Environment Variables

You can define environment variables that handle required configuration. In HPE Machine Learning Data Management, you can define the following types of environment variables:

  • pachd variables: Used for your HPE Machine Learning Data Management daemon container.
  • HPE Machine Learning Data Management worker variables: Used by the Kubernetes pods that run your pipeline code.
Tip
You can reference environment variables in your code. For example, if your code writes data to an external system and you want to know the current job ID, you can use the PACH_JOB_ID environment variable to refer to the current job ID.

pachd Environment Variables

You can find the list of pachd environment variables in the pachd manifest by running the following command:

kubectl get deploy pachd -o yaml

The following tables list all the pachd environment variables.

Global Configuration

Environment Variable Default Value Description
ETCD_SERVICE_HOST N/A The host on which the etcd service runs.
ETCD_SERVICE_PORT N/A The etcd port number.
PPS_WORKER_GRPC_PORT 80 The GRPs port number.
PORT 650 The pachd port number.
HTTP_PORT 652 The HTTP port number.
PEER_PORT 653 The port for pachd-to-pachd communication.
NAMESPACE default The namespace in which HPE Machine Learning Data Management is deployed.

PachD Configuration

Environment Variable Default Value Description
NUM_SHARDS 32 The max number of pachd pods that can run in a
single cluster.
STORAGE_BACKEND "" The storage backend defined for the HPE Machine Learning Data Management cluster.
STORAGE_HOST_PATH "" The host path to storage.
KUBERNETES_PORT_443_TCP_ADDR none An IP address that Kubernetes exports
automatically for your code to communicate with
the Kubernetes API. Read access only. Most variables
that have use the PORT_ADDRESS_TCP_ADDR pattern
are Kubernetes environment variables. For more information,
see Kubernetes environment variables.
METRICS true Defines whether anonymous HPE Machine Learning Data Management metrics are being
collected or not.
BLOCK_CACHE_BYTES 1G The size of the block cache in pachd.
WORKER_IMAGE "" The base Docker image that is used to run your pipeline.
WORKER_SIDECAR_IMAGE "" The pachd image that is used as a worker sidecar.
WORKER_IMAGE_PULL_POLICY IfNotPresent The pull policy that defines how Docker images are
pulled. You can set
a Kubernetes image pull policy as needed.
LOG_LEVEL info Verbosity of the log output. If you want to disable
logging, set this variable to 0. Viable Options
debug
info
error
For more information, see Go logrus log levels.
IAM_ROLE "" The role that defines permissions for HPE Machine Learning Data Management in AWS.
IMAGE_PULL_SECRET "" The Kubernetes secret for image pull credentials.
EXPOSE_OBJECT_API false Controls access to internal HPE Machine Learning Data Management API.
WORKER_USES_ROOT true Controls root access in the worker container.
S3GATEWAY_PORT 600 The S3 gateway port number
DISABLE_COMMIT_PROGRESS_COUNTER false A feature flag that disables commit propagation
progress counter. If you have a large DAG,
setting this parameter to true might help
improve etcd performance. You only need to set
this parameter on the pachd pod. HPE Machine Learning Data Management passes
this parameter to worker containers automatically.

Storage Configuration

Environment Variable Default Value Description
STORAGE_MEMORY_THRESHOLD N/A Defines the storage memory threshold.
STORAGE_SHARD_THRESHOLD N/A Defines the storage shard threshold.

Pipeline Worker Environment Variables

HPE Machine Learning Data Management defines many environment variables for each HPE Machine Learning Data Management worker that runs your pipeline code. You can print the list of environment variables into your HPE Machine Learning Data Management logs by including the env command into your pipeline specification. For example, if you have an images repository, you can configure your pipeline specification like this:

{
    "pipeline": {
        "name": "env"
    },
    "input": {
        "pfs": {
            "glob": "/",
            "repo": "images"
        }
    },
    "transform": {
        "cmd": ["sh" ],
        "stdin": ["env"],
        "image": "ubuntu:14.04"
    }
}

Run this pipeline and, upon completion, you can view the log with variables by running the following command:

pachctl logs --pipeline=env
PPS_WORKER_IP=172.17.0.7
DASH_PORT_8081_TCP_PROTO=tcp
PACHD_PORT_600_TCP_PORT=600
KUBERNETES_SERVICE_PORT=443
KUBERNETES_PORT=tcp://10.96.0.1:443
...

