base_slurm_runner

BaseSlurmRunner

Bases: BaseRunner

Source code in fractal_server/app/runner/executors/slurm_common/base_slurm_runner.py

class BaseSlurmRunner(BaseRunner):
    shutdown_file: Path
    common_script_lines: list[str]
    user_cache_dir: str
    root_dir_local: Path
    root_dir_remote: Path
    poll_interval: int
    poll_interval_internal: float
    jobs: dict[str, SlurmJob]
    python_worker_interpreter: str
    slurm_runner_type: Literal["ssh", "sudo"]

    def __init__(
        self,
        root_dir_local: Path,
        root_dir_remote: Path,
        slurm_runner_type: Literal["ssh", "sudo"],
        python_worker_interpreter: str,
        common_script_lines: Optional[list[str]] = None,
        user_cache_dir: Optional[str] = None,
        poll_interval: Optional[int] = None,
    ):
        self.slurm_runner_type = slurm_runner_type
        self.root_dir_local = root_dir_local
        self.root_dir_remote = root_dir_remote
        self.common_script_lines = common_script_lines or []
        self._check_slurm_account()
        self.user_cache_dir = user_cache_dir
        self.python_worker_interpreter = python_worker_interpreter

        settings = Inject(get_settings)

        self.poll_interval = (
            poll_interval or settings.FRACTAL_SLURM_POLL_INTERVAL
        )
        self.poll_interval_internal = self.poll_interval / 10.0

        self.check_fractal_server_versions()

        # Create job folders. Note that the local one may or may not exist
        # depending on whether it is a test or an actual run
        try:
            if not self.root_dir_local.is_dir():
                self._mkdir_local_folder(self.root_dir_local.as_posix())
            self._mkdir_remote_folder(self.root_dir_remote.as_posix())
        except Exception as e:
            error_msg = (
                f"Could not mkdir {self.root_dir_local.as_posix()} or "
                f"{self.root_dir_remote.as_posix()}. "
                f"Original error: {str(e)}."
            )
            logger.error(error_msg)
            raise RuntimeError(error_msg)

        self.shutdown_file = self.root_dir_local / SHUTDOWN_FILENAME
        self.jobs = {}

    def __enter__(self):
        return self

    def __exit__(self, exc_type, exc_val, exc_tb):
        return False

    def _run_local_cmd(self, cmd: str) -> str:
        raise NotImplementedError("Implement in child class.")

    def _run_remote_cmd(self, cmd: str) -> str:
        raise NotImplementedError("Implement in child class.")

    def run_squeue(self, job_ids: list[str]) -> tuple[bool, str]:
        # NOTE: see issue 2482

        if len(job_ids) == 0:
            return (False, "")

        job_id_single_str = ",".join([str(j) for j in job_ids])
        cmd = (
            f"squeue --noheader --format='%i %T' --jobs {job_id_single_str}"
            " --states=all"
        )

        try:
            if self.slurm_runner_type == "sudo":
                stdout = self._run_local_cmd(cmd)
            else:
                stdout = self._run_remote_cmd(cmd)
            return (True, stdout)
        except Exception as e:
            logger.info(f"{cmd=} failed with {str(e)}")
            return (False, "")

    def _get_finished_jobs(self, job_ids: list[str]) -> set[str]:
        #  If there is no Slurm job to check, return right away

        if not job_ids:
            return set()
        id_to_state = dict()

        success, stdout = self.run_squeue(job_ids)
        if success:
            id_to_state = {
                out.split()[0]: out.split()[1] for out in stdout.splitlines()
            }
        else:
            id_to_state = dict()
            for j in job_ids:
                success, res = self.run_squeue([j])
                if not success:
                    logger.info(f"Job {j} not found. Marked it as completed")
                    id_to_state.update({str(j): "COMPLETED"})
                else:
                    id_to_state.update(
                        {res.stdout.split()[0]: res.stdout.split()[1]}
                    )

        # Finished jobs only stay in squeue for a few mins (configurable). If
        # a job ID isn't there, we'll assume it's finished.
        return {
            j
            for j in job_ids
            if id_to_state.get(j, "COMPLETED") in STATES_FINISHED
        }

    def _mkdir_local_folder(self, folder: str) -> None:
        raise NotImplementedError("Implement in child class.")

    def _mkdir_remote_folder(self, folder: str) -> None:
        raise NotImplementedError("Implement in child class.")

