Codalab

Bugfixes:

* `cl mount` command is now fixed and operational on systems that support FUSE filesystems
* URL based bundle uploads now save the URL to the `source_url` metadata field of the bundle
* When a Codalab worker is initialized with incorrect credentials, it immediately quits with a helpfu error message rather than infinitely looping while throwin indecipherable error messages

Background improvements

* Workers now use the official Docker Python SDK to operate docker rather than an in-house HTTP API wrapper. Expect better docker image handling and fewer issues from workers
* **(backwards incompatible):**Worker now use version 2 of nvidia-docker rather than version 1 as version 1 is deprecated. If you're planning to run your own GPU workers please make sure you have nvidia-docker v2 and not v1. You can upgrade per the instructions [here](https://github.com/nvidia/nvidia-docker/wiki/Installation-(version-2.0))

Minor bugfix release.

- Make files open in browser instead of download.
- Now can clone repo to Windows machines.
- New CodaLab default docker images
- Worker reliability improvements.

Worksheets/Server Changes

* Disk and time quotas for run bundles that do not specify them are now set to the total left for the user. Note that if a user starts multiple runs without specifying these quotas and together the runs go over the quota, one of them will fail
* Monitor script is now more lenient with runtime of runs in case they stay staged for a long time.

CLI Changes

* The CLI may now be configured to pass extra HTTP headers with its REST requests. These may be set up in the CodaLab config file.

Worker Bugfixes

* Docker image management fixes: Docker image size limiting process now doesn't touch images actively being used (or expected) by a run. This prevents infinite looping docker image downloads even in pathological cases
* Dependency manager is now more resilient to external filesystem changes and just overwrites if a dependency directory it expects to be empty happens to be there.
* Dependency manager now prunes failed dependencies and retries downloading them if they are requested >10 seconds after the initial failure. This prevents dependency download failures.

Worker bugfixes
* When the docker container for a run cannot start for a reason (like too little memory requested), the run fails instead of retrying forever
* Rewritten concurrency model in Dependency Manager fixes 929, dependency download speeds should be back to normal. All versions >v0.2.26 and <v0.2.34 are affected from this bug, so if you have any of these versions, please update your worker to 0.2.34

Features
Worker
* Workers now have a `--exit-when-idle` flag which kills the worker process if it is ever idle after checking in with the server. This is part of an ongoing effort to allow ephemeral/one-use workers in shared computation environments like Slurm-managed clusters.

Bug fixes
Worksheets web app
* Fixed a bug where consecutive Markdown blocks were not merged, leading to rendering issues

Worker
* Fixed worker dependency manager initialization so the worker can correctly pick up existing cached dependencies upon restart
* Fixed a bug in worker dependency manager state saving that prevented any dependency from getting cleaned up, eventually making the dependency manager hit the disk limit and hang all dependency downloads.
* Fixed a bug where if the Docker container for a run cannot be started, the worker would forever keep attempting instead of failing the bundle. The bundles are failed immediately now if their Docker containers cannot be started.

Bundle server
* Fixed a bug that failed worker check-ins if the check-in included any run that had since been deleted server-side. Now the deleted run is ignored.

Misc changes
* codalab-cli python code is now auto formatted with [Black](https://github.com/ambv/black) with the exception of string normalization and a more liberal 100 character line limit to reduce the amount of formatting changes
* The repo comes with a `pyproject.toml` file to configure Black with these settings and a `.pre-commit-config.yaml` file to run Black before each commit if you have pre-commit installed.
* This also means CodaLab code is now written in a Python3 compatible syntax, but no further work has been done for Python3 compatibility beyond fixing the syntax differences.
* Updated package dependency versions to newer versions as long as there were no breaking changes, mainly as the previous version of gunicorn used had security vulnerabilities.