Skip to content

Materials for the "AI on Kubernetes" tutorial at O'Reilly AI SF 2018

Notifications You must be signed in to change notification settings

dwhitena/oreilly-ai-k8s-tutorial

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

AI on Kubernetes - O'Reilly AI Tutorial

This tutorial will guide you through the deployment and utilization of an end-to-end AI workflow on Kubernetes. Specifically, we will utilize KubeFlow, Pachyderm, and Seldon to deploy a pipeline that:

  • pre-processes a training data set containing GitHub issue data,
  • trains a sequence-to-sequence model to summarize GitHub issues,
  • version controls model binaries,
  • builds Docker images for serving the model, and
  • exports those Docker images to a Docker registry and to Seldon for serving.

This example is based on the modeling code from another KubeFlow example. However, by using Pachyderm to manage this example's data and pipeline, we can create a distributed pipeline that can:

  • Scale to large data sets (by distributing pre-processing and other tasks via Pachyderm built in distributed processing capabilities and support for resources such as GPUs),
  • Flexibly utilize non-distributed TensorFlow outside of a TFJob, non-distributed TensorFlow managed via a TFJob, or distributed TensorFlow (and, actually, any other framework such as PyTorch, Caffe, etc., because Pachyderm pipeline stages support any language/framework and can be managed independently),
  • Version data sets and models (via Pachyderm's built in data versioning, backed by an object store), which is extremely important for compliance and sustainability of pipelines over time,
  • Deploy all the components via KubeFlow, and
  • Still have compatibility to serve the model via a framework like Seldon.

To get the example up and running:

  1. Connect to the tutorial machine
  2. Deploy KubeFlow
  3. Deploy Pachyderm and Seldon on top of KubeFlow
  4. Create a versioned data repository with the training data set
  5. Deploy the pre-processing pipelines
  6. Deploy the training pipeline
  7. Deploy the model build and export pipelines
  8. Serve the model
  9. Update the model

If you get stuck on the example, please reach out the community via:

We also include some resources at the bottom of the tutorial, so you can dig in a little deeper.

1. Connect to the tutorial machine

You should have been given an IP for a cloud instance at the beginning of the course. The cloud instance is already connected to a running Kubernetes cluster and it has all of the command line tools we will be needing throughout the tutorial. Normally, you would be doing all of the following from your local machine, but we've set up these cloud instances to avoid annoying local installations on conference WiFi (and to standardize environments).

Note - If you are following along with a video recording of this tutorial or trying to replicate things after the conference, you should be able to run everything below by deploying KubeFlow, Pachyderm, and Seldon locally via minikube or on GCP via GKE. More information on deploying KubeFlow and Seldon can be found here, and these docs will walk you through deploying Pachyderm locally or in the cloud.

To log into the tutorial cloud instance on Linux or Mac, open a terminal and:

$ ssh pachrat@<remote machine IP>

On Windows you can use PuTTY or another ssh client. You will be asked for a password, which you should also be given during the tutorial.

To verify that everything is running correctly on the machine, you should be able to connect to Kubernetes using the command line tool kubectl:

$ kubectl get all
NAME                 TYPE        CLUSTER-IP   EXTERNAL-IP   PORT(S)   AGE
service/kubernetes   ClusterIP   10.96.0.1    <none>        443/TCP   32m

You will also need to clone the tutorial materials from GitHub to this machine:

$ git clone https://github.com/dwhitena/oreilly-ai-k8s-tutorial.git

2. Deploy KubeFlow

Note - We will be using ksonnet to deploy KubeFlow manually. However, if you are using GKE or minikube in other scenarios, KubeFlow has some convenient scripts and bootstrappers documented here.

First, let's create a namespace for the KubeFlow deployment:

$ NAMESPACE=kubeflow
$ kubectl create namespace ${NAMESPACE}

Then we can initialize a ksonnet app and set our default namespace for the ksonnet app:

$ APP_NAME=my-kubeflow
$ ks init ${APP_NAME}
$ cd ${APP_NAME}
$ ks env set default --namespace ${NAMESPACE}

We need to "install" ksonnet packages for the KubeFlow components we will be using. That requires pulling some files from GitHub. To avoid getting rate limited during the tutorial, you will need to export your GitHub API token to an environment variable. Create a new token or use an existing one from here, then run:

$ export GITHUB_TOKEN=<your-token-from-gh>

It's recommended to install the core Kubeflow infrastructure, which includes the ability to train models with a TFJob CRD. In addition to that, we are going to go ahead and add in the Pachyderm and Seldon components:

