Mount notebooks into a mounted fs volume to avoid data loss

[timothyjlaurent]

The current way this is done:

docker run -p 8888:8888 -it --rm $USER/assignments

creates an ephemeral container due to the --rm flag. In the first notebook you are doing very expensive operations (gunziping) several GBs. All of this data will be lost when the container shuts down. Using the filesystem as a mounted volume, allows subsequent sessions to have access to the files created (eg pickled data). This is especially important because this is course material so want to avoid time consuming gotchas for the students.

[tensorflow-jenkins]

Can one of the admins verify this patch?

[googlebot]

Thanks for your pull request. It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

📝 Please visit https://cla.developers.google.com/ to sign.

Once you've signed, please reply here (e.g. I signed it!) and we'll verify. Thanks.

---

• If you've already signed a CLA, it's possible we don't have your GitHub username or you're using a different email address. Check your existing CLA data and verify that your email is set on your git commits.
• If you signed the CLA as a corporation, please let us know the company's name.

[timothyjlaurent]

I signed it!

[googlebot]

CLAs look good, thanks!

[vincentvanhoucke]

@craigcitro any gotcha going down that route?
Should I also be advertising not using --rm for users of the pre-built Docker container?

[craigcitro]

No, these both seem pretty reasonable. Two thoughts:

1. The --rm thing is definitely a smart move -- I tend to spend more time debugging docker images than using them right now. For this case, though, it does make it a bit too easy to lose notebooks.
2. Mounting the notebooks dir as a volume also makes sense. This is one where mounting it ("I want to make changes that get saved") and not mounting it ("I'm playing around with something and it's a throwaway") both make sense, so maybe some text explaining when you would or wouldn't want the option? As a minor thing, no need to cd -- just pass the subdir path instead of $(pwd)?

[vincentvanhoucke]

@timothyjlaurent thanks for the fix. Would you mind changing the doc for the hosted container as well and removing the 'cd'?

[timothyjlaurent]

Ok @vincentvanhoucke I updated the readme.

[vincentvanhoucke]

s/tesorflow/tensorflow/

[timothyjlaurent]

I fixed the typo

[vincentvanhoucke]

LGTM @vrv

[vrv]

Thanks! Squash the commits and we'll merge.

[timothyjlaurent]

Commits have been squashed.

There is one caveat to this PR -- if the user needs to rebuild the docker image -- all the files underneath will need to be added to the context for the build. This is several GB for the first assignment so it takes a LONG time.

[craigcitro]

We aren't running any docker build commands from the top of the tree, though, are we? (Just the one docker run?)

[vincentvanhoucke]

Merged. Thanks!

[timothyjlaurent]

Well in the readme it says:

 Building a local Docker container
 ---------------------------------

     cd tensorflow/examples/udacity
     docker build -t $USER/assignments .

If you use that same directory to mount the notebooks then it will fill up with data files. Then if you need to remake the image all of that will have to be put in the context. When I had to do this, I just moved the data directories temporarily to .. and then back in.

So maybe we should recommend copying the udacity dir to some work dir location and then mount that into the container or mention the issue with context and explain that they should move extraneous files elsewhere while rebuilding the image.

[vincentvanhoucke]

I'm not sure why you, as a class user, would ever want to build the container?
These instructions are for us, when we want to update the hosted containers.

[timothyjlaurent]

The course materials instructed me to build the image:

If there is an image I could pull, that would definitely be better, but there is no mention of this image in the course materials as far as I can see.

[vincentvanhoucke]

Arpan, there should be no need for anyone to build a container. Was the doc updated?

[tensorflow-jenkins]

Can one of the admins verify this patch?

[tensorflow-jenkins]

Can one of the admins verify this patch?

[timothyjlaurent]

If there is a hosted image that should be used for the class, the command should look something like this:

docker run --rm -v <path to notebooks>:/notebooks tensorflow/udacity

[vincentvanhoucke]

@timothyjlaurent Access to the hosted version is described on the first line in the document you edited:
docker run -p 8888:8888 -it --rm b.gcr.io/tensorflow-udacity/assignments
Maybe I should separate the rest of this readme so that it stands out more.

[timothyjlaurent]

Oh I see it now that you point it out. Maybe the course should be updated to instruct users to use that image rather than build their own.

[timothyjlaurent]

Also now that I see the first line (although it wasn't shown in the course material)
This should probably also mount the working directory as a volume.
It takes maybe 20-30 minutes to decompress the data files. It also loads all this data into memory -- I had to upgrade my docker machine VM to have 8GB of RAM to avoid running out of memory. All of this should be disclosed to the student.

[vincentvanhoucke]

@timothyjlaurent we pushed some changes this weekend that should help with memory.