You should see a lengthy list of variables. Many of them define internal networking parameters that most probably you will not need to use.

Most users find the following environment variables particularly useful:

Environment Variable Description
AWS_ACCESS_KEY_ID The ID that contains your AWS access key; requires pfs.s3: true or s3Out:true in your pipeline spec.
AWS_SECRET_ACCESS_KEY The name of the secret which contains your AWS access key; requires pfs.s3: true or s3Out:true in your pipeline spec.
PACH_JOB_ID The ID of the current job. For example,
PACH_JOB_ID=8991d6e811554b2a8eccaff10ebfb341.
PACH_DATUM_ID The ID of the current Datum.
FILESET_ID The ID of the file set which contains the input files for a given job.
PACHD_PEER_SERVICE_HOST The host on which a pachd peer service runs. Used by the Pachyderm SDK.
PACHD_PEER_SERVICE_PORT The port number of a pachd peer service. Used by the Pachyderm SDK.
PACH_DATUM_<input.name>_JOIN_ON Exposes the join_on match to the pipeline’s job.
PACH_DATUM_<input.name>_GROUP_BY Expose the group_by match to the pipeline’s job.
PACH_OUTPUT_COMMIT_ID The ID of the commit in the output repo for
the current job. For example,
PACH_OUTPUT_COMMIT_ID=a974991ad44d4d37ba5cf33b9ff77394.
PPS_NAMESPACE The PPS namespace. For example,
PPS_NAMESPACE=default.
PPS_SPEC_COMMIT The hash of the pipeline specification commit.
This value is tied to the pipeline version. Therefore, jobs that use
the same version of the same pipeline have the same spec commit.
For example, PPS_SPEC_COMMIT=3596627865b24c4caea9565fcde29e7d.
PPS_POD_NAME The name of the pipeline pod. For example,
pipeline-env-v1-zbwm2.
PPS_PIPELINE_NAME The name of the pipeline that this pod runs.
For example, env.
PIPELINE_SERVICE_PORT_PROMETHEUS_METRICS The port that you can use to
exposed metrics to Prometheus from within your pipeline. The default value is 9090.
HOME The path to the home directory. The default value is /root
<input-repo>=<path/to/input/repo> The path to the filesystem that is
defined in the input in your pipeline specification. HPE Machine Learning Data Management defines
such a variable for each input. The path is defined by the glob pattern in the
spec. For example, if you have an input images and a glob pattern of /,
HPE Machine Learning Data Management defines the images=/pfs/images variable. If you
have a glob pattern of /*, HPE Machine Learning Data Management matches
the files in the images repository and, therefore, the path is
images=/pfs/images/liberty.png.
input_COMMIT The ID of the commit that is used for the input.
For example, images_COMMIT=fa765b5454e3475f902eadebf83eac34.
S3_ENDPOINT A HPE Machine Learning Data Management S3 gateway sidecar container endpoint.
If you have an S3 enabled pipeline, this parameter specifies a URL that
you can use to access the pipeline’s repositories state when a
particular job was run. The URL has the following format:
http://<job-ID>-s3:600.
An example of accessing the data by using AWS CLI looks like this: `echo foo_data

In addition to these environment variables, Kubernetes injects others for Services that run inside the cluster. These variables enable you to connect to those outside services, which can be powerful but might also result in processing being retried multiple times.

For example, if your code writes a row to a database, that row might be written multiple times because of retries. Interaction with outside services must be idempotent to prevent unexpected behavior. Furthermore, one of the running services that your code can connect to is HPE Machine Learning Data Management itself. This is generally not recommended as very little of the HPE Machine Learning Data Management API is idempotent, but in some specific cases it can be a viable approach.