diff --git a/qlib/model/ens/__init__.py b/qlib/model/ens/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/qlib/workflow/task/__init__.py b/qlib/workflow/task/__init__.py
new file mode 100644
index 000000000..cc338cca4
--- /dev/null
+++ b/qlib/workflow/task/__init__.py
@@ -0,0 +1,13 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+"""
+Task related workflow is implemented in this folder
+
+A typical task workflow
+
+| Step                  | Description                                    |
+|-----------------------+------------------------------------------------|
+| TaskGen               | Generating tasks.                              |
+| TaskManager(optional) | Manage generated tasks                         |
+| run task              | retrive  tasks from TaskManager and run tasks. |
+"""
diff --git a/qlib/workflow/task/collect.py b/qlib/workflow/task/collect.py
new file mode 100644
index 000000000..13b5869de
--- /dev/null
+++ b/qlib/workflow/task/collect.py
@@ -0,0 +1,52 @@
+from qlib.workflow import R
+import pandas as pd
+from typing import Union
+from tqdm.auto import tqdm
+
+
+class RollingEnsemble:
+    '''
+    Rolling Models Ensemble based on (R)ecord
+
+    This shares nothing with Ensemble
+    '''
+    # TODO: 这边还可以加加速
+    def __init__(self, get_key_func, flt_func=None):
+        self.get_key_func = get_key_func
+        self.flt_func = flt_func
+
+    def __call__(self, exp_name) -> Union[pd.Series, dict]:
+        # TODO;
+        # Should we split the scripts into several sub functions?
+        exp = R.get_exp(experiment_name=exp_name)
+
+        # filter records
+        recs = exp.list_recorders()
+
+        recs_flt = {}
+        for rid, rec in tqdm(recs.items(), desc="Loading data"):
+            # rec = exp.get_recorder(recorder_id=rid)
+            params = rec.load_object("param")
+            if rec.status == rec.STATUS_FI:
+                if self.flt_func is None or self.flt_func(params):
+                    rec.params = params
+                    recs_flt[rid] = rec
+
+        # group
+        recs_group = {}
+        for _, rec in recs_flt.items():
+            params = rec.params
+            group_key = self.get_key_func(params)
+            recs_group.setdefault(group_key, []).append(rec)
+
+        # reduce group
+        reduce_group = {}
+        for k, rec_l in recs_group.items():
+            pred_l = []
+            for rec in rec_l:
+                pred_l.append(rec.load_object('pred.pkl').iloc[:, 0])
+            pred = pd.concat(pred_l).sort_index()
+            reduce_group[k] = pred
+
+        return reduce_group
+
diff --git a/qlib/workflow/task/gen.py b/qlib/workflow/task/gen.py
new file mode 100644
index 000000000..66529f3a5
--- /dev/null
+++ b/qlib/workflow/task/gen.py
@@ -0,0 +1,133 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+'''
+this is a task generator
+'''
+import abc
+import copy
+import typing
+from .utils import TimeAdjuster
+
+
+class TaskGen(metaclass=abc.ABCMeta):
+    @abc.abstractmethod
+    def __call__(self, *args, **kwargs) -> typing.List[dict]:
+        """
+        generate
+
+        Parameters
+        ----------
+        args, kwargs:
+            The info for generating tasks
+            Example 1):
+                input: a specific task template
+                output: rolling version of the tasks
+            Example 2):
+                input: a specific task template
+                output: a set of tasks with different losses
+
+        Returns
+        -------
+        typing.List[dict]:
+            A list of tasks
+        """
+        pass
+
+
+class RollingGen(TaskGen):
+
+    ROLL_EX = TimeAdjuster.SHIFT_EX
+    ROLL_SD = TimeAdjuster.SHIFT_SD
+
+    def __init__(self, step: int = 40, rtype: str = ROLL_EX):
+        """
+        Generate tasks for rolling
+
+        Parameters
+        ----------
+        step : int
+            step to rolling
+        rtype : str
+            rolling type (expanding, rolling)
+        """
+        self.step = step
+        self.rtype = rtype
+        self.ta = TimeAdjuster(future=True)  # 为了保证test最后的日期不是None, 所以这边要改一改
+
+        self.test_key = 'test'
+        self.train_key = 'train'
+
+    def __call__(self, task: dict):
+        """
+        Converting the task into a rolling task
+
+        Parameters
+        ----------
+        task : dict
+            A dict describing a task. For example.
