Use custom containers

Managed Service for Apache Spark serverless runs workloads within Docker containers. The container provides the runtime environment for the workload's driver and executor processes. By default, Managed Service for Apache Spark uses a container image that includes the default Spark, Java, Python and R packages associated with a runtime release version. The Managed Service for Apache Spark batches API lets you use a custom container image instead of the default image. Typically, a custom container image adds Spark workload Java or Python dependencies not provided by the default container image. Important: Do not include Spark in your custom container image; Managed Service for Apache Spark will mount Spark into the container at runtime.

Submit a Spark batch workload using a custom container image

gcloud

Use the gcloud dataproc batches submit spark command with the --container-image flag to specify your custom container image when you submit a Spark batch workload.

gcloud dataproc batches submit spark \
    --container-image=custom-image, for example, "gcr.io/my-project-id/my-image:1.0.1" \
    --region=region \
    --jars=path to user workload jar located in Cloud Storage or included in the custom container \
    --class=The fully qualified name of a class in the jar file, such as org.apache.spark.examples.SparkPi \
    -- add any workload arguments here

Notes:

  • Custom-image: Specify the custom container image using the following Container Registry image naming format: {hostname}/{project-id}/{image}:{tag}, for example, "gcr.io/my-project-id/my-image:1.0.1". Note: You must host your custom container image on Container Registry or Artifact Registry (GAR). (Managed Service for Apache Spark cannot fetch containers from other registries).
  • --jars: Specify a path to a user workload included in your custom container image or located in Cloud Storage, for example, file:///opt/spark/jars/spark-examples.jar or gs://my-bucket/spark/jars/spark-examples.jar.
  • Other batches command options: You can add other optional batches command flags, for example, to use a Persistent History Server (PHS). Note: The PHS must be located in the region where you run batch workloads.
  • workload arguments: You can add any workload arguments by adding a "--" to the end of the command, followed by the workload arguments.

REST

The custom container image is provided through the RuntimeConfig.containerImage field as part of a batches.create API request.

The following example shows how to use a custom container to submit a batch workload using the Managed Service for Apache Spark batches.create API.

Before using any of the request data, make the following replacements:

  • project-id: Google Cloud project ID
  • region: region
  • custom-container-image: Specify the custom container image using the following Container Registry image naming format: {hostname}/{project-id}/{image}:{tag}, for example, "gcr.io/my-project-id/my-image:1.0.1". Note: You must host your custom container on Container Registry or Artifact Registry . (Managed Service for Apache Spark cannot fetch containers from other registries).
  • jar-uri: Specify a path to a workload jar included in your custom container image or located in Cloud Storage, for example, "/opt/spark/jars/spark-examples.jar" or "gs:///spark/jars/spark-examples.jar".
  • class: The fully qualified name of a class in the jar file, such as "org.apache.spark.examples.SparkPi".
  • Other options: You can use other batch workload resource fields, for example, use the sparkBatch.args field to pass arguments to your workload (see the Batch resource documentation for more information). To use a Persistent History Server (PHS), see Setting up a Persistent History Server. Note: The PHS must be located in the region where you run batch workloads.

HTTP method and URL: 

POST https://dataproc.googleapis.com/v1/projects/project-id/locations/region/batches

Request JSON body: 

{
  "runtimeConfig":{
    "containerImage":"custom-container-image
  },
  "sparkBatch":{
    "jarFileUris":[
      "jar-uri"
    ],
    "mainClass":"class"
  }
}

To send your request, expand one of these options:

You should receive a JSON response similar to the following:

{
"name":"projects/project-id/locations/region/batches/batch-id",
  "uuid":",uuid",
  "createTime":"2021-07-22T17:03:46.393957Z",
  "runtimeConfig":{
    "containerImage":"gcr.io/my-project/my-image:1.0.1"
  },
  "sparkBatch":{
    "mainClass":"org.apache.spark.examples.SparkPi",
    "jarFileUris":[
      "/opt/spark/jars/spark-examples.jar"
    ]
  },
  "runtimeInfo":{
    "outputUri":"gs://dataproc-.../driveroutput"
  },
  "state":"SUCCEEDED",
  "stateTime":"2021-07-22T17:06:30.301789Z",
  "creator":"account-email-address",
  "runtimeConfig":{
    "properties":{
      "spark:spark.executor.instances":"2",
      "spark:spark.driver.cores":"2",
      "spark:spark.executor.cores":"2",
      "spark:spark.app.name":"projects/project-id/locations/region/batches/batch-id"
    }
  },
  "environmentConfig":{
    "peripheralsConfig":{
      "sparkHistoryServerConfig":{
      }
    }
  },
  "operation":"projects/project-id/regions/region/operation-id"
}

Build a custom container image

Managed Service for Apache Spark custom container images are Docker images. You can use the tools for building Docker images to build custom container images, but there are conditions the images must meet to be compatible with Managed Service for Apache Spark. The following sections explain these conditions.

