: {self._resume_state}") return None def _telemetry_get_framework(self) -> str: """Get telemetry data for internal config structure.""" # detect framework by checking what is loaded imports: telemetry.TelemetryImports if self._telemetry_obj.HasField("imports_finish"): imports = self._telemetry_obj.imports_finish elif self._telemetry_obj.HasField("imports_init"): imports = self._telemetry_obj.imports_init else: return "" framework = next( (n for f, n in _framework_priority() if getattr(imports, f, False)), "" ) return framework def _config_backend_dict(self) -> sender_config.BackendConfigDict: config = self._consolidated_config or sender_config.ConfigState() return config.to_backend_dict( telemetry_record=self._telemetry_obj, framework=self._telemetry_get_framework(), start_time_millis=self._start_time, metric_pbdicts=self._config_metric_pbdict_list, environment_record=self._environment_obj, ) def _config_save( self, config_value_dict: sender_config.BackendConfigDict, ) -> None: config_path = os.path.join(self._settings.files_dir, "config.yaml") config_util.save_config_file_from_dict(config_path, config_value_dict) def _sync_spell(self) -> None: """Sync this run with spell.""" if not self._run: return try: env = os.environ self._interface.publish_config( key=("_wandb", "spell_url"), val=env.get("SPELL_RUN_URL") ) url = f"{self._api.app_url}/{self._run.entity}/{self._run.project}/runs/{self._run.run_id}" requests.put( env.get("SPELL_API_URL", "https://api.spell.run") + "/wandb_url", json={"access_token": env.get("WANDB_ACCESS_TOKEN"), "url": url}, timeout=2, ) except requests.RequestException: pass # TODO: do something if sync spell is not successful? def _setup_fork(self, server_run: dict): assert self._run assert self._run.branch_point first_step = int(self._run.branch_point.value) + 1 self._resume_state.step = first_step self._resume_state.history = server_run.get("historyLineCount", 0) self._run.forked = True self._run.starting_step = first_step def _load_rewind_state(self, run: RunRecord): assert run.branch_point self._rewind_response = self._api.rewind_run( run_name=run.run_id, entity=run.entity or None, project=run.project or None, metric_name=run.branch_point.metric, metric_value=run.branch_point.value, program_path=self._settings.program or None, ) self._resume_state.history = self._rewind_response.get("historyLineCount", 0) self._resume_state.config = json.loads( self._rewind_response.get("config", "{}") ) def _install_rewind_state(self): assert self._run assert self._run.branch_point assert self._rewind_response first_step = int(self._run.branch_point.value) + 1 self._resume_state.step = first_step # We set the fork flag here because rewind uses the forking # infrastructure under the hood. Setting `forked` here # ensures that run._step is properly set in the user process. self._run.forked = True self._run.starting_step = first_step def _handle_error( self, record: Record, error: wandb_internal_pb2.ErrorInfo, run: RunRecord, ) -> None: if record.control.req_resp or record.control.mailbox_slot: result = proto_util._result_from_record(record) result.run_result.run.CopyFrom(run) result.run_result.error.CopyFrom(error) self._respond_result(result) else: logger.error("Got error in async mode: %s", error.message) def send_run(self, record: Record, file_dir: str | None = None) -> None: run = record.run error = None is_wandb_init = self._run is None # save start time of a run self._start_time = int(run.start_time.ToMicroseconds() // 1e6) # update telemetry if run.telemetry: self._telemetry_obj.MergeFrom(run.telemetry) if self._settings.x_sync: self._telemetry_obj.feature.sync = True # build config dict config_value_dict: sender_config.BackendConfigDict | None = None if run.config: self._consolidated_config.update_from_proto(run.config) config_value_dict = self._config_backend_dict() self._config_save(config_value_dict) do_rewind = run.branch_point.run == run.run_id do_fork = not do_rewind and run.branch_point.run != "" do_resume = bool(self._settings.resume) num_resume_options_set = sum([do_fork, do_rewind, do_resume]) if num_resume_options_set > 1: error = wandb_internal_pb2.ErrorInfo() error.code = wandb_internal_pb2.ErrorInfo.ErrorCode.USAGE error.message = ( "Multiple resume options specified. " "Please specify only one of `fork_from`, `resume`, or `resume_from`." ) self._handle_error(record, error, run) if is_wandb_init: # Ensure we have a project to query for status if run.project == "": run.project = util.auto_project_name(self._settings.program) # Only check resume status on `wandb.init` if do_resume: error = self._setup_resume(run) elif do_rewind: error = self._load_rewind_state(run) if error is not None: self._handle_error(record, error, run) return # Save the resumed config if self._resume_state.config is not None: self._consolidated_config.merge_resumed_config( config_util.dict_strip_value_dict(self._resume_state.config) ) config_value_dict = self._config_backend_dict() self._config_save(config_value_dict) # handle empty config # TODO(jhr): consolidate the 4 ways config is built: # (passed config, empty config, resume config, send_config) if not config_value_dict: config_value_dict = self._config_backend_dict() self._config_save(config_value_dict) try: server_run = self._init_run(run, config_value_dict) except (CommError, UsageError) as e: logger.error(e, exc_info=True) error = ProtobufErrorHandler.from_exception(e) self._handle_error(record, error, run) return assert self._run # self._run is configured in _init_run() if do_fork: error = self._setup_fork(server_run) if error is not None: self._handle_error(record, error, run) return if record.control.req_resp or record.control.mailbox_slot: result = proto_util._result_from_record(record) # TODO: we could do self._interface.publish_defer(resp) to notify # the handler not to actually perform server updates for this uuid # because the user process will send a summary update when we resume result.run_result.run.CopyFrom(self._run) self._respond_result(result) # Only spin up our threads on the first run message if is_wandb_init: self._start_run_threads(file_dir) else: logger.info("updated run: %s", self._run.run_id) def _update_resume_state(self, is_rewinding: bool, inserted: bool): assert self._run if self._resume_state.resumed: self._run.resumed = True if self._resume_state.wandb_runtime is not None: self._run.runtime = self._resume_state.wandb_runtime elif is_rewinding: # because is_rewinding is mutually exclusive with self._resume_state.resumed, # this block will always execute if is_rewinding is set self._install_rewind_state() else: # If the user is not resuming, and we didn't insert on upsert_run then # it is likely that we are overwriting the run which we might want to # prevent in the future. This could be a false signal since an upsert_run # message which gets retried in the network could also show up as not # inserted. if not inserted: # no need to flush this, it will get updated eventually self._telemetry_obj.feature.maybe_run_overwrite = True def _init_run( self, run: RunRecord, config_dict: sender_config.BackendConfigDict | None, ) -> dict: # We subtract the previous runs runtime when resuming start_time = ( run.start_time.ToMicroseconds() / 1e6 ) - self._resume_state.runtime # TODO: we don't check inserted currently, ultimately we should make # the upsert know the resume state and fail transactionally if self._resume_state and self._resume_state.tags and not run.tags: run.tags.extend(self._resume_state.tags) is_rewinding = bool(self._settings.resume_from) if is_rewinding: assert self._rewind_response server_run = self._rewind_response server_messages = None inserted = True else: server_run, inserted, server_messages = self._api.upsert_run( name=run.run_id, entity=run.entity or None, project=run.project or None, group=run.run_group or None, job_type=run.job_type or None, display_name=run.display_name or None, notes=run.notes or None, tags=run.tags[:] or None, config=config_dict or None, sweep_name=run.sweep_id or None, host=run.host or None, program_path=self._settings.program or None, repo=run.git.remote_url or None, commit=run.git.commit or None, ) # TODO: we don't want to create jobs in sweeps, since