+
+            DEFAULT_TASK = {
+                "model": {
+                    "class": "LGBModel",
+                    "module_path": "qlib.contrib.model.gbdt",
+                },
+                "dataset": {
+                    "class": "DatasetH",
+                    "module_path": "qlib.data.dataset",
+                    "kwargs": {
+                        "handler": {
+                            "class": "Alpha158",
+                            "module_path": "qlib.contrib.data.handler",
+                            "kwargs": data_handler_config,
+                        },
+                        "segments": {
+                            "train": ("2008-01-01", "2014-12-31"),
+                            "valid": ("2015-01-01", "2016-12-20"),  # Please avoid leaking the future test data into validation
+                            "test": ("2017-01-01", "2020-08-01"),
+                        },
+                    },
+                },
+                # You shoud record the data in specific sequence
+                # "record": ['SignalRecord', 'SigAnaRecord', 'PortAnaRecord'],
+            }
+        """
+        res = []
+
+        prev_seg = None
+        test_end = None
+        while True:
+            t = copy.deepcopy(task)
+
+            # calculate segments
+            if prev_seg is None:
+                # First rolling
+                # 1) prepare the end porint
+                segments = copy.deepcopy(self.ta.align_seg(t['dataset']['kwargs']['segments']))
+                test_end = self.ta.max() if segments[self.test_key][1] is None else segments[self.test_key][1]
+                # 2) and the init test segments
+                test_start_idx = self.ta.align_idx(segments[self.test_key][0])
+                segments[self.test_key] = (self.ta.get(test_start_idx), self.ta.get(test_start_idx + self.step - 1))
+            else:
+                segments = {}
+                try:
+                    for k, seg in prev_seg.items():
+                        # 决定怎么shift
+                        if k == self.train_key and self.rtype == self.ROLL_EX:
+                            rtype = self.ta.SHIFT_EX
+                        else:
+                            rtype = self.ta.SHIFT_SD
+                        # 整段数据做shift
+                        segments[k] = self.ta.shift(seg, step=self.step, rtype=rtype)
+                    if segments[self.test_key][0] > test_end:
+                        break
+                except  KeyError:
+                    # We reach the end of tasks
+                    # No more rolling
+                    break
+
+            t['dataset']['kwargs']['segments'] = copy.deepcopy(segments)
+            prev_seg = segments
+            res.append(t)
+        return res
+
+
diff --git a/qlib/workflow/task/manage.py b/qlib/workflow/task/manage.py
new file mode 100644
index 000000000..6407279f0
--- /dev/null
+++ b/qlib/workflow/task/manage.py
@@ -0,0 +1,290 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+"""
+A task consists of 2 parts
+- tasks description: the desc will define the task
+- tasks status: the status of the task
+- tasks result information : A user can get the task with the task description and task result.
+
+"""
+from bson.binary import Binary
+import pickle
+from pymongo.errors import InvalidDocument
+from fire import Fire
+from bson.objectid import ObjectId
+from contextlib import contextmanager
+from loguru import logger
+from tqdm.cli import tqdm
+import time
+import concurrent
+import pymongo
+from qlib.config import C
+
+
+class TaskManager:
+    """TaskManager
+    here is the what will a task looks like
+    {
+        'def': pickle serialized task definition.  using pickle will make it easier
+        'filter': json-like data. This is for filtering the tasks.
+        'status': 'waiting' | 'running' | 'done'
+        'res': pickle serialized task result,
+    }
+
+    The tasks manager assume that you will only update the tasks you fetched.
+    The mongo fetch one and update will make it date updating secure.
+
+    Usage Examples from the CLI.
+    python -m blocks.tasks.__init__ task_stat --task_pool meta_task_rule
+
+
+    NOTE:
+    - 假设： 存储在db里面的都是encode过的， 拿出来的都是decode过的
+    """
+    STATUS_WAITING = 'waiting'
+    STATUS_RUNNING = 'running'
+    STATUS_DONE = 'done'
+    STATUS_PART_DONE = 'part_done'
+
+    ENCODE_FIELDS_PREFIX = ['def', 'res']
+
+    def __init__(self, task_pool=None):
+        self.mdb = get_mongodb()
+        self.task_pool = task_pool
+
+    def list(self):
+        return self.mdb.list_collection_names()
+
+    def _encode_task(self, task):
+        for prefix in self.ENCODE_FIELDS_PREFIX:
+            for k in list(task.keys()):
+                if k.startswith(prefix):
+                    task[k] = Binary(pickle.dumps(task[k]))
+        return task
+
+    def _decode_task(self, task):
+        for prefix in self.ENCODE_FIELDS_PREFIX:
+            for k in list(task.keys()):
+                if k.startswith(prefix):
+                    task[k] = pickle.loads(task[k])
+        return task
+
+    def _get_task_pool(self, task_pool=None):
+        if task_pool is None:
+            task_pool = self.task_pool
+        if task_pool is None:
+            raise ValueError('You must specify a task pool.')
