empty dataset error for benchmark dynamics using gbdt meta

[2young-2simple-sometimes-naive]

🐛 Bug Description

When changing the workflow yaml to most recent (2021-06-10 for test data), and using --sim_task_model="gbdt", reports this data empty error:
raise ValueError("Empty data from dataset, please check your dataset config.")

To Reproduce

Steps to reproduce the behavior:

1. change workflow yaml to the following workflow yaml (only thing changed is the end date of the qlib data). workflow_config_lightgbm_Alpha158.txt

2. python workflow.py --forecast_model="gbdt" --sim_task_model="gbdt" run_all

Environment

Linux
x86_64
Linux-4.18.0-147.el8.x86_64-x86_64-with-glibc2.2.5
#1 SMP Wed Dec 4 21:51:45 UTC 2019

Python version: 3.8.6 (default, Oct 22 2020, 17:03:03)  [GCC 9.3.0]

Qlib version: 0.8.3.99
numpy==1.22.1
pandas==1.3.5
scipy==1.7.3
requests==2.27.1
sacred==0.8.2
python-socketio==5.5.1
redis==4.1.1
python-redis-lock==3.7.0
schedule==1.1.0
cvxpy==1.1.18
hyperopt==0.1.2
fire==0.4.0
statsmodels==0.13.1
xlrd==2.0.1
plotly==5.5.0
matplotlib==3.5.1
tables==3.7.0
pyyaml==5.4.1
mlflow==1.23.0
tqdm==4.62.3
loguru==0.5.3
lightgbm==3.3.2
tornado==6.1
joblib==1.1.0
fire==0.4.0
ruamel.yaml==0.17.20

Additional Notes

output attatched
debug.txt

[you-n-g]

@wendili-cs will help to answer this question.

[wendili-cs]

There are two reasons that may cause this problem.
The first reason is using cache in rolling_benchmark.py (https://github.com/microsoft/qlib/blob/main/examples/benchmarks_dynamic/baseline/rolling_benchmark.py#L55). The code will pickle the handler for future use, now it does not check whether the data is up-to-date, so deleting the pkl file in the baseline folder will solve this problem.
The second reason is the horizon setting. Currently, the default forecasting horizon of rolling_benchmark.py is 20 trading days (https://github.com/microsoft/qlib/blob/main/examples/benchmarks_dynamic/baseline/rolling_benchmark.py#L50). If you want to calculate the label on time t, you need data 20 trading days later on time t, because you need 20 more future data. If the latest available data with 1 horizon day is 2021-06-10, the latest available data with 20 horizon days is 20 trading days earlier, so it should be set to 2021-05-17 in this case.