Operating system

You can choose any operating system base image for your custom container image.

Recommendation: Use the default Debian 12 images, for example, debian:12-slim, since they have been tested to avoid compatibility issues.

Utilities

You must include the following utility packages, which are required to run Spark, in your custom container image:

  • procps
  • tini

To run XGBoost from Spark (Java or Scala), you must include libgomp1

Container user

Managed Service for Apache Spark runs containers as the spark Linux user with a 1099 UID and a 1099 GID. USER directives set in custom container image Dockerfiles are ignored at runtime. Use the UID and GID for file system permissions. For example, if you add a jar file at /opt/spark/jars/my-lib.jar in the image as a workload dependency, you must give the spark user read permission to the file.

Image streaming

Managed Service for Apache Spark normally begins a workload requiring a custom container image by downloading the entire image to disk. This can mean a delay in initialization time, especially for customers with large images.

You can instead use image streaming, which is a method to pull image data on an as-needed basis. This lets the workload start up without waiting for the entire image to download, potentially improving initialization time. To enable image streaming, you must enable the Container File System API. You must also store your container images in Artifact Registry, and the Artifact Registry repository must be in the same region as your Managed Service for Apache Spark workload or in a multi-region that corresponds with the region where your workload is running. If Managed Service for Apache Spark does not support the image or the image streaming service is not available, our streaming implementation downloads the entire image.

Note that we don't support the following for image streaming:

In these cases, Managed Service for Apache Spark pulls the entire image before starting the workload.

Additional requirements

  • If VPC Service Controls protects your container images and you use Image streaming, you must also include the Image streaming API (containerfilesystem.googleapis.com) in the service perimeter.

  • If the Spark workload doesn't use the default service account, you must ensure that your custom service account has the Service Usage Consumer (roles/serviceusage.serviceUsageConsumer) IAM role in the project that hosts the container image.

Spark

Don't include Spark in your custom container image. At runtime, Managed Service for Apache Spark mounts Spark binaries and configs from the host into the container: binaries are mounted to the /usr/lib/spark directory and configs are mounted to the /etc/spark/conf directory. Existing files in these directories are overridden by Managed Service for Apache Spark at runtime.

Java Runtime Environment

Don't include your own Java Runtime Environment (JRE) in your custom container image. At run time, Managed Service for Apache Spark mounts OpenJDK from the host into the container. If you include a JRE in your custom container image, it will be ignored.

Java packages

You can include jar files as Spark workload dependencies in your custom container image, and you can set the SPARK_EXTRA_CLASSPATH env variable to include the jars. Managed Service for Apache Spark will add the env variable value in the classpath of Spark JVM processes. Recommendation: put jars under the /opt/spark/jars directory and set SPARK_EXTRA_CLASSPATH to /opt/spark/jars/*.

You can include the workload jar in your custom container image, then reference it with a local path when submitting the workload, for example file:///opt/spark/jars/my-spark-job.jar (see Submit a Spark batch workload using a custom container image for an example).

Python packages

By default, Managed Service for Apache Spark mounts a Conda environment build using an OSS Conda-Forge repo from the host to the /opt/managed-spark/conda directory in the container at runtime. PYSPARK_PYTHON is set to /opt/managed-spark/conda/bin/python. Its base directory, /opt/managed-spark/conda/bin, is included in PATH.

You can include your Python environment with packages in a different directory in your custom container image, for example in /opt/conda, and set the PYSPARK_PYTHON environment variable to /opt/conda/bin/python.

Your custom container image can include other Python modules that are not part of the Python environment, for example, Python scripts with utility functions. Set the PYTHONPATH environment variable to include the directories where the modules are located.

R environment

You can customize the R environment in your custom container image using one of the following options:

  • Use Conda to manage and install R packages from conda-forge channel.
  • Add an R repository for your container image Linux OS, and install R packages using the Linux OS package manager (see the R Software package index).

When you use either option, you must set the R_HOME environment variable to point to your custom R environment. Exception: If you are using Conda to both manage your R environment and customize your Python environment, you don't need to set the R_HOME environment variable; it is automatically set based on the PYSPARK_PYTHON environment variable.

Example custom container image build

This section includes custom container image build examples, which include sample Dockerfiles, followed by build commands. One sample includes the minimum configuration required to build an image. The other sample includes examples of extra configuration, including Python and R libraries.