$ ks registry add kubeflow github.com/katacoda/kubeflow-ksonnet/tree/master/kubeflow
$ ks pkg install kubeflow/core
$ ks pkg install kubeflow/seldon
$ ks pkg install kubeflow/pachyderm

Now we can deploy the core of KubeFlow:

$ ks generate kubeflow-core kubeflow-core --namespace=${NAMESPACE}
INFO Writing component at '/home/pachrat/my-kubeflow/components/kubeflow-core.jsonnet'

$ ks apply default -c kubeflow-core
INFO Applying configmaps kubeflow.kubeflow-version
INFO Creating non-existent configmaps kubeflow.kubeflow-version
INFO Applying services kubeflow.tf-hub-0
INFO Creating non-existent services kubeflow.tf-hub-0
INFO Applying services kubeflow.tf-hub-lb
INFO Creating non-existent services kubeflow.tf-hub-lb
INFO Applying clusterrolebindings kubeflow.centraldashboard
INFO Creating non-existent clusterrolebindings kubeflow.centraldashboard
INFO Applying roles kubeflow.jupyter-role
...
...

You can then verify that KubeFlow was deployed successfully as follows:

$ kubectl -n kubeflow get svc
NAME               TYPE        CLUSTER-IP       EXTERNAL-IP   PORT(S)    AGE
ambassador         ClusterIP   10.107.0.185     <none>        80/TCP     1m
ambassador-admin   ClusterIP   10.106.237.102   <none>        8877/TCP   1m
centraldashboard   ClusterIP   10.105.218.24    <none>        80/TCP     1m
k8s-dashboard      ClusterIP   10.104.55.209    <none>        443/TCP    1m
tf-hub-0           ClusterIP   None             <none>        8000/TCP   1m
tf-hub-lb          ClusterIP   10.100.25.109    <none>        80/TCP     1m
tf-job-dashboard   ClusterIP   10.104.119.108   <none>        80/TCP     1m

3. Deploy Pachyderm and Seldon on top of KubeFlow

We will be using Pachyderm and Seldon for data storage, versioning, pipelining, and serving. We already added these components to our ksonnet app, we just need to deploy/prep them for use.

To deploy Seldon:

# generate the template for ksonnet
$ ks generate seldon seldon --namespace=${NAMESPACE}

# create cluster role binding seldon-admin
$ kubectl create clusterrolebinding seldon-admin --clusterrole=cluster-admin --serviceaccount=${NAMESPACE}:default

# deploy seldon cluster manager, etc.
$ ks apply default -c seldon

If the Seldon cluster manager deployed successfully, you should see the following pod running:

$ kubectl get pods -n kubeflow | grep "seldon"
seldon-cluster-manager-7f5ddbcf7d-trvfp     1/1       Running   0          1m

To deploy Pachyderm, we follow a similar pattern:

# generate the template
$ ks generate pachyderm pachyderm

# deploy pachyderm
$ ks apply default -c pachyderm

# update pachyderm to the latest version
$ pachctl deploy local --namespace kubeflow

Once deployed, you should see the following Pachyderm daemon running:

$ kubectl get pods -n kubeflow | grep "pach"
pachd-6b589d9988-4zx7v                      1/1       Running   0          2m

And you should be able to communicate with Pachyderm via their CLI pachctl:

$ pachctl version
COMPONENT           VERSION
pachctl             1.7.5
pachd               1.7.5

4. Create a versioned data repository with the training data set

If you followed steps 1-3, you should have an operational KubeFlow + Pachyderm + Seldon cluster. Now we can start managing our data with Pachyderm! Pachyderm version controls all your data sets, so you don't have to worry about committing bad/corrupt data, updating training data, etc. You can always see the full history of your data sets and revert to old versions. Think about this like "git for data" (even though Pachyderm doesn't actually use git under the hood).

To start versioning some data in Pachyderm, we need to create a "data repository":

$ pachctl create-repo raw_data

You will now see this repository listed. It won't have any data versioned in it yet, because we haven't put any data there:

$ pachctl list-repo
NAME                CREATED             SIZE
raw_data            8 seconds ago       0B

Now let's put the data we will use for training into this repo. You could use the full data set of GitHub issues from here, but our model will take a little while to train on that data. As such, we have created a sampled version of the data set that's much smaller and good for experimentation. To "commit" this smaller version of the data set into our raw_data repository:

$ pachctl put-file raw_data master github_issues_medium.csv -f https://nyc3.digitaloceanspaces.com/workshop-data/github_issues_medium.csv

We will now see that file versioned in the data repository:

$ pachctl list-repo
NAME     CREATED        SIZE
raw_data 54 seconds ago 2.561MiB
$ pachctl list-file raw_data master
NAME                     TYPE SIZE
github_issues_medium.csv file 2.561MiB

5. Deploy the pre-processing pipelines

Now, we need to process this training data in the following ways:

  • split the data set into training and tests sets, and
  • pre-process the training data in preparation for training.

