nvidia-docker 2 is now used as the runtime engine.
This allows using images built on more recent CUDA setups, such as the official Tensorflow 1.13 images.
S3 inputs are much faster to resolve.
Enterprise users with custom environments can now set the grace period after which servers can be scaled down.
It defaults to 15 minutes.
When available, the original creation and modification times of Datums are now acquired from workers.
When enough memory is available, /tmp is now a tmpfs (RAM disk) on agent machines by default.
You can explicitly enable or disable this per-execution by setting the VH_TMPFS environment variable
to 0 or 1.
Carriage Return (\r) characters are now better supported by the execution agent, so progress bars such
as those emitted by the popular tqdm module should work much better.
Projects' queued/started execution counts are now shown in the dashboard.
Docker image formats are now validated.
Multiple input defaults are now supported in the UI.
The job running agent won't attempt to download all inputs in parallel anymore.
Input files are cached for a longer time.
Jobs should finish faster, since less redundant information is saved.
The agent software will now consider 301 Redirects without a Location header erroneous.
(This mainly concerns S3 URLs with an invalid addressing scheme.)
A new version of the CLI tool was also released with a lot of new features and quality-of-life improvements.
You can now use Valohai-CLI without a linked project aka. API driver mode.