the # executable doesn't appear to be consistent if run.sweep_id: self._job_builder.disable = True self._server_messages = server_messages or [] self._run = run if self._resume_state.resumed and is_rewinding: # this should not ever be possible to hit, since we check for # resumption above and raise an error if resumption is specified # twice. raise ValueError( "Cannot attempt to rewind and resume a run - only one of " "`resume` or `resume_from` can be specified." ) self._update_resume_state(is_rewinding, inserted) self._run.starting_step = self._resume_state.step self._run.start_time.FromMicroseconds(int(start_time * 1e6)) self._run.config.CopyFrom(self._interface._make_config(config_dict)) if self._resume_state.summary is not None: self._run.summary.CopyFrom( self._interface._make_summary_from_dict(self._resume_state.summary) ) storage_id = server_run.get("id") if storage_id: self._run.storage_id = storage_id id = server_run.get("name") if id: self._api.set_current_run_id(id) display_name = server_run.get("displayName") if display_name: self._run.display_name = display_name project = server_run.get("project") # TODO: remove self._api.set_settings, and make self._project a property? if project: project_name = project.get("name") if project_name: self._run.project = project_name self._project = project_name self._api_settings["project"] = project_name self._api.set_setting("project", project_name) entity = project.get("entity") if entity: entity_name = entity.get("name") if entity_name: self._run.entity = entity_name self._entity = entity_name self._api_settings["entity"] = entity_name self._api.set_setting("entity", entity_name) sweep_id = server_run.get("sweepName") if sweep_id: self._run.sweep_id = sweep_id if os.getenv("SPELL_RUN_URL"): self._sync_spell() return server_run def _start_run_threads(self, file_dir: str | None = None) -> None: assert self._run # self._run is configured by caller self._fs = file_stream.FileStreamApi( self._api, self._run.run_id, self._run.start_time.ToMicroseconds() / 1e6, timeout=self._settings.x_file_stream_timeout_seconds or 0, settings=self._api_settings, ) # Ensure the streaming polices have the proper offsets self._fs.set_file_policy("wandb-summary.json", file_stream.SummaryFilePolicy()) self._fs.set_file_policy( "wandb-history.jsonl", file_stream.JsonlFilePolicy(start_chunk_id=self._resume_state.history), ) self._fs.set_file_policy( "wandb-events.jsonl", file_stream.JsonlFilePolicy(start_chunk_id=self._resume_state.events), ) self._fs.set_file_policy( "output.log", file_stream.CRDedupeFilePolicy(start_chunk_id=self._resume_state.output), ) # hack to merge run_settings and self._settings object together # so that fields like entity or project are available to be attached to Sentry events. run_settings = message_to_dict(self._run) _settings = dict(self._settings) _settings.update(run_settings) get_sentry().configure_scope(tags=_settings, process_context="internal") self._fs.start() self._pusher = FilePusher(self._api, self._fs, settings=self._settings) self._dir_watcher = DirWatcher(self._settings, self._pusher, file_dir) logger.info( "run started: %s with start time %s", self._run.run_id, self._run.start_time.ToMicroseconds() / 1e6, ) def _save_history(self, history_dict: dict[str, Any]) -> None: if self._fs: self._fs.push(filenames.HISTORY_FNAME, json.dumps(history_dict)) def send_history(self, record: Record) -> None: history = record.history history_dict = proto_util.dict_from_proto_list(history.item) self._save_history(history_dict) def _update_summary_record(self, summary: SummaryRecord) -> None: summary_dict = proto_util.dict_from_proto_list(summary.update) self._cached_summary = summary_dict self._update_summary() def send_summary(self, record: Record) -> None: self._update_summary_record(record.summary) def send_request_summary_record(self, record: Record) -> None: self._update_summary_record(record.request.summary_record.summary) def _update_summary(self) -> None: summary_dict = self._cached_summary.copy() summary_dict.pop("_wandb", None) if self._metadata_summary: summary_dict["_wandb"] = self._metadata_summary # merge with consolidated summary self._consolidated_summary.update(summary_dict) json_summary = json.dumps(self._consolidated_summary) if self._fs: self._fs.push(filenames.SUMMARY_FNAME, json_summary) # TODO(jhr): we should only write this at the end of the script summary_path = os.path.join(self._settings.files_dir, filenames.SUMMARY_FNAME) with open(summary_path, "w") as f: f.write(json_summary) self._save_file(filesystem.GlobStr(filenames.SUMMARY_FNAME)) def send_stats(self, record: Record) -> None: stats = record.stats if stats.stats_type != wandb_internal_pb2.StatsRecord.StatsType.SYSTEM: return if not self._fs: return if not self._run: return now_us = stats.timestamp.ToMicroseconds() start_us = self._run.start_time.ToMicroseconds() d = dict() for item in stats.item: try: d[item.key] = json.loads(item.value_json) except json.JSONDecodeError: logger.exception("error decoding stats json: %s", item.value_json) row: dict[str, Any] = dict(system=d) self._flatten(row) row["_wandb"] = True row["_timestamp"] = now_us / 1e6 row["_runtime"] = (now_us - start_us) / 1e6 self._fs.push(filenames.EVENTS_FNAME, json.dumps(row)) # TODO(jhr): check fs.push results? def _output_raw_finish(self) -> None: for stream, output_raw in self._output_raw_streams.items(): output_raw._stopped.set() # shut down threads output_raw._writer_thr.join(timeout=5) if output_raw._writer_thr.is_alive(): logger.info("processing output...") output_raw._writer_thr.join() output_raw._reader_thr.join() # flush output buffers and files self._output_raw_flush(stream) self._output_raw_streams = {} if self._output_raw_file: self._output_raw_file.close() self._output_raw_file = None def _output_raw_writer_thread(self, stream: StreamLiterals) -> None: while True: output_raw = self._output_raw_streams[stream] if output_raw._queue.empty(): if output_raw._stopped.is_set(): return time.sleep(0.5) continue data = [] while not output_raw._queue.empty(): data.append(output_raw._queue.get()) if output_raw._stopped.is_set() and sum(map(len, data)) > 100000: logger.warning("Terminal output too large. Logging without processing.") self._output_raw_flush(stream) for line in data: self._output_raw_flush(stream, line) # TODO: lets mark that this happened in telemetry return try: output_raw._emulator.write("".join(data)) except Exception as e: logger.warning(f"problem writing to output_raw emulator: {e}") def _output_raw_reader_thread(self, stream: StreamLiterals) -> None: output_raw = self._output_raw_streams[stream] while not (output_raw._stopped.is_set() and output_raw._queue.empty()): self._output_raw_flush(stream) time.sleep(_OUTPUT_MIN_CALLBACK_INTERVAL) def _output_raw_flush( self, stream: StreamLiterals, data: str | None = None ) -> None: if data is None: output_raw = self._output_raw_streams[stream] try: data = output_raw._emulator.read() except Exception as e: logger.warning(f"problem reading from output_raw emulator: {e}") if data: self._send_output_line(stream, data) if self._output_raw_file: self._output_raw_file.write(data.encode("utf-8")) def send_request_python_packages(self, record: Record) -> None: import os from wandb.sdk.lib.filenames import REQUIREMENTS_FNAME installed_packages_list = sorted( f"{r.name}=={r.version}" for r in record.request.python_packages.package ) with open(os.path.join(self._settings.files_dir, REQUIREMENTS_FNAME), "w") as f: f.write("\n".join(installed_packages_list)) def send_output(self, record: Record) -> None: if not self._fs: return out = record.output stream: StreamLiterals = "stdout" if out.output_type == wandb_internal_pb2.OutputRecord.OutputType.STDERR: stream = "stderr" line = out.line self._send_output_line(stream, line) def send_output_raw(self, record: Record) -> None: if not self._fs: return out = record.output_raw stream: StreamLiterals = "stdout" if out.output_type == wandb_internal_pb2.OutputRawRecord.OutputType.STDERR: stream = "stderr" line = out.line output_raw = self._output_raw_streams.get(stream) if not output_raw: output_raw = _OutputRawStream(stream=stream, sm=self) self._output_raw_streams[stream] = output_raw # open the console output file shared between both streams if not self._output_raw_file: output_log_path = os.path.join( self._settings.files_dir, filenames.OUTPUT_FNAME ) output_raw_file = None try: output_raw_file = filesystem.CRDedupedFile( open(output_log_path, "wb") ) except OSError as e: logger.warning(f"could not open output_raw_file: {e}") if output_raw_file: self._output_raw_file = output_raw_file output_raw.start() output_raw._queue.put(line) def _send_output_line(self, stream: StreamLiterals, line: str) -> None: """Combined writer for raw and non raw output lines. This is combined because they are both post emulator. """ prepend = "" if stream == "stderr": prepend = "ERROR " if not line.endswith("\n"): self._partial_output.setdefault(stream, "") if line.startswith("\r"): # TODO: maybe we shouldn't just drop this, what if there was some \ns in the partial # that should probably be the check instead of not line.endswith(\n") # logger.info(f"Dropping data {self._partial_output[stream]}") self._partial_output[stream] = "" self._partial_output[stream] += line # TODO(jhr): how do we make sure this gets flushed? # we might need this for other stuff like telemetry else: # TODO(jhr): use time from timestamp proto # TODO(jhr): do we need to make sure we write full lines? # seems to be some issues with line breaks cur_time = time.time() timestamp = datetime.utcfromtimestamp(cur_time).isoformat() + " " prev_str = self._partial_output.get(stream, "") line = f"{prepend}{timestamp}{prev_str}{line}" if self._fs: self._fs.push(filenames.OUTPUT_FNAME, line) self._partial_output[stream] = "" def _update_config(self) -> None: self._config_needs_debounce = True def send_config(self, record: Record) -> None: self._consolidated_config.update_from_proto(record.config) self._update_config() def send_metric(self, record: Record) -> None: metric = record.metric if metric.glob_name: logger.warning("Seen metric with glob (shouldn't happen)") return # merge or overwrite old_metric = self._config_metric_dict.get( metric.name, wandb_internal_pb2.MetricRecord() ) if metric._control.overwrite: old_metric.CopyFrom(metric) else: old_metric.MergeFrom(metric) self._config_metric_dict[metric.name] = old_metric metric = old_metric # convert step_metric to index if metric.step_metric: find_step_idx = self._config_metric_index_dict.get(metric.step_metric) if find_step_idx is not None: # make a copy of this metric as we will be modifying it rec = wandb_internal_pb2.Record() rec.metric.CopyFrom(metric) metric = rec.metric metric.ClearField("step_metric") metric.step_metric_index = find_step_idx + 1 md: dict[int, Any] = proto_util.proto_encode_to_dict(metric) find_idx = self._config_metric_index_dict.get(metric.name) if find_idx is not None: self._config_metric_pbdict_list[find_idx] = md else: next_idx = len(self._config_metric_pbdict_list) self._config_metric_pbdict_list.append(md) self._config_metric_index_dict[metric.name] = next_idx self._debounce_config() def _update_telemetry_record(self, telemetry: telemetry.TelemetryRecord) -> None: self._telemetry_obj.MergeFrom(telemetry) self._debounce_config() def send_telemetry(self, record: Record) -> None: self._update_telemetry_record(record.telemetry) def send_request_telemetry_record(self, record: Record) -> None: self._update_telemetry_record(record.request.telemetry_record.telemetry) def _save_file( self, fname: filesystem.GlobStr, policy: filesystem.PolicyName = "end" ) -> None: logger.info("saving file %s with policy %s", fname, policy) if self._dir_watcher: self._dir_watcher.update_policy(fname, policy) def send_files(self, record: Record) -> None: files = record.files for k in files.files: # TODO(jhr): fix paths with directories self._save_file( filesystem.GlobStr(glob.escape(k.path)), interface.file_enum_to_policy(k.policy), ) def send_header(self, record: Record) -> None: pass def send_footer(self, record: Record) -> None: pass def send_tbrecord(self, record: Record) -> None: # tbrecord watching threads are handled by handler.py pass def _update_environment_record(self, environment: EnvironmentRecord) -> None: self._environment_obj.MergeFrom(environment) self._debounce_config() def send_environment(self, record: Record) -> None: """Inject environment info into config and upload as a JSON file.""" self._update_environment_record(record.environment) environment_json = json.dumps(proto_util.message_to_dict(self._environment_obj)) with open( os.path.join(self._settings.files_dir, filenames.METADATA_FNAME), "w" ) as f: f.write(environment_json) self._save_file(filesystem.GlobStr(filenames.METADATA_FNAME), policy="now") def send_request_link_artifact(self, record: Record) -> None: if not (record.control.req_resp or record.control.mailbox_slot): raise ValueError( f"Expected either `req_resp` or `mailbox_slot`, got: {record.control!r}" ) result = proto_util._result_from_record(record) link = record.request.link_artifact client_id = link.client_id server_id = link.server_id portfolio_name = link.portfolio_name entity = link.portfolio_entity project = link.portfolio_project aliases = link.portfolio_aliases organization = link.portfolio_organization logger.debug( f"link_artifact params - client_id={client_id}, server_id={server_id}, " f"portfolio_name={portfolio_name}, entity={entity}, project={project}, " f"organization={organization}" ) if (client_id or server_id) and portfolio_name and entity and project: try: response = self._api.link_artifact( client_id, server_id, portfolio_name, entity, project, aliases, organization, ) result.response.link_artifact_response.version_index = response[ "versionIndex" ] except Exception as e: org_or_entity = organization or entity result.response.link_artifact_response.error_message = ( f"error linking artifact to " f'"{org_or_entity}/{project}/{portfolio_name}"; error: {e}' ) logger.warning("Failed to link artifact to portfolio: %s", e) self._respond_result(result) def send_use_artifact(self, record: Record) -> None: """Pretend to send a used artifact. This function doesn't actually send anything, it is just used internally. """ use = record.use_artifact if use.type == "job" and not use.partial.job_name: self._job_builder.disable = True elif use.partial.job_name: # job is partial, let job builder rebuild job, set job source dict self._job_builder.set_partial_source_id(use.id) def send_request_log_artifact(self, record: Record) -> None: result = proto_util._result_from_record(record) artifact = record.request.log_artifact.artifact history_step = record.request.log_artifact.history_step try: res = self._send_artifact(artifact, history_step) assert res, "Unable to send artifact" result.response.log_artifact_response.artifact_id = res["id"] logger.info(f"logged artifact {artifact.name} - {res}") except Exception as e: result.response.log_artifact_response.error_message = ( f'error logging artifact "{artifact.type}/{artifact.name}": {e}' ) self._respond_result(result) def send_artifact(self, record: Record) -> None: artifact = record.artifact try: res = self._send_artifact(artifact) logger.info(f"sent artifact {artifact.name} - {res}") except Exception: logger.exception( f'send_artifact: failed for artifact "{artifact.type}/{artifact.name}"' ) def _send_artifact( self, artifact: ArtifactRecord, history_step: int | None = None ) -> dict | None: from packaging.version import parse assert self._pusher saver = ArtifactSaver( api=self._api, digest=artifact.digest, manifest_json=_manifest_json_from_proto(artifact.manifest), file_pusher=self._pusher, is_user_created=artifact.user_created, ) if artifact.distributed_id: max_cli_version = self._max_cli_version() if max_cli_version is None or parse(max_cli_version) < parse("0.10.16"): logger.warning( "This W&B Server doesn't support distributed artifacts, " "have your administrator install wandb/local >= 0.9.37" ) return None metadata = json.loads(artifact.metadata) if artifact.metadata else None res = saver.save( entity=artifact.entity, project=artifact.project, type=artifact.type, name=artifact.name, client_id=artifact.client_id, sequence_client_id=artifact.sequence_client_id, metadata=metadata, ttl_duration_seconds=artifact.ttl_duration_seconds or None, description=artifact.description