+        if isinstance(task_pool, str):
+            return getattr(self.mdb, task_pool)
+        return task_pool
+
+    def _dict_to_str(self, flt):
+        return {k: str(v) for k, v in flt.items()}
+
+    def replace_task(self, task, new_task, task_pool=None):
+        # 这里的假设是从接口拿出来的都是decode过的，在接口内部的都是 encode过的
+        new_task = self._encode_task(new_task)
+        task_pool = self._get_task_pool(task_pool)
+        query = {'_id': ObjectId(task['_id'])}
+        try:
+            task_pool.replace_one(query, new_task)
+        except InvalidDocument:
+            task['filter'] = self._dict_to_str(task['filter'])
+            task_pool.replace_one(query, new_task)
+
+    def insert_task(self, task, task_pool=None):
+        task_pool = self._get_task_pool(task_pool)
+        try:
+            task_pool.insert_one(task)
+        except InvalidDocument:
+            task['filter'] = self._dict_to_str(task['filter'])
+            task_pool.insert_one(task)
+
+    def insert_task_def(self, task_def, task_pool=None):
+        task_pool = self._get_task_pool(task_pool)
+        task = self._encode_task({
+            'def': task_def,
+            'filter': task_def,  # FIXME: catch the raised error
+            'status': self.STATUS_WAITING,
+        })
+        self.insert_task(task, task_pool)
+
+    def create_task(self, task_def_l, task_pool=None, dry_run=False, print_nt=False):
+        task_pool = self._get_task_pool(task_pool)
+        new_tasks = []
+        for t in task_def_l:
+            try:
+                r = task_pool.find_one({'filter': t})
+            except InvalidDocument:
+                r = task_pool.find_one({'filter': self._dict_to_str(t)})
+            if r is None:
+                new_tasks.append(t)
+        print("Total Tasks, New Tasks:", len(task_def_l), len(new_tasks))
+
+        if print_nt:  # print new task
+            for t in new_tasks:
+                print(t)
+
+        if dry_run:
+            return
+
+        for t in new_tasks:
+            self.insert_task_def(t, task_pool)
+
+    def fetch_task(self, query={}, task_pool=None):
+        task_pool = self._get_task_pool(task_pool)
+        query = query.copy()
+        if '_id' in query:
+            query['_id'] = ObjectId(query['_id'])
+        query.update({'status': self.STATUS_WAITING})
+        task = task_pool.find_one_and_update(query, {'$set': {
+            'status': self.STATUS_RUNNING
+        }},
+                                             sort=[('priority', pymongo.DESCENDING)])
+        # 这里我的 priority 必须是 高数优先级更高，因为 null会被在 ASCENDING时被排在最前面
+        if task is None:
+            return None
+        task['status'] = self.STATUS_RUNNING
+        return self._decode_task(task)
+
+    @contextmanager
+    def safe_fetch_task(self, query={}, task_pool=None):
+        task = self.fetch_task(query=query, task_pool=task_pool)
+        try:
+            yield task
+        except Exception:
+            if task is not None:
+                logger.info('Returning task before raising error')
+                self.return_task(task)
+                logger.info('Task returned')
+            raise
+
+    def task_fetcher_iter(self, query={}, task_pool=None):
+        while True:
+            with self.safe_fetch_task(query=query, task_pool=task_pool) as task:
+                if task is None:
+                    break
+                yield task
+
+    def query(self, query={}, decode=True, task_pool=None):
+        """query
+        This function may raise exception `pymongo.errors.CursorNotFound: cursor id not found` if it takes too long to iterate the generator
+
+        :param query:
+        :param decode:
+        :param task_pool:
+        """
+        query = query.copy()
+        if '_id' in query:
+            query['_id'] = ObjectId(query['_id'])
+        task_pool = self._get_task_pool(task_pool)
+        for t in task_pool.find(query):
+            yield self._decode_task(t)
+
+    def commit_task_res(self, task, res, status=None, task_pool=None):
+        task_pool = self._get_task_pool(task_pool)
+        # A workaround to use the class attribute.
