Kubeflow

  • Kubeflow Charmers | bundle
  • Cloud
Channel Revision Published
latest/candidate 294 24 Jan 2022
latest/beta 430 30 Aug 2024
latest/edge 423 26 Jul 2024
1.9/stable 432 03 Dec 2024
1.9/beta 420 19 Jul 2024
1.9/edge 431 03 Dec 2024
1.8/stable 414 22 Nov 2023
1.8/beta 411 22 Nov 2023
1.8/edge 413 22 Nov 2023
1.7/stable 409 27 Oct 2023
1.7/beta 408 27 Oct 2023
1.7/edge 407 27 Oct 2023
juju deploy kubeflow --channel 1.9/stable
Show information

Platform:

This guide describes how to back up and restore the Charmed Kubeflow (CKF) control plane data to a compatible S3 storage.

It is expected that these steps are followed all at once, backing up all databases, pipelines MinIO bucket, and ML Metadata database at the same time. Failing to do so may result in data loss.

Running Kubeflow pipelines and Katib experiments can affect the outcome of the backup, please make sure all pipelines and experiments are stopped and no other processes are calling them, such as Jupyter Notebooks.

User workloads in user namespaces are not backed up.

Requirements

  • Access to a S3 storage used for the backup data - only AWS S3 and S3 RadosGW are supported.
  • Admin access to the Kubernetes cluster where CKF is deployed.
  • Juju admin access to the kubeflow model.
  • yq binary.
  • Ensure the local storage is big enough to back up the data.

Configure rclone

rclone is a tool that allows file management in cloud storage. This tool will be used for backing up several files throughout this guide and it can be installed as a snap:

sudo snap install rclone

Connect to a shared S3 storage

  1. Configure rclone to connect to the shared S3 storage. The following can be used as reference:
[remote-s3]
type = s3
provider = AWS
env_auth = true
access_key_id = ...
secret_access_key = ...
region = eu-central-1
acl = private
server_side_encryption = AES256

You can check where this configuration file is located with rclone config file.

  1. Save the name of the S3 remote in an ENV variable:
RCLONE_S3_REMOTE=remote-s3

Connect to CKF MinIO

  1. The following steps require an accessible MinIO endpoint, which can be done port forwarding the minio service:
kubectl port-forward -n kubeflow svc/minio 9000:9000
  1. Get minio’s secret-key value:
juju show-unit kfp-ui/0 \
        | yq '.kfp-ui/0.relation-info.[] | select (.endpoint == "object-storage") | .application-data.data' \
        | yq '.secret-key'
  1. Get minio’s access-key:
juju config minio access-key
  1. Configure rclone to connect to CKF MinIO. The following can be used as reference:
[minio-ckf]
type = s3
provider = Minio
access_key_id = minio
secret_access_key = ...
endpoint = http://localhost:9000
acl = private
  1. Save the name of the MinIO remote in an ENV variable:
RCLONE_MINIO_CKF_REMOTE=minio-ckf

Back up CKF databases to S3 storage

CKF uses katib-db and kfp-db as databases for Katib and Kubeflow pipelines respectively.

  1. Deploy and configure the s3-integrator to connect to the shared S3 storage.

See S3 AWS and S3 Radowsg configuration guides for this step.

  1. Scale up kfp-db and katib-db.

This step avoids the Primary database from becoming unavailable during backup.

juju scale-application kfp-db 2
juju scale-application katib-db 2
  1. Create a backup for each database.

Replace mysql-k8s with the name of the database you intend to create a backup for in the commands from that guide.

Back up ML Metadata using sqlite3

The mlmd charm uses a SQLite database to store ML metadata generated from Kubeflow pipelines.

  1. Install the required tools inside the application container:

This step expects the mlmd application container to have Internet access.

# MLMD > 1.14, CKF 1.9
MLMD_POD="mlmd-0"
MLMD_CONTAINER="mlmd-grpc-server"

# MLMD 1.14, CKF 1.8
MLMD_POD="mlmd-0"
MLMD_CONTAINER="mlmd"

kubectl exec -n kubeflow $MLMD_POD -c $MLMD_CONTAINER -- \
    /bin/bash -c "apt update && apt install sqlite3 -y"
  1. Scale down kfp-metadata-writer. This is done to prevent any additional writes to mlmd.
juju scale-application kfp-metadata-writer 0
  1. Perform a database backup.

This moves all the database contents into a compressed text file inside the mlmd-0 container:

MLMD_BACKUP=mlmd-$(date -d "today" +"%Y-%m-%d-%H-%M").dump.gz

kubectl exec -n kubeflow $MLMD_POD -c $MLMD_CONTAINER -- \
	/bin/bash -c \
	"sqlite3 /data/mlmd.db .dump | gzip -c >/tmp/$MLMD_BACKUP"
  1. Copy the backup file to local storage:
kubectl cp -n kubeflow -c $MLMD_CONTAINER \
	$MLMD_POD:/tmp/$MLMD_BACKUP \
	./$MLMD_BACKUP
  1. Copy the mlmd backup data to the S3 storage:
S3_BUCKET=backup-bucket-2024
RCLONE_S3_REMOTE=remote-s3
RCLONE_BWIDTH_LIMIT=20M

rclone --size-only copy \
	--bwlimit $RCLONE_BWIDTH_LIMIT \
	./$MLMD_BACKUP \
	$RCLONE_S3_REMOTE:$S3_BUCKET

Optionally, you can remove the mlmd data from your local machine:

rm -rf $MLMD_BACKUP
  1. Scale up kfp-metadata-writer:
juju scale-application kfp-metadata-writer 1

Back up mlpipeline MinIO bucket

Sync all files from minio to the shared S3 storage:

S3_BUCKET=backup-bucket-2024
RCLONE_S3_REMOTE=remote-s3
RCLONE_BWIDTH_LIMIT=20M

rclone --size-only sync \
	--bwlimit $RCLONE_BWIDTH_LIMIT \
	$RCLONE_MINIO_REMOTE:mlpipeline \
	$RCLONE_S3_REMOTE:$S3_BUCKET/mlpipeline

Back up ML Metadata with kubectl

You can also perform the backup using kubectl.

  1. Scale down kfp-metadata-writer. This is done to prevent any additional writes to mlmd:
juju scale-application kfp-metadata-writer 0
  1. Copy the backup file to local storage:
# MLMD > 1.14, CKF 1.9
MLMD_POD="mlmd-0"
MLMD_CONTAINER="mlmd-grpc-server"

# MLMD 1.14, CKF 1.8
MLMD_POD="mlmd-0"
MLMD_CONTAINER="mlmd"

kubectl cp -n kubeflow -c $MLMD_CONTAINER \
	$MLMD_POD:/data/mlmd.db \
	./$MLMD_BACKUP
  1. Copy the mlmd backup data to the S3 storage:
S3_BUCKET=backup-bucket-2024
RCLONE_S3_REMOTE=remote-s3
RCLONE_BWIDTH_LIMIT=20M

rclone --size-only copy \
	--bwlimit $RCLONE_BWIDTH_LIMIT \
	./$MLMD_BACKUP \
	$RCLONE_S3_REMOTE:$S3_BUCKET

Optionally, you can remove the `mlmd’ backup data from your local machine:

rm -rf $MLMD_BACKUP
  1. Scale up kfp-metadata-writer:
juju scale-application kfp-metadata-writer 1