    def _submit_single_sbatch(
        self,
        func,
        slurm_job: SlurmJob,
        slurm_config: SlurmConfig,
    ) -> str:
        logger.info("[_submit_single_sbatch] START")
        # Prepare input pickle(s)
        versions = dict(
            python=sys.version_info[:3],
            cloudpickle=cloudpickle.__version__,
            fractal_server=__VERSION__,
        )
        for task in slurm_job.tasks:
            # Write input pickle
            _args = []
            _kwargs = dict(
                parameters=task.parameters,
                remote_files=task.task_files.remote_files_dict,
            )
            funcser = cloudpickle.dumps((versions, func, _args, _kwargs))
            with open(task.input_pickle_file_local, "wb") as f:
                f.write(funcser)
            logger.info(
                "[_submit_single_sbatch] Written "
                f"{task.input_pickle_file_local=}"
            )

            if self.slurm_runner_type == "ssh":
                # Send input pickle (only relevant for SSH)
                self.fractal_ssh.send_file(
                    local=task.input_pickle_file_local,
                    remote=task.input_pickle_file_remote,
                )
                logger.info(
                    "[_submit_single_sbatch] Transferred "
                    f"{task.input_pickle_file_local=}"
                )

        # Prepare commands to be included in SLURM submission script
        cmdlines = []
        for task in slurm_job.tasks:
            if self.slurm_runner_type == "ssh":
                input_pickle_file = task.input_pickle_file_remote
            else:
                input_pickle_file = task.input_pickle_file_local
            output_pickle_file = task.output_pickle_file_remote
            cmdlines.append(
                (
                    f"{self.python_worker_interpreter}"
                    " -m fractal_server.app.runner."
                    "executors.slurm_common.remote "
                    f"--input-file {input_pickle_file} "
                    f"--output-file {output_pickle_file}"
                )
            )

        # Set ntasks
        num_tasks_max_running = slurm_config.parallel_tasks_per_job
        ntasks = min(len(cmdlines), num_tasks_max_running)
        slurm_config.parallel_tasks_per_job = ntasks

        # Prepare SLURM preamble based on SlurmConfig object
        script_lines = slurm_config.to_sbatch_preamble(
            remote_export_dir=self.user_cache_dir
        )

        # Extend SLURM preamble with variable which are not in SlurmConfig, and
        # fix their order
        script_lines.extend(
            [
                f"#SBATCH --err={slurm_job.slurm_stderr_remote}",
                f"#SBATCH --out={slurm_job.slurm_stdout_remote}",
                f"#SBATCH -D {slurm_job.workdir_remote}",
            ]
        )
        script_lines = slurm_config.sort_script_lines(script_lines)
        logger.info(script_lines)

        # Always print output of `uname -n` and `pwd`
        script_lines.append('\necho "Hostname: $(uname -n)"')
        script_lines.append('echo "Current directory: $(pwd)"')
        script_lines.append(
            'echo "Start time: $(date +"%Y-%m-%dT%H:%M:%S%z")"'
        )

        # Complete script preamble
        script_lines.append("\n")

        # Include command lines
        mem_per_task_MB = slurm_config.mem_per_task_MB
        for cmd in cmdlines:
            script_lines.append(
                "srun --ntasks=1 --cpus-per-task=$SLURM_CPUS_PER_TASK "
                f"--mem={mem_per_task_MB}MB "
                f"{cmd} &"
            )
        script_lines.append("wait\n")
        script = "\n".join(script_lines)
        script_lines.append(
            'echo "End time:   $(date +"%Y-%m-%dT%H:%M:%S%z")"'
        )

        # Write submission script
        with open(slurm_job.slurm_submission_script_local, "w") as f:
            f.write(script)
        logger.info(
            "[_submit_single_sbatch] Written "
            f"{slurm_job.slurm_submission_script_local=}"
        )

        if self.slurm_runner_type == "ssh":
            self.fractal_ssh.send_file(
                local=slurm_job.slurm_submission_script_local,
                remote=slurm_job.slurm_submission_script_remote,
            )
            submit_command = (
                "sbatch --parsable "
                f"{slurm_job.slurm_submission_script_remote}"
            )
        else:
            submit_command = (
                "sbatch --parsable "
                f"{slurm_job.slurm_submission_script_local}"
            )
        # Run sbatch
        pre_submission_cmds = slurm_config.pre_submission_commands
        if len(pre_submission_cmds) == 0:
            logger.info(f"Now run {submit_command=}")
            sbatch_stdout = self._run_remote_cmd(submit_command)
        else:
            logger.info(f"Now using {pre_submission_cmds=}")
            script_lines = pre_submission_cmds + [submit_command]
            wrapper_script_contents = "\n".join(script_lines)
            wrapper_script_contents = f"{wrapper_script_contents}\n"
            if self.slurm_runner_type == "ssh":
                wrapper_script = (
                    f"{slurm_job.slurm_submission_script_remote}_wrapper.sh"
                )
                self.fractal_ssh.write_remote_file(
                    path=wrapper_script, content=wrapper_script_contents
                )
            else:
                wrapper_script = (
                    f"{slurm_job.slurm_submission_script_local}_wrapper.sh"
                )
                with open(wrapper_script, "w") as f:
                    f.write(wrapper_script_contents)
            logger.info(f"Now run {wrapper_script=}")
            sbatch_stdout = self._run_remote_cmd(f"bash {wrapper_script}")