To do this, we have create two corresponding Pachyderm pipeline "specifications," split.json and pre_process.json. These specification declaratively tell Pachyderm what Docker image to use to process data, what command to run to process data, which data (repositories) to process, etc. You can specify a bunch of things in these specifications, but these should give you a basic idea.

To create the split and pre_process pipelines (to process the data in the raw_data repository),

$ cd /home/pachrat/oreilly-ai-k8s-tutorial
$ pachctl create-pipeline -f split.json
$ pachctl create-pipeline -f pre_process.json

Immediately a few things will happen:

  • Pachyderm will spin up the necessary pods under the hood to perform our processing:

  • Pachyderm "knows" that we want it to process any data in raw_data with these pipelines, so it will go ahead and spin up these jobs to perform the processing:

    $ pachctl list-job
    ID                               OUTPUT COMMIT                                STARTED            DURATION       RESTART PROGRESS  DL       UL       STATE
    51a94ad5cdc447339e762d1e2f4a61a1 pre_process/a808c5b602f64302a679c7c369593a2a About a minute ago -              0       0 + 0 / 1 0B       0B       running
    ba2f6b00127f484da17353b053b9d0c0 split/778896608e71461db596a93bcf96f7c1       About a minute ago About a minute 0       1 + 0 / 1 2.561MiB 1.026MiB success
    
    $ pachctl list-job
    ID                               OUTPUT COMMIT                                STARTED            DURATION       RESTART PROGRESS  DL       UL       STATE
    51a94ad5cdc447339e762d1e2f4a61a1 pre_process/a808c5b602f64302a679c7c369593a2a About a minute ago About a minute 0       1 + 0 / 1 1.026MiB 1.198MiB success
    ba2f6b00127f484da17353b053b9d0c0 split/778896608e71461db596a93bcf96f7c1       About a minute ago About a minute 0       1 + 0 / 1 2.561MiB 1.026MiB success
    
  • Pachyderm also knows that we have input and output data for each of these split and pre-process stages that needs to be stored and versioned. As such, it creates an output data repository for each of the stages and automatically gathers/versions the output data:

    $ pachctl list-repo
    NAME        CREATED       SIZE
    pre_process 2 minutes ago 1.198MiB
    split       2 minutes ago 1.026MiB
    raw_data    5 minutes ago 2.561MiB
    

5. Deploy the training pipeline

We're now ready to train a sequence-to-sequence model for issue summarization! With Pachyderm + KubeFlow, this is as simple as creating another pipeline specification, train.json, and creating the pipeline:

$ pachctl create-pipeline -f train.json

Similar to our other pipelines, Pachyderm will automatically run the training job, collect the results, and create the output repository to version the results (in this case, the model binary):

$ pachctl list-job
ID                               OUTPUT COMMIT                                STARTED        DURATION       RESTART PROGRESS  DL       UL       STATE
0d25d4d1b120460d8c336a0e293f53e4 train/b2c0d3eddb9e4551871c502ebf01ffb9       5 minutes ago  About a minute 0       1 + 0 / 1 1.198MiB 68.94MiB success
51a94ad5cdc447339e762d1e2f4a61a1 pre_process/a808c5b602f64302a679c7c369593a2a 10 minutes ago About a minute 0       1 + 0 / 1 1.026MiB 1.198MiB success
ba2f6b00127f484da17353b053b9d0c0 split/778896608e71461db596a93bcf96f7c1       10 minutes ago About a minute 0       1 + 0 / 1 2.561MiB 1.026MiB success

$ pachctl list-repo
NAME        CREATED        SIZE
train       5 minutes ago  68.94MiB
pre_process 10 minutes ago 1.198MiB
split       10 minutes ago 1.026MiB
raw_data    13 minutes ago 2.561MiB

$ pachctl list-file train master
NAME                  TYPE SIZE
IssueSummarization.py file 1.224KiB
output_model.h5       file 68.93MiB
requirements.txt      file 74B
seq2seq_utils.py      file 13.81KiB

7. Deploy the model build and export pipelines

Now that we have our model stored and versioned, we want to serve it with some framework like Seldon or TensorFlow serving. Here, we will prepare the model for serving using Seldon, which involves:

  • Converting the model binary to a Seldon "build" artifact, and
  • Building and push a Docker image for Seldon serving based on the build artifact.