+        if status is None:
+            status = TaskManager.STATUS_DONE
+        task_pool.update_one({"_id": task['_id']}, {'$set': {'status': status, 'res': Binary(pickle.dumps(res))}})
+
+    def return_task(self, task, status=None, task_pool=None):
+        task_pool = self._get_task_pool(task_pool)
+        if status is None:
+            status = TaskManager.STATUS_WAITING
+        update_dict = {'$set': {'status': status}}
+        task_pool.update_one({"_id": task['_id']}, update_dict)
+
+    def remove(self, query={}, task_pool=None):
+        query = query.copy()
+        task_pool = self._get_task_pool(task_pool)
+        if '_id' in query:
+            query['_id'] = ObjectId(query['_id'])
+        task_pool.delete_many(query)
+
+    def task_stat(self, query={}, task_pool=None):
+        query = query.copy()
+        if '_id' in query:
+            query['_id'] = ObjectId(query['_id'])
+        tasks = self.query(task_pool=task_pool, query=query, decode=False)
+        status_stat = {}
+        for t in tasks:
+            status_stat[t['status']] = status_stat.get(t['status'], 0) + 1
+        return status_stat
+
+    def reset_waiting(self, query={}, task_pool=None):
+        query = query.copy()
+        # default query
+        if 'status' not in query:
+            query['status'] = self.STATUS_RUNNING
+        return self.reset_status(query=query, status=self.STATUS_WAITING, task_pool=task_pool)
+
+    def reset_status(self, query, status, task_pool=None):
+        query = query.copy()
+        task_pool = self._get_task_pool(task_pool)
+        if '_id' in query:
+            query['_id'] = ObjectId(query['_id'])
+        print(task_pool.update_many(query, {"$set": {"status": status}}))
+
+    def _get_undone_n(self, task_stat):
+        return task_stat.get(self.STATUS_WAITING, 0) + task_stat.get(self.STATUS_RUNNING, 0)
+
+    def _get_total(self, task_stat):
+        return sum(task_stat.values())
+
+    def wait(self, query={}, task_pool=None):
+        task_stat = self.task_stat(query, task_pool)
+        total = self._get_total(task_stat)
+        last_undone_n = self._get_undone_n(task_stat)
+        with tqdm(total=total, initial=total - last_undone_n) as pbar:
+            while True:
+                time.sleep(10)
+                undone_n = self._get_undone_n(self.task_stat(query, task_pool))
+                pbar.update(last_undone_n - undone_n)
+                last_undone_n = undone_n
+                if undone_n == 0:
+                    break
+
+    def __str__(self):
+        return f"TaskManager({self.task_pool})"
+
+
+def run_task(task_func, task_pool, force_release=False, *args, **kwargs):
+    """run_task.
