DC/OS Data Science Engine 2.1.0 was released on 16, November 2020
New Features
- Bumped SDK version to
0.58.0
.
- DC/OS Data Science Engine now allows log rotate options for
stdout
and stderr
.
- DC/OS Data Science Engine now allows external volume as a storage. Supported driver names would be
pxd
for Portworx, netapp
for NetApp, and any other generic driver name.
DC/OS Data Science Engine 2.0.1 was released on 31, July 2020
New Features
- DC/OS Data Science Engine now allows multiple Spark UI instances to be accessible
DC/OS Data Science Engine 2.0.0 was released on 12, June 2020
New Features
- Horovod for supporting distributed training with TensorFlow, PyTorch, and MXNet
- Horovod-on-Spark integration for dynamic allocation of cluster resources
- User management and filesystem permissions by adding the ability to access host path volumes with custom UID/GID
- Seamless upgrade process
- EdgeLB integration (limited to notebooks only)
- Pre-baked Notebook images with different flavors of ML libraries
- TensorFlow, PyTorch, MXNet as the main ML distributions
- Added Horovod + Spark for distributed use-cases
- Notebook image has two flavors: local and distributed
- Each image has CPU and GPU variant
- Mounting host path support
- Upgrading Jupyter and other frameworks to the latest versions:
- JupyterLab 2.0
- TensorFlow 1.15
- TensorFlow 2.1.0
- PyTorch 1.4.0
- MXNet 1.6.0
- Horovod 0.19.0
Bug Fixes
- Fixed allowing spaces in the notebook name
- Fixed Conda packages directory path for conda virtual environment creation
Limitations
- BeakerX kernels have been removed from the distribution. Now only IPython, Toree, and R kernels supported
Release Notes for DC/OS Data Science Engine version 1.0.2
New Features
- Updated miniconda version from
4.6.14
to 4.7.12.1
Bug Fixes
- Fixed LD_LIBRARY_PATH to point to correct version of CUDA libraries
- Fixed Spark Environment to have SPARK_DIST_CLASSPATH set
- Fixed Service Account file path to allow notebooks to run from any directory
- Fixed BeakerX configuration file path to allow BeakerX Spark Magic to work from any directory
Limitations
- DC/OS Data Science Engine does not support notebook having spaces in the name. It is recommended to use underscore (
_
) instead of (
) in the notebook name.
Release Notes for DC/OS Data Science Engine version 1.0.1
New Features
- Added automation for enabling/disabling verbose logging for Spark resource offers
- Notebook Docker image is configurable via service options.
- DC/OS Data Science Engine now stores fetched Hadoop configuration files, hdfs-site.xml and core-site.xml, in
$MESOS_SANDBOX/hadoop_conf
folder so users can modify them without root privileges.
Bug Fixes
- Fixed Spark configuration dialog in BeakerX kernel.
- Fixed folder permissions so users can install new packages with
conda
or pip
.
- Fixed static resources paths in host mode for components with UI: Spark, History Server, Tensorboard
Breaking Changes
- CUDA 9 images will no longer be provided with a bundle; CUDA 10 is the only version used for GPU images.
jaas_secret
option removed from security configuration. extra_spark_secrets
should be used instead.
Limitations
- DC/OS Data Science Engine does not fully support
root
service user. It is recommended to use the default user nobody
.
- DC/OS Data Science Engine does not support the installation of Python packages on CoreOS. It is recommended to use CentOS.
- DC/OS Data Science Engine does not support AWS Classic Load Balancer, because it does not support WebSockets. It is recommended to configure load balancers and proxies to allow WebSockets.