        # Submit SLURM job and retrieve job ID
        logger.info(f"[_submit_single_sbatc] {sbatch_stdout=}")
        stdout = sbatch_stdout.strip("\n")
        submitted_job_id = int(stdout)
        slurm_job.slurm_job_id = str(submitted_job_id)

        # Add job to self.jobs
        self.jobs[slurm_job.slurm_job_id] = slurm_job
        logger.info(
            "[_submit_single_sbatch] Added "
            f"{slurm_job.slurm_job_id} to self.jobs."
        )
        logger.info("[_submit_single_sbatch] END")

    def _fetch_artifacts(
        self,
        finished_slurm_jobs: list[SlurmJob],
    ) -> None:
        raise NotImplementedError("Implement in child class.")

    def _check_slurm_account(self) -> None:
        """
        Check that SLURM account is not set here in `common_script_lines`.
        """
        try:
            invalid_line = next(
                line
                for line in self.common_script_lines
                if line.startswith("#SBATCH --account=")
            )
            raise RuntimeError(
                "Invalid line in `common_script_lines`: "
                f"'{invalid_line}'.\n"
                "SLURM account must be set via the request body of the "
                "apply-workflow endpoint, or by modifying the user properties."
            )
        except StopIteration:
            pass

    def _postprocess_single_task(
        self,
        *,
        task: SlurmTask,
        was_job_scancelled: bool = False,
    ) -> tuple[Any, Exception]:
        try:
            with open(task.output_pickle_file_local, "rb") as f:
                outdata = f.read()
            success, output = cloudpickle.loads(outdata)
            if success:
                # Task succeeded
                result = output
                return (result, None)
            else:
                # Task failed in a controlled way, and produced an `output`
                # object which is a dictionary with required keys
                # `exc_type_name` and `traceback_string` and with optional
                # keys `workflow_task_order`, `workflow_task_id` and
                # `task_name`.
                exc_type_name = output.get("exc_type_name")
                logger.debug(
                    f"Output pickle contains a '{exc_type_name}' exception."
                )
                traceback_string = output.get("traceback_string")
                kwargs = {
                    key: output[key]
                    for key in [
                        "workflow_task_order",
                        "workflow_task_id",
                        "task_name",
                    ]
                    if key in output.keys()
                }
                exception = TaskExecutionError(traceback_string, **kwargs)
                return (None, exception)

        except Exception as e:
            exception = JobExecutionError(f"ERROR, {str(e)}")
            # If job was scancelled and task failed, replace
            # exception with a shutdown-related one.
            if was_job_scancelled:
                logger.debug(
                    "Replacing exception with a shutdown-related one, "
                    f"for {task.index=}."
                )
                exception = SHUTDOWN_EXCEPTION
            return (None, exception)
        finally:
            Path(task.input_pickle_file_local).unlink(missing_ok=True)
            Path(task.output_pickle_file_local).unlink(missing_ok=True)

    def is_shutdown(self) -> bool:
        return self.shutdown_file.exists()

    @property
    def job_ids(self) -> list[str]:
        return list(self.jobs.keys())

    def wait_and_check_shutdown(self) -> list[str]:
        """
        Wait at most `self.poll_interval`, while also checking for shutdown.
        """
        # Sleep for `self.poll_interval`, but keep checking for shutdowns
        start_time = time.perf_counter()
        max_time = start_time + self.poll_interval
        can_return = False
        logger.debug(
            "[wait_and_check_shutdown] "
            f"I will wait at most {self.poll_interval} s, "
            f"in blocks of {self.poll_interval_internal} s."
        )