Architecture co-location recommendation

To ensure rapid container initialization using Google Cloud Image streaming (Container File System API) and to prevent cross-regional data movement fees, the following services should all be located in the same region:

  • Artifact Registry
  • Cloud Storage bucket
  • Managed Service for Apache Spark batch workload

Step 1: Environment configuration

Open your terminal or Google Cloud Shell and set up the following environment variables to simplify the deployment process.

export PROJECT_ID="YOUR_PROJECT_ID"
export REGION="us-east1"
export REPO_NAME="spark-images"
export BUCKET_NAME="spark-jobs-bucket-${PROJECT_ID}-${REGION}"

Step 2: Dockerfile configuration

Minimum configuration

# Recommendation: Use Debian 12.
FROM debian:12-slim

# Suppress interactive prompts.
ENV DEBIAN_FRONTEND=noninteractive

# Install utilities required by Spark scripts.
RUN apt update && apt install -y procps tini libjemalloc2

# Enable jemalloc as default memory allocator.
ENV LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so.2

# Create the 'spark' group/user.
# The GID and UID must be 1099. Home directory is required.
RUN groupadd -g 1099 spark
RUN useradd -u 1099 -g 1099 -d /home/spark -m spark

USER spark

Extra configuration

# Recommendation: Use Debian 12.
FROM debian:12-slim

# Suppress interactive prompts.
ENV DEBIAN_FRONTEND=noninteractive

# Install utilities required by Spark scripts.
RUN apt update && apt install -y procps tini libjemalloc2

# Enable jemalloc as default memory allocator
ENV LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so.2

# Install utilities required by XGBoost for Spark.
RUN apt install -y procps libgomp1

# Install and configure Miniconda3.
ENV CONDA_HOME=/opt/miniforge3
ENV PYSPARK_PYTHON=${CONDA_HOME}/bin/python
ENV PATH=${CONDA_HOME}/bin:${PATH}
ADD https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-x86_64.sh .
RUN bash Miniforge3-Linux-x86_64.sh -b -p /opt/miniforge3 \
  && ${CONDA_HOME}/bin/conda config --system --set always_yes True \
  && ${CONDA_HOME}/bin/conda config --system --set auto_update_conda False \
  && ${CONDA_HOME}/bin/conda config --system --set channel_priority strict
# Packages ipython and ipykernel are required if using custom conda and want to
# use this container for running notebooks.
RUN ${CONDA_HOME}/bin/mamba install ipython ipykernel

# Install Google Cloud SDK.
RUN ${CONDA_HOME}/bin/mamba install -n base google-cloud-sdk

# Install Conda packages.
#
# The following packages are installed in the default image.
# Recommendation: include all packages.
#
# Use mamba to quickly install packages.
RUN ${CONDA_HOME}/bin/mamba install -n base \
    accelerate \
    bigframes \
    cython \
    deepspeed \
    evaluate \
    fastavro \
    fastparquet \
    gcsfs \
    google-cloud-aiplatform \
    google-cloud-bigquery-storage \
    google-cloud-bigquery[pandas] \
    google-cloud-bigtable \
    google-cloud-container \
    google-cloud-datacatalog \
    google-cloud-dataproc \
    google-cloud-datastore \
    google-cloud-language \
    google-cloud-logging \
    google-cloud-monitoring \
    google-cloud-pubsub \
    google-cloud-redis \
    google-cloud-spanner \
    google-cloud-speech \
    google-cloud-storage \
    google-cloud-texttospeech \
    google-cloud-translate \
    google-cloud-vision \
    langchain \
    lightgbm \
    koalas \
    matplotlib \
    mlflow \
    nltk \
    numba \
    numpy \
    openblas \
    orc \
    pandas \
    pyarrow \
    pynvml \
    pysal \
    pytables \
    python \
    pytorch-cpu \
    regex \
    requests \
    rtree \
    scikit-image \
    scikit-learn \
    scipy \
    seaborn \
    sentence-transformers \
    sqlalchemy \
    sympy \
    tokenizers \
    transformers \
    virtualenv \
    xgboost

