Dockerfile

FROM continuumio/miniconda3:4.12.0 AS env-builder
SHELL ["/bin/bash", "-c"]

ARG MLSERVER_ENV_NAME="mlserver-custom-env" \
    MLSERVER_ENV_TARBALL="./envs/base.tar.gz"

RUN conda config --add channels conda-forge && \
    conda install conda-pack

# The `[]` character range will ensure that Docker doesn't complain if the
# files don't exist:
# https://stackoverflow.com/a/65138098/5015573
COPY \
    ./environment.ym[l] \
    ./environment.yam[l] \
    ./conda.ym[l] \
    ./conda.yam[l] \
    .
RUN mkdir $(dirname $MLSERVER_ENV_TARBALL); \
    for envFile in environment.yml environment.yaml conda.yml conda.yaml; do \
        if [[ -f $envFile ]]; then \
            conda env create                 --name $MLSERVER_ENV_NAME \
                --file $envFile; \
            conda-pack --ignore-missing-files --quiet                 -n $MLSERVER_ENV_NAME \
                -o $MLSERVER_ENV_TARBALL; \
        fi \
    done; \
    chmod -R 776 $(dirname $MLSERVER_ENV_TARBALL)

FROM seldonio/mlserver:1.3.5-slim
SHELL ["/bin/bash", "-c"]

# Copy all potential sources for custom environments
COPY \
    --chown=1000 \
    --from=env-builder \
    /envs/base.tar.g[z] \
    ./envs/base.tar.gz
COPY \
    ./settings.jso[n] \
    ./model-settings.jso[n] \
    ./requirements.tx[t] \
    .

USER root
# Install dependencies system-wide, to ensure that they are available for every
# user and give permissions to (future) environment folder.
RUN ./hack/build-env.sh . && \
    mkdir -p ./envs/base && \
    chown -R 1000:0 ./envs/base && \
    chmod -R 776 ./envs/base && \
    rm -rf /root/.cache/pip
USER 1000

# Copy everything else
COPY . .

# Override MLServer's own `CMD` to activate the embedded environment
# (optionally activating the hot-loaded one as well).
CMD source ./hack/activate-env.sh ./envs/base.tar.gz && \
    mlserver start $MLSERVER_MODELS_DIR