        while (time.perf_counter() < max_time) or (can_return is False):
            # Handle shutdown
            if self.is_shutdown():
                logger.info("[wait_and_check_shutdown] Shutdown file detected")
                scancelled_job_ids = self.scancel_jobs()
                logger.info(f"[wait_and_check_shutdown] {scancelled_job_ids=}")
                return scancelled_job_ids
            can_return = True
            time.sleep(self.poll_interval_internal)

        logger.debug("[wait_and_check_shutdown] No shutdown file detected")
        return []

    def submit(
        self,
        func: callable,
        parameters: dict[str, Any],
        history_unit_id: int,
        task_files: TaskFiles,
        config: SlurmConfig,
        task_type: Literal[
            "non_parallel",
            "converter_non_parallel",
            "compound",
            "converter_compound",
        ],
    ) -> tuple[Any, Exception]:
        logger.info("[submit] START")

        workdir_local = task_files.wftask_subfolder_local
        workdir_remote = task_files.wftask_subfolder_remote

        if self.jobs != {}:
            raise JobExecutionError("Unexpected branch: jobs should be empty.")

        if self.is_shutdown():
            with next(get_sync_db()) as db:
                update_status_of_history_unit(
                    history_unit_id=history_unit_id,
                    status=HistoryUnitStatus.FAILED,
                    db_sync=db,
                )

            return None, SHUTDOWN_EXCEPTION

        # Validation phase
        self.validate_submit_parameters(
            parameters=parameters,
            task_type=task_type,
        )

        # Create task subfolder
        logger.info("[submit] Create local/remote folders - START")
        self._mkdir_local_folder(folder=workdir_local.as_posix())
        self._mkdir_remote_folder(folder=workdir_remote.as_posix())
        logger.info("[submit] Create local/remote folders - END")

        # Submission phase
        slurm_job = SlurmJob(
            prefix=task_files.prefix,
            workdir_local=workdir_local,
            workdir_remote=workdir_remote,
            tasks=[
                SlurmTask(
                    prefix=task_files.prefix,
                    index=0,
                    component=task_files.component,
                    parameters=parameters,
                    workdir_remote=workdir_remote,
                    workdir_local=workdir_local,
                    task_files=task_files,
                )
            ],
        )

        config.parallel_tasks_per_job = 1
        self._submit_single_sbatch(
            func,
            slurm_job=slurm_job,
            slurm_config=config,
        )
        logger.info(f"[submit] END submission phase, {self.job_ids=}")

        # NOTE: see issue 2444
        settings = Inject(get_settings)
        sleep_time = settings.FRACTAL_SLURM_INTERVAL_BEFORE_RETRIEVAL
        logger.warning(f"[submit] Now sleep {sleep_time} seconds.")
        time.sleep(sleep_time)

        # Retrieval phase
        logger.info("[submit] START retrieval phase")
        scancelled_job_ids = []
        while len(self.jobs) > 0:
            # Look for finished jobs
            finished_job_ids = self._get_finished_jobs(job_ids=self.job_ids)
            logger.debug(f"[submit] {finished_job_ids=}")
            finished_jobs = [
                self.jobs[_slurm_job_id] for _slurm_job_id in finished_job_ids
            ]
            self._fetch_artifacts(finished_jobs)
            with next(get_sync_db()) as db:
                for slurm_job_id in finished_job_ids:
                    logger.debug(f"[submit] Now process {slurm_job_id=}")
                    slurm_job = self.jobs.pop(slurm_job_id)
                    was_job_scancelled = slurm_job_id in scancelled_job_ids
                    result, exception = self._postprocess_single_task(
                        task=slurm_job.tasks[0],
                        was_job_scancelled=was_job_scancelled,
                    )

                    if exception is not None:
                        update_status_of_history_unit(
                            history_unit_id=history_unit_id,
                            status=HistoryUnitStatus.FAILED,
                            db_sync=db,
                        )
                    else:
                        if task_type not in ["compound", "converter_compound"]:
                            update_status_of_history_unit(
                                history_unit_id=history_unit_id,
                                status=HistoryUnitStatus.DONE,
                                db_sync=db,
                            )

            if len(self.jobs) > 0:
                scancelled_job_ids = self.wait_and_check_shutdown()

        logger.info("[submit] END")
        return result, exception

    def multisubmit(
        self,
        func: callable,
        list_parameters: list[dict],
        history_unit_ids: list[int],
        list_task_files: list[TaskFiles],
        task_type: Literal["parallel", "compound", "converter_compound"],
        config: SlurmConfig,
    ) -> tuple[dict[int, Any], dict[int, BaseException]]:
        """
        Note: `list_parameters`, `list_task_files` and `history_unit_ids`
        have the same size. For parallel tasks, this is also the number of
        input images, while for compound tasks these can differ.
        """