or None, aliases=artifact.aliases, tags=artifact.tags, use_after_commit=artifact.use_after_commit, distributed_id=artifact.distributed_id, finalize=artifact.finalize, incremental=artifact.incremental_beta1, history_step=history_step, base_id=artifact.base_id or None, ) self._job_builder._handle_server_artifact(res, artifact) if artifact.manifest.manifest_file_path: with contextlib.suppress(FileNotFoundError): os.remove(artifact.manifest.manifest_file_path) return res def send_alert(self, record: Record) -> None: from packaging.version import parse alert = record.alert max_cli_version = self._max_cli_version() if max_cli_version is None or parse(max_cli_version) < parse("0.10.9"): logger.warning( "This W&B server doesn't support alerts, " "have your administrator install wandb/local >= 0.9.31" ) else: try: self._api.notify_scriptable_run_alert( title=alert.title, text=alert.text, level=alert.level, wait_duration=alert.wait_duration, ) except Exception: logger.exception(f"send_alert: failed for alert {alert.title!r}") def finish(self) -> None: logger.info("shutting down sender") # if self._tb_watcher: # self._tb_watcher.finish() self._output_raw_finish() if self._dir_watcher: self._dir_watcher.finish() self._dir_watcher = None if self._pusher: self._pusher.finish() self._pusher.join() self._pusher = None if self._fs: self._fs.finish(self._exit_code) self._fs = None get_sentry().end_session() def _max_cli_version(self) -> str | None: server_info = self.get_server_info() max_cli_version = server_info.get("cliVersionInfo", {}).get( "max_cli_version", None ) if not isinstance(max_cli_version, str): return None return max_cli_version def get_viewer_server_info(self) -> None: if self._cached_server_info and self._cached_viewer: return self._cached_viewer, self._cached_server_info = self._api.viewer_server_info() def get_viewer_info(self) -> dict[str, Any]: if not self._cached_viewer: self.get_viewer_server_info() return self._cached_viewer def get_server_info(self) -> dict[str, Any]: if not self._cached_server_info: self.get_viewer_server_info() return self._cached_server_info def get_local_info(self) -> LocalInfo: """Queries the server to get the local version information. First, we perform an introspection, if it returns empty we deduce that the docker image is out-of-date. Otherwise, we use the returned values to deduce the state of the local server. """ local_info = wandb_internal_pb2.LocalInfo() if self._settings._offline: local_info.out_of_date = False return local_info latest_local_version = "latest" # Assuming the query is successful if the result is empty it indicates that # the backend is out of date since it doesn't have the desired field server_info = self.get_server_info() latest_local_version_info = server_info.get("latestLocalVersionInfo", {}) if latest_local_version_info is None: local_info.out_of_date = False else: local_info.out_of_date = latest_local_version_info.get("outOfDate", True) local_info.version = latest_local_version_info.get( "latestVersionString", latest_local_version ) return local_info def _flush_job(self) -> None: if self._job_builder.disable or self._settings._offline: return self._job_builder.set_config(self._consolidated_config.non_internal_config()) summary_dict = self._cached_summary.copy() summary_dict.pop("_wandb", None) self._job_builder.set_summary(summary_dict) artifact = self._job_builder.build(api=self._api) if artifact is not None and self._run is not None: proto_artifact = self._interface._make_artifact(artifact) proto_artifact.run_id = self._run.run_id proto_artifact.project = self._run.project proto_artifact.entity = self._run.entity # TODO: this should be removed when the latest tag is handled # by the backend (WB-12116) proto_artifact.aliases.append("latest") # add docker image tag for alias in self._job_builder._aliases: proto_artifact.aliases.append(alias) proto_artifact.user_created = True proto_artifact.use_after_commit = True proto_artifact.finalize = True self._interface._publish_artifact(proto_artifact) def __next__(self) -> Record: return self._record_q.get(block=True) next = __next__