ServiceX_transformer Repo for input-output transformation code for ServiceX

Allows a user to extract columns from a Root file. There are two supported transformers available in this repo:

1. xaod_transformer - Uses pyroot to extract data from xAOD files
2. uproot_transformer - Uses uproot to pull data out of a tree in a flat Root file.

Both of these are based a Docker image found here: https://hub.docker.com/r/atlas/analysisbase

The transformer accepts a list of columns (Root Branches) to extract from the file along with Root files. It extracts the data into awkward arrays and publishes arrow buffers to a backend messaging system.

How to Build

The base docker image can be built as:

docker build -t sslhep/servicex-transformer-base:develop .

Once you have this image, you can build the two variations as

docker build --build-arg BASE_VERSION=develop -f Dockerfile.uprootTransformer -t sslhep/servicex-transformer-uproot:develop .
docker build --build-arg BASE_VERSION=develop -f Dockerfile.xAODTransformer -t sslhep/servicex-transformer-xaod:develop .

How to Run

The container is designed to run inside a kubernetes cluster as part of the serviceX application. It can also be run standalone as a CLI utility for transforming ROOT files.

You can launch a container with an X509 proxy mounted in a docker volume as:

docker run --rm \
    --mount type=bind,source=$HOME/.globus,readonly,target=/etc/grid-certs \
    --mount type=bind,source="$(pwd)"/secrets/secrets.txt,target=/servicex/secrets.txt \
    --mount type=volume,source=x509,target=/etc/grid-security \
    --name=x509-secrets sslhep/x509-secrets:latest


docker run --rm -it \
    --mount type=volume,source=x509,target=/etc/grid-security-ro \
    sslhep/servicex-transformer:develop bash  

Transformer Scripts

Both transformer scripts are included in the container:

• uproot_transformer.py
• xaod_transformer.py

Both of them have a common set of command line arguments. The scripts are usually launched by the ServiceX Transformer Manager as job where each instance of the transformer pulls requests for ROOT files from a RabbitMQ.

The scripts can also be called from the command line with a path to a Root file.

Transformed Result Output

Command line arguments determine a destination for the results as well as an output format.

• Kafka - Streaming system. Write messages formatted as Arrow tables. The chunks parameter determines how many events are included in each message.
• Object Store - Each transformed file is written as an object to an S3 compatible object store. The only currently supported output file format is parquet. The objects are stored in a bucket named after the transformation request ID.

Command Line Reference

Development

For ease, we often will launch a bash shell with this repo's source code mounted as a volume. You can edit the source code using your desktop's tools and execute inside the container.

To launch this container, cd to root of this repo.

docker run -it \
    --mount type=bind,source=$(pwd),target=/code \
    --mount type=bind,source=$(pwd)/../data,target=/data \
    --mount type=volume,source=x509,target=/etc/grid-security-ro \
    sslhep/servicex-transformer:develop bash

This assumes that you have a directory above this repo called data that has some sample ROOT files.

To run the transformer from the command line you can cd /code to move to the copy of the code served up from your host.

python xaod_branches.py --path /data/AOD.11182705._000001.pool.root.1 --result-destination kafka --request-id my-bucket --chunks 5000

If you want to run it from the command line and use the object storage:

export MINIO_ACCESS_KEY='miniouser'
export MINIO_SECRET_KEY='leftfoot1'
export MINIO_URL='host.docker.internal:9000'
python xaod_branches.py --path /data/AOD.11182705._000001.pool.root.1 --result-destination object-store --result-format parquet --request-id my-bucket --chunks 5000

To Debug RabbitMQ interactions

It's possible to run the transformer from the command line and have it interact with RabbitMQ deployed into the local kubernetes cluster.

1. Turn off the App's transformer manager

• Run a local copy of the ServiceX_App on port 5000
• In app.conf set TRANSFORMER_MANAGER_ENABLED = False
• Submit a transform request
• Inside the container run transformer against the rabbit mq topic

python xaod_branches.py --rabbit-uri amqp://user:leftfoot1@host.docker.internal:30672/%2F?heartbeat=9000 --result-destination kafka --request-id 52b40f91-103c-4c64-b732-a7ac96992682 --chunks 5000

Debuggging Kafka

We have a debugging pod deployed to the cluster that has some handy tools. You can get a shell into this pod with the command:

% kubectl -n kafka exec -it testclient bash

From here you can:

1. Delete the topic to clear out old data

% bin/kafka-topics.sh --bootstrap-server servicex-kafka.kafka.svc.cluster.local:9092  --delete --topic servicex 

2. Follow a topic to print the binary (opaque) data to the console:

% bin/kafka-console-consumer.sh --bootstrap-server servicex-kafka:9092 --topic servicex

3. See the size of each partition in the topic:

%  bin/kafka-log-dirs.sh --describe --bootstrap-server servicex-kafka-0.slateci.net:19092 --topic-list servicex | grep '{' | jq '.brokers[].logDirs[].partitions[].size' | grep '^[^0]'

Running Tests

Validation of the code logic is performed using pytest and pytest-mock. Unit test fixtures are in test directories inside each package. The tests depend on imports that are found inside the atlas docker container, so they are usually run inside a container.

Start up the container and run unit tests with:

 docker run sslhep/servicex-transformer:rabbitmq bash -c "source /home/atlas/.bashrc && pytest -s"

The tests are instrumented with code coverage reporting via codecov. The travis job has a the codecov upload token set as an environment variable which is passed into the docker container so the report can be uploaded upon successful conclusion of the tests.

Coding Standards

To make it easier for multiple people to work on the codebase, we enforce PEP8 standards, verified by flake8. The community has found that the 80 character limit is a bit awkward, so we have a local config setting the max_line_length to 99.