        if len(self.jobs) > 0:
            raise RuntimeError(
                f"Cannot run `multisubmit` when {len(self.jobs)=}"
            )

        if self.is_shutdown():
            if task_type == "parallel":
                with next(get_sync_db()) as db:
                    bulk_update_status_of_history_unit(
                        history_unit_ids=history_unit_ids,
                        status=HistoryUnitStatus.FAILED,
                        db_sync=db,
                    )
            results = {}
            exceptions = {
                ind: SHUTDOWN_EXCEPTION for ind in range(len(list_parameters))
            }
            return results, exceptions

        self.validate_multisubmit_parameters(
            list_parameters=list_parameters,
            task_type=task_type,
            list_task_files=list_task_files,
            history_unit_ids=history_unit_ids,
        )

        logger.info(f"[multisubmit] START, {len(list_parameters)=}")

        workdir_local = list_task_files[0].wftask_subfolder_local
        workdir_remote = list_task_files[0].wftask_subfolder_remote

        # Create local&remote task subfolders
        if task_type == "parallel":
            self._mkdir_local_folder(workdir_local.as_posix())
            self._mkdir_remote_folder(folder=workdir_remote.as_posix())

        # Execute tasks, in chunks of size `parallel_tasks_per_job`
        # TODO Pick a data structure for results and exceptions, or review the
        # interface
        results: dict[int, Any] = {}
        exceptions: dict[int, BaseException] = {}

        tot_tasks = len(list_parameters)

        # NOTE: chunking has already taken place in `get_slurm_config`,
        # so that `config.tasks_per_job` is now set.

        # Divide arguments in batches of `tasks_per_job` tasks each
        args_batches = []
        batch_size = config.tasks_per_job
        for ind_chunk in range(0, tot_tasks, batch_size):
            args_batches.append(
                list_parameters[ind_chunk : ind_chunk + batch_size]  # noqa
            )
        if len(args_batches) != math.ceil(tot_tasks / config.tasks_per_job):
            raise RuntimeError("Something wrong here while batching tasks")

        # Part 1/3: Iterate over chunks, prepare SlurmJob objects
        logger.info("[multisubmit] Prepare `SlurmJob`s.")
        jobs_to_submit = []
        for ind_batch, chunk in enumerate(args_batches):
            # Read prefix based on the first task of this batch
            prefix = list_task_files[ind_batch * batch_size].prefix
            tasks = []
            for ind_chunk, parameters in enumerate(chunk):
                index = (ind_batch * batch_size) + ind_chunk
                tasks.append(
                    SlurmTask(
                        prefix=prefix,
                        index=index,
                        component=list_task_files[index].component,
                        workdir_local=workdir_local,
                        workdir_remote=workdir_remote,
                        parameters=parameters,
                        zarr_url=parameters["zarr_url"],
                        task_files=list_task_files[index],
                    ),
                )
            jobs_to_submit.append(
                SlurmJob(
                    prefix=prefix,
                    workdir_local=workdir_local,
                    workdir_remote=workdir_remote,
                    tasks=tasks,
                )
            )

        # NOTE: see issue 2431
        logger.info("[multisubmit] Transfer files and submit jobs.")
        for slurm_job in jobs_to_submit:
            self._submit_single_sbatch(
                func,
                slurm_job=slurm_job,
                slurm_config=config,
            )

        logger.info(f"END submission phase, {self.job_ids=}")

        settings = Inject(get_settings)
        sleep_time = settings.FRACTAL_SLURM_INTERVAL_BEFORE_RETRIEVAL
        logger.warning(f"[submit] Now sleep {sleep_time} seconds.")
        time.sleep(sleep_time)

        # Retrieval phase
        logger.info("[multisubmit] START retrieval phase")
        scancelled_job_ids = []
        while len(self.jobs) > 0:
            # Look for finished jobs
            finished_job_ids = self._get_finished_jobs(job_ids=self.job_ids)
            logger.debug(f"[multisubmit] {finished_job_ids=}")
            finished_jobs = [
                self.jobs[_slurm_job_id] for _slurm_job_id in finished_job_ids
            ]
            self._fetch_artifacts(finished_jobs)

            with next(get_sync_db()) as db:
                for slurm_job_id in finished_job_ids:
                    logger.info(f"[multisubmit] Now process {slurm_job_id=}")
                    slurm_job = self.jobs.pop(slurm_job_id)
                    for task in slurm_job.tasks:
                        logger.info(f"[multisubmit] Now process {task.index=}")
                        was_job_scancelled = slurm_job_id in scancelled_job_ids
                        result, exception = self._postprocess_single_task(
                            task=task,
                            was_job_scancelled=was_job_scancelled,
                        )

