tensorflow: tensorflow-cpu install from conda-forge fails with 'requires grpc-cpp >=1.39.1,<1.40.0a0, but none of the providers can be installed'

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): 20.04.3 LTS
• TensorFlow installed from (source or binary): binary, tensorflow-cpu latest as of 12 Dec 2021
• TensorFlow version: latest (should install 2.6.2, installs 2.6.0 due to existing packages)
• Python version: 3.9.7
• Installed using virtualenv? pip? conda?: mamba / conda
• GPU model and memory: none, CPU only

Describe the problem

Run following Docker container:

docker container run -it --rm --entrypoint=bash --user=0 \
  jupyter/all-spark-notebook

Run these commands:

cat << EOF > req.txt
jupyterhub
nbgitpuller
pyspark
boto3
elasticsearch
altair
beautifulsoup4
bokeh
bottleneck
cloudpickle
cython
dask
h5py
ipympl
ipywidgets
matplotlib-base
nltk
gensim
numba
numexpr
pandas
patsy
protobuf
pytables
scikit-learn
scikit-image
scipy
seaborn
statsmodels
xlrd
EOF

mamba install -q --file req.txt -y
mamba install -v -y -c conda-forge tensorflow-cpu

Will result in:

info     Problem count: 1
Encountered problems while solving:
  - package tensorflow-base-2.6.0-cpu_py39h7e79a0b_2 requires grpc-cpp >=1.39.1,<1.40.0a0, but none of the providers can be installed

Notice that:

• 2.6.2 is latest version of tensorflow, but this appears to install 2.6.0
• grpc-cpp 1.42.0 is installed already prior to tensorflow-cpu attempted install

So let’s try:

mamba install -v -y -c conda-forge tensorflow-cpu==2.6.2

Result:

info     Problem count: 1
Encountered problems while solving:
  - package grpc-cpp-1.42.0-ha1441d3_1 requires libprotobuf >=3.19.1,<3.20.0a0, but none of the providers can be installed

Funny thing is, libprotobuf 3.19.1 is installed:

mamba list | grep proto
# libprotobuf               3.19.1               h780b84a_0    conda-forge