+    While task pool is not empty, use task_func to fetch and run tasks in task_pool
+
+    Parameters
+    ----------
+    task_func : def (task_def, *args, **kwargs) -> <res which will be committed>
+        the function to run the task
+    task_pool :
+        The name of the task pool
+    force_release :
+        will the program force to release the resource
+    args :
+        args
+    kwargs :
+        kwargs
+    """
+    tm = TaskManager(task_pool)
+
+    ever_run = False
+
+    while True:
+        with tm.safe_fetch_task() as task:
+            if task is None:
+                break
+            logger.info(task['def'])
+            if force_release:
+                with concurrent.futures.ProcessPoolExecutor(max_workers=1) as executor:
+                    res = executor.submit(task_func, task['def'], *args, **kwargs).result()
+            else:
+                res = task_func(task['def'], *args, **kwargs)
+            tm.commit_task_res(task, res)
+            ever_run = True
+
+    return ever_run
+
+
+if __name__ == '__main__':
+    Fire(TaskManager)
diff --git a/qlib/workflow/task/utils.py b/qlib/workflow/task/utils.py
new file mode 100644
index 000000000..3d8fe8996
--- /dev/null
+++ b/qlib/workflow/task/utils.py
@@ -0,0 +1,134 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+import bisect
+import pandas as pd
+from qlib.data import D
+from qlib.config import C
+from qlib.log import get_module_logger
+from pymongo import MongoClient
+
+
+def get_mongodb():
+    try:
+        cfg = C['mongo']
+    except KeyError:
+        get_module_logger("task").error("Please configure `C['mongo']` before using TaskManager")
+        raise
+
+    client = MongoClient(cfg['task_url'])
+    return client.get_database(name=cfg['task_db_name'])
+
+
+class TimeAdjuster:
+    '''找到合适的日期，然后adjust date'''
+    def __init__(self, future=False):
+        self.cals = D.calendar(future=future)
+
+    def get(self, idx: int):
+        """
+        Get datetime by index
+
+        Parameters
+        ----------
+        idx : int
+            index of the calendar
+        """
+        if idx >= len(self.cals):
+            return None
+        return self.cals[idx]
+
+    def max(self):
+        """
+        Return return the max calendar date
+        """
+        return max(self.cals)
+
+    def align_idx(self, time_point, tp_type="start"):
+        time_point = pd.Timestamp(time_point)
+        if tp_type == 'start':
+            idx = bisect.bisect_left(self.cals, time_point)
+        elif tp_type == 'end':
+            idx = bisect.bisect_right(self.cals, time_point) - 1
+        else:
+            raise NotImplementedError(f"This type of input is not supported")
+        return idx
+
+    def align_time(self, time_point, tp_type="start"):
+        """
+        Align a timepoint to calendar  weekdays
+
+        Parameters
+        ----------
+        time_point :
+            Time point
+        tp_type : str
+            time point type (`"start"`, `"end"`)
+        """
+        return self.cals[self.align_idx(time_point, tp_type=tp_type)]
+
+    def align_seg(self, segment):
+        if isinstance(segment, dict):
+            return {k: self.align_seg(seg) for k, seg in segment.items()}
+        elif isinstance(segment, tuple):
+            return self.align_time(segment[0], tp_type="start"), self.align_time(segment[1], tp_type="end")
+        else:
+            raise NotImplementedError(f"This type of input is not supported")
+
+    def truncate(self, segment, test_start, days: int):
+        """
+        truncate the segment based on the test_start date
+
+        Parameters
+        ----------
+        segment :
+            time segment
+        days : int
+            The trading days to be truncated
+            大部分情况是因为这个时间段的数据(一般是特征)会用到 `days` 天的数据
+        """
+        test_idx = self.align_idx(test_start)
+        if isinstance(segment, tuple):
+            new_seg = []
+            for time_point in segment:
+                tp_idx = min(self.align_idx(time_point), test_idx - days)
+                assert (tp_idx > 0)
+                new_seg.append(self.get(tp_idx))
+            return tuple(new_seg)
+        else:
+            raise NotImplementedError(f"This type of input is not supported")
+
+    SHIFT_SD = "sliding"
+    SHIFT_EX = "expanding"
+
+    def shift(self, seg, step: int, rtype=SHIFT_SD):
+        """
+        shift the datatiem of segment
+
+        Parameters
+        ----------
+        seg :
+            datetime segment
+        step : int
+            rolling step
+        rtype : str
+            rolling type ("sliding" or "expanding")
+
+        Raises
+        ------
+        KeyError:
+            shift will raise error if the index(both start and end) is out of self.cal
+        """
+        if isinstance(seg, tuple):
+            start_idx, end_idx = self.align_idx(seg[0], tp_type="start"), self.align_idx(seg[1], tp_type="end")
+            if rtype == self.SHIFT_SD:
+                start_idx += step
+                end_idx += step
+            elif rtype == self.SHIFT_EX:
+                end_idx += step
+            else:
+                raise NotImplementedError(f"This type of input is not supported")
+            if start_idx > len(self.cals):
+                raise KeyError("The segment is out of valid calendar")
+            return self.get(start_idx), self.get(end_idx)
+        else:
+            raise NotImplementedError(f"This type of input is not supported")
diff --git a/setup.py b/setup.py
index f759945fd..c83d092a3 100644
--- a/setup.py
+++ b/setup.py
@@ -55,6 +55,7 @@ REQUIRED = [
     "tornado",
     "joblib>=0.17.0",
     "ruamel.yaml>=0.16.12",
+    "pymongo==3.7.2",  # For task management
 ]
 
 # Numpy include