                        # Note: the relevant done/failed check is based on
                        # whether `exception is None`. The fact that
                        # `result is None` is not relevant for this purpose.
                        if exception is not None:
                            exceptions[task.index] = exception
                            if task_type == "parallel":
                                update_status_of_history_unit(
                                    history_unit_id=history_unit_ids[
                                        task.index
                                    ],
                                    status=HistoryUnitStatus.FAILED,
                                    db_sync=db,
                                )
                        else:
                            results[task.index] = result
                            if task_type == "parallel":
                                update_status_of_history_unit(
                                    history_unit_id=history_unit_ids[
                                        task.index
                                    ],
                                    status=HistoryUnitStatus.DONE,
                                    db_sync=db,
                                )

            if len(self.jobs) > 0:
                scancelled_job_ids = self.wait_and_check_shutdown()

        logger.info("[multisubmit] END")
        return results, exceptions

    def check_fractal_server_versions(self) -> None:
        """
        Compare fractal-server versions of local/remote Python interpreters.
        """

        # Skip check when the local and remote interpreters are the same
        # (notably for some sudo-slurm deployments)
        if self.python_worker_interpreter == sys.executable:
            return

        # Fetch remote fractal-server version
        cmd = (
            f"{self.python_worker_interpreter} "
            "-m fractal_server.app.runner.versions"
        )
        stdout = self._run_remote_cmd(cmd)
        remote_version = json.loads(stdout.strip("\n"))["fractal_server"]

        # Verify local/remote version match
        if remote_version != __VERSION__:
            error_msg = (
                "Fractal-server version mismatch.\n"
                "Local interpreter: "
                f"({sys.executable}): {__VERSION__}.\n"
                "Remote interpreter: "
                f"({self.python_worker_interpreter}): {remote_version}."
            )
            logger.error(error_msg)
            raise RuntimeError(error_msg)

    def scancel_jobs(self) -> list[str]:
        logger.info("[scancel_jobs] START")

        if self.jobs:
            scancelled_job_ids = self.job_ids
            scancel_string = " ".join(scancelled_job_ids)
            scancel_cmd = f"scancel {scancel_string}"
            logger.warning(f"Now scancel-ing SLURM jobs {scancel_string}")
            try:
                self._run_remote_cmd(scancel_cmd)
            except Exception as e:
                logger.warning(
                    "[scancel_jobs] `scancel` command failed. "
                    f"Original error:\n{str(e)}"
                )
        logger.info("[scancel_jobs] END")
        return scancelled_job_ids

    def validate_slurm_jobs_workdirs(
        self,
        slurm_jobs: list[SlurmJob],
    ) -> None:
        """
        Check that a list of `SlurmJob`s have homogeneous working folders.
        """
        set_workdir_local = set(_job.workdir_local for _job in slurm_jobs)
        set_workdir_remote = set(_job.workdir_remote for _job in slurm_jobs)
        if len(set_workdir_local) > 1:
            raise ValueError(f"Non-unique values in {set_workdir_local=}.")
        if len(set_workdir_remote) > 1:
            raise ValueError(f"Non-unique values in {set_workdir_remote=}.")

_check_slurm_account()

Check that SLURM account is not set here in common_script_lines.

Source code in fractal_server/app/runner/executors/slurm_common/base_slurm_runner.py

def _check_slurm_account(self) -> None:
    """
    Check that SLURM account is not set here in `common_script_lines`.
    """
    try:
        invalid_line = next(
            line
            for line in self.common_script_lines
            if line.startswith("#SBATCH --account=")
        )
        raise RuntimeError(
            "Invalid line in `common_script_lines`: "
            f"'{invalid_line}'.\n"
            "SLURM account must be set via the request body of the "
            "apply-workflow endpoint, or by modifying the user properties."
        )
    except StopIteration:
        pass

check_fractal_server_versions()

Compare fractal-server versions of local/remote Python interpreters.

