AWS SageMaker Notebook安装gcc提示环境不一致如何解决

问题场景 #

• 开发环境：AWS SageMaker Jupyter Notebook，选用conda_python3内核
• 开发目标：使用Prophet库实现时间序列预测，安装Prophet前需配置gcc环境
• 故障现象：在Notebook中执行%conda install gcc时，触发环境不一致报错，依赖求解过程长时间卡顿，手动中断后报错如下：

Collecting package metadata (current_repodata.json): done
Solving environment: /
The environment is inconsistent, please check the package plan carefully
The following packages are causing the inconsistency:
- conda-forge/noarch::nbclient==0.5.2=pyhd8ed1ab_0
- conda-forge/linux-64::matplotlib==3.3.4=py36h5fab9bb_0
- conda-forge/noarch::qdarkstyle==2.8.1=pyhd8ed1ab_2
- conda-forge/linux-64::scikit-image==0.16.2=py36hb3f55d8_0
- conda-forge/noarch::python-language-server==0.36.2=pyhd8ed1ab_0
- conda-forge/linux-64::widgetsnbextension==3.5.1=py36h5fab9bb_4
- conda-forge/noarch::flake8==3.8.4=py_0
- conda-forge/noarch::ipywidgets==7.6.3=pyhd3deb0d_0
- conda-forge/noarch::typing-extensions==3.7.4.3=0
- conda-forge/noarch::path.py==12.5.0=0
- conda-forge/noarch::dask==2021.2.0=pyhd8ed1ab_0
- conda-forge/noarch::nbformat==5.1.2=pyhd8ed1ab_1
- conda-forge/linux-64::path==15.1.2=py36h5fab9bb_0
- conda-forge/linux-64::nbconvert==6.0.7=py36h5fab9bb_3
- conda-forge/linux-64::distributed==2021.2.0=py36h5fab9bb_0
- conda-forge/noarch::anaconda-client==1.7.2=py_0
- conda-forge/noarch::aioitertools==0.7.1=pyhd8ed1ab_0
- conda-forge/linux-64::matplotlib-base==3.3.4=py36hd391965_0
- conda-forge/linux-64::pluggy==0.13.1=py36h5fab9bb_4
- conda-forge/noarch::black==20.8b1=py_1
- conda-forge/linux-64::blaze==0.11.3=py36_0
- conda-forge/noarch::pyls-spyder==0.3.2=pyhd8ed1ab_0
- conda-forge/noarch::odo==0.5.1=py_1
- conda-forge/linux-64::keyring==22.0.1=py36h5fab9bb_0
- conda-forge/noarch::anaconda-project==0.9.1=pyhd8ed1ab_0
- conda-forge/noarch::importlib_metadata==3.7.0=hd8ed1ab_0
- conda-forge/linux-64::jupyter==1.0.0=py36h5fab9bb_6
- conda-forge/noarch::jupyterlab_server==2.3.0=pyhd8ed1ab_0
- conda-forge/noarch::seaborn-base==0.11.1=pyhd8ed1ab_1
- conda-forge/noarch::imageio==2.9.0=py_0
- conda-forge/noarch::numpydoc==1.1.0=py_1
- conda-forge/linux-64::yarl==1.6.3=py36h8f6f2f9_1
- conda-forge/noarch::jsonschema==3.2.0=py_2
- conda-forge/noarch::flask==1.1.2=pyh9f0ad1d_0
- conda-forge/noarch::seaborn==0.11.1=hd8ed1ab_1
- conda-forge/noarch::helpdev==0.7.1=pyhd8ed1ab_0
- conda-forge/linux-64::nb_conda==2.2.1=py36h5fab9bb_4
- conda-forge/noarch::nbclassic==0.2.6=pyhd8ed1ab_0
- conda-forge/noarch::sphinx==3.5.1=pyhd8ed1ab_0
- conda-forge/noarch::jupyterlab_launcher==0.13.1=py_2
- conda-forge/linux-64::spyder==4.2.0=py36h5fab9bb_0
- conda-forge/linux-64::importlib-metadata==3.7.0=py36h5fab9bb_0
- conda-forge/linux-64::pytest==6.2.2=py36h5fab9bb_0
- conda-forge/noarch::pyls-black==0.4.6=pyh9f0ad1d_0
/ ^C
failed with initial frozen solve. Retrying with flexible solve.
CondaError: KeyboardInterrupt
Note: you may need to restart the kernel to use updated packages.

故障根因 #

报错列出的所有冲突包全是Python3.6版本的依赖，说明当前选用的conda_python3是SageMaker预置的老旧Python3.6环境，出厂时就混合了conda默认源和conda-forge源的包，依赖树已经损坏。直接执行conda install gcc会触发全量依赖一致性校验，不仅求解速度极慢，还会报环境不一致错误。
另外SageMaker底层的Amazon Linux系统本身已经预装了gcc环境，完全满足Prophet的编译要求，不需要重复通过conda安装gcc。

解决方法 #

按优先级从高到低选即可：

方案1：切换到新版预置内核（最省事，零冲突） #

老版py36环境的混源问题是出厂自带的，没必要硬修：

• 关闭当前Notebook，在启动页选择Python3.10及以上版本的conda_python3内核新建Notebook
• 新环境已经预装了完整的编译工具链，不需要单独装gcc，直接执行%conda install -c conda-forge prophet -y即可完成安装，重启内核后就能正常导入使用

方案2：复用系统gcc，跳过conda依赖校验（适合必须用老环境的场景） #

如果不能切换当前py36内核，不用在conda里装gcc：

1. 先执行!gcc --version验证系统gcc存在，正常会输出版本号信息
2. 跳过conda严格依赖校验直接安装Prophet：

   %conda install -c conda-forge prophet --no-deps -y
   

3. 补装Prophet运行必需的Python依赖：

   %pip install pystan==2.19.1.1 numpy pandas matplotlib
   

4. 重启内核即可正常使用Prophet

方案3：修复老环境依赖（耗时长，不推荐） #

如果一定要在当前老环境里通过conda安装gcc，先统一源再修复依赖：

1. 先配置conda仅使用conda-forge源，开启严格通道优先级：

   !conda config --remove-key channels
   !conda config --add channels conda-forge
   !conda config --set channel_priority strict
   

2. 执行全量环境更新修复损坏的依赖树，过程大概10-15分钟，不要手动中断：

   %conda update --all -y
   

3. 修复完成后再执行%conda install gcc -y就不会报环境不一致错误

避坑提示 #

不要在老版py36的conda环境里强行求解依赖，很容易把Jupyter本身的依赖搞坏，导致内核连接失败、Notebook无法使用的问题。

内容的提问来源于stack exchange，提问作者oligato