⚠️

HPE ML Data Management uses FUSE to mount repositories as local filesystems. Because Apple has announced phasing out support for macOS kernel extensions, including FUSE, this functionality is no longer stable on macOS Catalina (10.15) or later.

HPE ML Data Management enables you to mount a repository as a local filesystem on your computer by using the pachctl mount command. This command uses the Filesystem in Userspace (FUSE) user interface to export a HPE ML Data Management File System (PFS) to a Unix computer system. This functionality is useful when you want to pull data locally to experiment, review the results of a pipeline, or modify the files in the input repository directly.

You can mount a HPE ML Data Management repo in one of the following modes:

• Read-only — you can read the mounted files to further experiment with them locally, but cannot modify them.
• Read-write — you can read mounted files, modify their contents, and push them back into your centralized HPE ML Data Management input repositories.

Prerequisites #

You must have the following configured for this functionality to work:

• Unix or Unix-like operating system, such as Ubuntu 16.04 or macOS Yosemite or later.

• FUSE for your operating system installed:

  • On macOS, run:

    brew install osxfuse

    

  • On Ubuntu, run:

    sudo apt-get install -y fuse

    

    For more information, see:

  • FUSE for macOS

Mounting Repositories in Read-Only Mode #

By default, HPE ML Data Management mounts all repositories in read-only mode. You can access the files through your file browser or enable third-party applications access. Read-only access enables you to explore and experiment with the data, without modifying it. For example, you can mount your repo to a local computer and then open that directory in a Jupyter Notebook for exploration.

ℹ️

The pachctl mount command allows you to mount not only the default branch, typically a master branch, but also other HPE ML Data Management branches. By default, HPE ML Data Management mounts the master branch. However, if you add a branch to the name of the repo, the HEAD of that branch will be mounted.

Example:

pachctl mount images --repos images@staging

You can also mount a specific commit, but because commits might be on multiple branches, modifying them might result in data deletion in the HEAD of the branches. Therefore, you can only mount commits in read-only mode. If you want to write to a specific commit that is not the HEAD of a branch, you can create a new branch with that commit as HEAD.

Mounting Repositories in Read-Write Mode #

Running the pachctl mount command with the --write flag grants you write access to the mounted repositories, which means that you can open the files for editing and put them back to the HPE ML Data Management repository.

⚠️

Your changes are saved to the HPE ML Data Management repository only after you interrupt the pachctl mount with CTRL+C or with pachctl unmount, unmount /<path-to-mount>, or fusermount -u /<path-to-mount>.

For example, you have the OpenCV example pipeline up and running. If you want to edit files in the images repository, experiment with brightness and contrast settings in liberty.png, and finally have your edges pipeline process those changes. If you do not mount the images repo, you would have to first download the files to your computer, edit them, and then put them back to the repository. The pachctl mount command automates all these steps for you. You can mount just the images repo or all HPE ML Data Management repositories as directories on you machine, edit as needed, and, when done, exit the pachctl mount command. Upon exiting the pachctl mount command, HPE ML Data Management uploads all the changes to the corresponding repository.

If someone else modifies the files while you are working on them locally, their changes will likely be overwritten when you exit pachctl mount. This happens because Therefore, make sure that you do not work on the same files while someone else is working on them.

ℹ️

• Use writable mount ONLY when you have sole ownership over the mounted data. Otherwise, merge conflicts or unexpected data overwrites can occur.

• Because output repositories are created by the HPE ML Data Management pipelines, they are immutable. Only a pipeline can change and update files in these repositories. If you try to change a file in an output repo, you will get an error message.

How to Mount/Unmount a HPE ML Data Management Repo #

To mount a HPE ML Data Management repo on a local computer, complete the following steps:

1. In a terminal, go to a directory in which you want to mount a HPE ML Data Management repo. It can be any new empty directory on your local computer. For example, mydirectory.

2. Run pachctl mount for a repository and branch that you want to mount:

   pachctl mount <path-on-your-computer> [flags]

   

   Example:

   • If you want to mount all the repositories in your HPE ML Data Management cluster to a mydirectory directory on your computer and give WRITE access to them, run:

   pachctl mount mydirectory --write

   
   • If you want to mount the master branch of the images repo and enable file editing in this repository, run:

   pachctl mount mydirectory --repos images@master+w

   

   To give read-only access, omit +w.

   System Response:

   ro for images: &{Branch:master Write:true}
   ri: repo:<name:"montage" > created:<seconds:1591812554 nanos:348079652 > size_bytes:1345398 description:"Output repo for pipeline montage." branches:<repo:<name:"montage" > name:"master" >
   continue
   ri: repo:<name:"edges" > created:<seconds:1591812554 nanos:201592492 > size_bytes:136795 description:"Output repo for pipeline edges." branches:<repo:<name:"edges" > name:"master" >
   continue
   ri: repo:<name:"images" > created:<seconds:1591812554 nanos:28450609 > size_bytes:244068 branches:<repo:<name:"images" > name:"master" >
   MkdirAll /var/folders/jl/mm3wrxqd75l9r1_d0zktphdw0000gn/T/pfs201409498/images

   

   The command runs in your terminal until you terminate it by pressing CTRL+C.

   • Tip Mount multiple repos at once by appending each mount instruction to the same command. For example, the following will mount both repos to the /mydirectory directory.

   pachctl mount ./mydirectory -r first_repo@master -r second_repo@master

   

3. You can check that the repo was mounted by running the mount command in your terminal:

   mount
   /dev/disk1s1 on / (apfs, local, read-only, journaled)
   devfs on /dev (devfs, local, nobrowse)
   /dev/disk1s2 on /System/Volumes/Data (apfs, local, journaled, nobrowse)
   /dev/disk1s5 on /private/var/vm (apfs, local, journaled, nobrowse)
   map auto_home on /System/Volumes/Data/home (autofs, automounted, nobrowse)
   pachctl@osxfuse0 on /Users/testuser/mydirectory (osxfuse, nodev, nosuid, synchronous, mounted by testuser)

   

4. Access your mountpoint.

   For example, in macOS, open Finder, press CMD + SHIFT + G, and type the mountpoint location. If you have mounted the repo to ~/mydirectory, type ~/mydirectory.

   

5. Edit the files as needed.

6. When ready, add your changes to the HPE ML Data Management repo by stopping the pachctl mount command with CTRL+C or by running pachctl unmount <mountpoint> (or unmount /<path-to-mount>, or fusermount -u /<path-to-mount>).

   If you have mounted a writable HPE ML Data Management share, interrupting the pachctl mount command results in the upload of your changes to the corresponding repo and branch, which is equivalent to running the pachctl put file command. You can check that HPE ML Data Management runs a new job for this work by listing current jobs with pachctl list job.