Source code in fractal_server/app/runner/executors/slurm_common/base_slurm_runner.py

def check_fractal_server_versions(self) -> None:
    """
    Compare fractal-server versions of local/remote Python interpreters.
    """

    # Skip check when the local and remote interpreters are the same
    # (notably for some sudo-slurm deployments)
    if self.python_worker_interpreter == sys.executable:
        return

    # Fetch remote fractal-server version
    cmd = (
        f"{self.python_worker_interpreter} "
        "-m fractal_server.app.runner.versions"
    )
    stdout = self._run_remote_cmd(cmd)
    remote_version = json.loads(stdout.strip("\n"))["fractal_server"]

    # Verify local/remote version match
    if remote_version != __VERSION__:
        error_msg = (
            "Fractal-server version mismatch.\n"
            "Local interpreter: "
            f"({sys.executable}): {__VERSION__}.\n"
            "Remote interpreter: "
            f"({self.python_worker_interpreter}): {remote_version}."
        )
        logger.error(error_msg)
        raise RuntimeError(error_msg)

multisubmit(func, list_parameters, history_unit_ids, list_task_files, task_type, config)

Note: list_parameters, list_task_files and history_unit_ids have the same size. For parallel tasks, this is also the number of input images, while for compound tasks these can differ.

Source code in fractal_server/app/runner/executors/slurm_common/base_slurm_runner.py

def multisubmit(
    self,
    func: callable,
    list_parameters: list[dict],
    history_unit_ids: list[int],
    list_task_files: list[TaskFiles],
    task_type: Literal["parallel", "compound", "converter_compound"],
    config: SlurmConfig,
) -> tuple[dict[int, Any], dict[int, BaseException]]:
    """
    Note: `list_parameters`, `list_task_files` and `history_unit_ids`
    have the same size. For parallel tasks, this is also the number of
    input images, while for compound tasks these can differ.
    """

    if len(self.jobs) > 0:
        raise RuntimeError(
            f"Cannot run `multisubmit` when {len(self.jobs)=}"
        )

    if self.is_shutdown():
        if task_type == "parallel":
            with next(get_sync_db()) as db:
                bulk_update_status_of_history_unit(
                    history_unit_ids=history_unit_ids,
                    status=HistoryUnitStatus.FAILED,
                    db_sync=db,
                )
        results = {}
        exceptions = {
            ind: SHUTDOWN_EXCEPTION for ind in range(len(list_parameters))
        }
        return results, exceptions

    self.validate_multisubmit_parameters(
        list_parameters=list_parameters,
        task_type=task_type,
        list_task_files=list_task_files,
        history_unit_ids=history_unit_ids,
    )

    logger.info(f"[multisubmit] START, {len(list_parameters)=}")

    workdir_local = list_task_files[0].wftask_subfolder_local
    workdir_remote = list_task_files[0].wftask_subfolder_remote

    # Create local&remote task subfolders
    if task_type == "parallel":
        self._mkdir_local_folder(workdir_local.as_posix())
        self._mkdir_remote_folder(folder=workdir_remote.as_posix())

    # Execute tasks, in chunks of size `parallel_tasks_per_job`
    # TODO Pick a data structure for results and exceptions, or review the
    # interface
    results: dict[int, Any] = {}
    exceptions: dict[int, BaseException] = {}

    tot_tasks = len(list_parameters)

    # NOTE: chunking has already taken place in `get_slurm_config`,
    # so that `config.tasks_per_job` is now set.

    # Divide arguments in batches of `tasks_per_job` tasks each
    args_batches = []
    batch_size = config.tasks_per_job
    for ind_chunk in range(0, tot_tasks, batch_size):
        args_batches.append(
            list_parameters[ind_chunk : ind_chunk + batch_size]  # noqa
        )
    if len(args_batches) != math.ceil(tot_tasks / config.tasks_per_job):
        raise RuntimeError("Something wrong here while batching tasks")

    # Part 1/3: Iterate over chunks, prepare SlurmJob objects
    logger.info("[multisubmit] Prepare `SlurmJob`s.")
    jobs_to_submit = []
    for ind_batch, chunk in enumerate(args_batches):
        # Read prefix based on the first task of this batch
        prefix = list_task_files[ind_batch * batch_size].prefix
        tasks = []
        for ind_chunk, parameters in enumerate(chunk):
            index = (ind_batch * batch_size) + ind_chunk
            tasks.append(
                SlurmTask(
                    prefix=prefix,
                    index=index,
                    component=list_task_files[index].component,
                    workdir_local=workdir_local,
                    workdir_remote=workdir_remote,
                    parameters=parameters,
                    zarr_url=parameters["zarr_url"],
                    task_files=list_task_files[index],
                ),
            )
        jobs_to_submit.append(
            SlurmJob(
                prefix=prefix,
                workdir_local=workdir_local,
                workdir_remote=workdir_remote,
                tasks=tasks,
            )
        )