# Install pip packages.
RUN ${PYSPARK_PYTHON} -m pip install \
    spark-tensorflow-distributor \
    torcheval

# Install R and R libraries.
RUN ${CONDA_HOME}/bin/mamba install -n base \
    r-askpass \
    r-assertthat \
    r-backports \
    r-bit \
    r-bit64 \
    r-blob \
    r-boot \
    r-brew \
    r-broom \
    r-callr \
    r-caret \
    r-cellranger \
    r-chron \
    r-class \
    r-cli \
    r-clipr \
    r-cluster \
    r-codetools \
    r-colorspace \
    r-commonmark \
    r-cpp11 \
    r-crayon \
    r-curl \
    r-data.table \
    r-dbi \
    r-dbplyr \
    r-desc \
    r-devtools \
    r-digest \
    r-dplyr \
    r-ellipsis \
    r-evaluate \
    r-fansi \
    r-fastmap \
    r-forcats \
    r-foreach \
    r-foreign \
    r-fs \
    r-future \
    r-generics \
    r-ggplot2 \
    r-gh \
    r-glmnet \
    r-globals \
    r-glue \
    r-gower \
    r-gtable \
    r-haven \
    r-highr \
    r-hms \
    r-htmltools \
    r-htmlwidgets \
    r-httpuv \
    r-httr \
    r-hwriter \
    r-ini \
    r-ipred \
    r-isoband \
    r-iterators \
    r-jsonlite \
    r-kernsmooth \
    r-knitr \
    r-labeling \
    r-later \
    r-lattice \
    r-lava \
    r-lifecycle \
    r-listenv \
    r-lubridate \
    r-magrittr \
    r-markdown \
    r-mass \
    r-matrix \
    r-memoise \
    r-mgcv \
    r-mime \
    r-modelmetrics \
    r-modelr \
    r-munsell \
    r-nlme \
    r-nnet \
    r-numderiv \
    r-openssl \
    r-pillar \
    r-pkgbuild \
    r-pkgconfig \
    r-pkgload \
    r-plogr \
    r-plyr \
    r-praise \
    r-prettyunits \
    r-processx \
    r-prodlim \
    r-progress \
    r-promises \
    r-proto \
    r-ps \
    r-purrr \
    r-r6 \
    r-randomforest \
    r-rappdirs \
    r-rcmdcheck \
    r-rcolorbrewer \
    r-rcpp \
    r-rcurl \
    r-readr \
    r-readxl \
    r-recipes \
    r-recommended \
    r-rematch \
    r-remotes \
    r-reprex \
    r-reshape2 \
    r-rlang \
    r-rmarkdown \
    r-rodbc \
    r-roxygen2 \
    r-rpart \
    r-rprojroot \
    r-rserve \
    r-rsqlite \
    r-rstudioapi \
    r-rvest \
    r-scales \
    r-selectr \
    r-sessioninfo \
    r-shape \
    r-shiny \
    r-sourcetools \
    r-spatial \
    r-squarem \
    r-stringi \
    r-stringr \
    r-survival \
    r-sys \
    r-teachingdemos \
    r-testthat \
    r-tibble \
    r-tidyr \
    r-tidyselect \
    r-tidyverse \
    r-timedate \
    r-tinytex \
    r-usethis \
    r-utf8 \
    r-uuid \
    r-vctrs \
    r-whisker \
    r-withr \
    r-xfun \
    r-xml2 \
    r-xopen \
    r-xtable \
    r-yaml \
    r-zip

ENV R_HOME=/usr/lib/R

# Add extra Python modules.
ENV PYTHONPATH=/opt/python/packages
RUN mkdir -p "${PYTHONPATH}"

# Add extra jars.
ENV SPARK_EXTRA_JARS_DIR=/opt/spark/jars/
ENV SPARK_EXTRA_CLASSPATH='/opt/spark/jars/*'
RUN mkdir -p "${SPARK_EXTRA_JARS_DIR}"

# Uncomment below and replace EXTRA_JAR_NAME with the jar file name.
# COPY "EXTRA_JAR_NAME" "${SPARK_EXTRA_JARS_DIR}"

# Create the 'spark' group/user.
# The GID and UID must be 1099. Home directory is required.
RUN groupadd -g 1099 spark
RUN useradd -u 1099 -g 1099 -d /home/spark -m spark

USER spark

Step 3: Build & push commands

Execute these commands to build and upload your container image into Artifact Registry using Cloud Build.

  1. Create the Cloud Storage bucket:

    gcloud storage buckets create gs://${BUCKET_NAME} --project=${PROJECT_ID} \
  --location=${REGION}

  2. Create the Artifact Registry repository:

    gcloud artifacts repositories create ${REPO_NAME} \
  --repository-format=docker --location=${REGION} --description="Custom Dataproc Serverless images"

  3. Build and push the image using Cloud Build:

    gcloud builds submit --tag ${REGION}-docker.pkg.dev/${PROJECT_ID}/${REPO_NAME}/spark-custom:latest .

Step 4: Job submission

Deploy the job by defining the runtime --version=2.3 to avoid version drift issues. In production environments, providing the image version instead of the :latest tag is a best practice.

  • Permissions: Ensure the appropriate service account has the ** Artifact Registry Reader** (roles/artifactregistry.reader) role on the Artifact Registry repository (recommended) or the project that hosts the container image. This is especially important if the image is hosted in a different project than where the workload runs. 
gcloud dataproc batches submit pyspark \
  file:///usr/lib/spark/examples/src/main/python/pi.py \
  --region=${REGION} \
  --project=${PROJECT_ID} \
  --version=2.3 \
  --container-image=${REGION}-docker.pkg.dev/${PROJECT_ID}/${REPO_NAME}/spark-custom:latest \
  -- 10