    # NOTE: see issue 2431
    logger.info("[multisubmit] Transfer files and submit jobs.")
    for slurm_job in jobs_to_submit:
        self._submit_single_sbatch(
            func,
            slurm_job=slurm_job,
            slurm_config=config,
        )

    logger.info(f"END submission phase, {self.job_ids=}")

    settings = Inject(get_settings)
    sleep_time = settings.FRACTAL_SLURM_INTERVAL_BEFORE_RETRIEVAL
    logger.warning(f"[submit] Now sleep {sleep_time} seconds.")
    time.sleep(sleep_time)

    # Retrieval phase
    logger.info("[multisubmit] START retrieval phase")
    scancelled_job_ids = []
    while len(self.jobs) > 0:
        # Look for finished jobs
        finished_job_ids = self._get_finished_jobs(job_ids=self.job_ids)
        logger.debug(f"[multisubmit] {finished_job_ids=}")
        finished_jobs = [
            self.jobs[_slurm_job_id] for _slurm_job_id in finished_job_ids
        ]
        self._fetch_artifacts(finished_jobs)

        with next(get_sync_db()) as db:
            for slurm_job_id in finished_job_ids:
                logger.info(f"[multisubmit] Now process {slurm_job_id=}")
                slurm_job = self.jobs.pop(slurm_job_id)
                for task in slurm_job.tasks:
                    logger.info(f"[multisubmit] Now process {task.index=}")
                    was_job_scancelled = slurm_job_id in scancelled_job_ids
                    result, exception = self._postprocess_single_task(
                        task=task,
                        was_job_scancelled=was_job_scancelled,
                    )

                    # Note: the relevant done/failed check is based on
                    # whether `exception is None`. The fact that
                    # `result is None` is not relevant for this purpose.
                    if exception is not None:
                        exceptions[task.index] = exception
                        if task_type == "parallel":
                            update_status_of_history_unit(
                                history_unit_id=history_unit_ids[
                                    task.index
                                ],
                                status=HistoryUnitStatus.FAILED,
                                db_sync=db,
                            )
                    else:
                        results[task.index] = result
                        if task_type == "parallel":
                            update_status_of_history_unit(
                                history_unit_id=history_unit_ids[
                                    task.index
                                ],
                                status=HistoryUnitStatus.DONE,
                                db_sync=db,
                            )

        if len(self.jobs) > 0:
            scancelled_job_ids = self.wait_and_check_shutdown()

    logger.info("[multisubmit] END")
    return results, exceptions

validate_slurm_jobs_workdirs(slurm_jobs)

Check that a list of SlurmJobs have homogeneous working folders.

Source code in fractal_server/app/runner/executors/slurm_common/base_slurm_runner.py

def validate_slurm_jobs_workdirs(
    self,
    slurm_jobs: list[SlurmJob],
) -> None:
    """
    Check that a list of `SlurmJob`s have homogeneous working folders.
    """
    set_workdir_local = set(_job.workdir_local for _job in slurm_jobs)
    set_workdir_remote = set(_job.workdir_remote for _job in slurm_jobs)
    if len(set_workdir_local) > 1:
        raise ValueError(f"Non-unique values in {set_workdir_local=}.")
    if len(set_workdir_remote) > 1:
        raise ValueError(f"Non-unique values in {set_workdir_remote=}.")

wait_and_check_shutdown()

Wait at most self.poll_interval, while also checking for shutdown.

Source code in fractal_server/app/runner/executors/slurm_common/base_slurm_runner.py

def wait_and_check_shutdown(self) -> list[str]:
    """
    Wait at most `self.poll_interval`, while also checking for shutdown.
    """
    # Sleep for `self.poll_interval`, but keep checking for shutdowns
    start_time = time.perf_counter()
    max_time = start_time + self.poll_interval
    can_return = False
    logger.debug(
        "[wait_and_check_shutdown] "
        f"I will wait at most {self.poll_interval} s, "
        f"in blocks of {self.poll_interval_internal} s."
    )

    while (time.perf_counter() < max_time) or (can_return is False):
        # Handle shutdown
        if self.is_shutdown():
            logger.info("[wait_and_check_shutdown] Shutdown file detected")
            scancelled_job_ids = self.scancel_jobs()
            logger.info(f"[wait_and_check_shutdown] {scancelled_job_ids=}")
            return scancelled_job_ids
        can_return = True
        time.sleep(self.poll_interval_internal)

    logger.debug("[wait_and_check_shutdown] No shutdown file detected")
    return []