As you might have guessed, we have two more pipeline stages that perform these functions, build.json and export.json.

Note, there are a couple important points to consider here and put in place:

  • We are using the Pachyderm Job ID as the tag for the Seldon Docker image. This will allow us to link the Docker images back from Seldon to the jobs, data, Docker images, and pipelines that created them (i.e., full "data provenance").

  • You will need to modify the docker-config.yaml file to include your base64 encoded Docker Hub username and password (we are using Docker Hub here, although you could modify the example and use any registry), and use this file to create a k8s secret with our registry creds. Our pipeline will use these credentials to push the image built in the pipeline:

    $ kubectl create -f docker-config.yaml --namespace kubeflow
    

Once you have that secret, you can create the build and export pipelines as follows:

$ pachctl create-pipeline -f build.json
$ pachctl create-pipeline -f export.json

Again, Pachyderm will automatically run the build and export jobs and version any associated data (Note - the export stage may take some time, like 10+ min, because it's building rather large docker image for serving):

$ pachctl list-job
ID                               OUTPUT COMMIT                                STARTED           DURATION   RESTART PROGRESS  DL       UL       STATE
b496e8c500df43d58d667d113ee8aee7 export/98c3f65fcd3d4f41bebff0e83452f7db      About an hour ago 15 minutes 0       1 + 0 / 1 27.98MiB 0B       success
618dc8220d3b433b95dad3e1680f778d build/190f5440fed749d4aec87761f40e9bc6       About an hour ago 53 seconds 0       1 + 0 / 1 27.95MiB 27.98MiB success
435e6243b4374ba6888d598431200d5d train/a096f70a6ffc43b883285d26dccf4a29       About an hour ago 2 minutes  0       1 + 0 / 1 201KiB   27.75MiB success
226fec9ba9014032a30660b1b1caff38 pre_process/357232558c0a47eeb9f21ac949c06bb3 About an hour ago 2 minutes  0       1 + 0 / 1 104.2KiB 201KiB   success
70eb099852b543ce9818bf11dbe68c76 split/99468d94ac51433e8968362b1015fb2c       About an hour ago 2 minutes  0       1 + 0 / 1 2.561MiB 104.2KiB success
$ pachctl list-repo
NAME                CREATED             SIZE
build               About an hour ago   27.98MiB
export              About an hour ago   0B
pre_process         About an hour ago   201KiB
raw_data            About an hour ago   2.561MiB
split               About an hour ago   104.2KiB
train               About an hour ago   27.75MiB
$ pachctl list-file build master build
NAME                            TYPE                SIZE
build/Dockerfile                file                778B
build/Makefile                  file                74B
build/README.md                 file                367B
build/body_preprocessor.dpkl    file                113.3KiB
build/build_image.sh            file                95B
build/microservice.py           file                6.176KiB
build/model_microservice.py     file                3.536KiB
build/output_model.h5           file                27.75MiB
build/persistence.py            file                1.917KiB
build/proto                     dir                 76.69KiB
build/push_image.sh             file                73B
build/requirements.txt          file                74B
build/seldon_requirements.txt   file                79B
build/title_preprocessor.dpkl   file                29.75KiB

You should also be able to see the Seldon model serving image on Docker Hub under:

<your user>/issuesummarization:<pachyderm job id>

That image is now ready to be used in Seldon to serve the model.

8. Serve the model

We have already deployed the Seldon cluster manager. Now we just need to let Seldon know about the model image that we built in Pachyderm, such that it can spin up inference servers to utilize that image.

Check Docker Hub to find the image tag that you want to utilize for serving (should look like <your user>/issuesummarization:<pachyderm job id>), and then run the following (replacing <your-docker-tag> with the corresponding pachyderm job id that tags your docker image):

$ cd /home/pachrat/my-kubeflow

$ ks generate seldon-serve-simple issue-summarization-model-serving \
   --name=issue-summarization \
   --image=pachyderm/issuesummarization:<your-docker-tag> \
   --namespace=kubeflow \
   --replicas=2

$ ks apply default -c issue-summarization-model-serving

This will spin up two replica inference servers in Kubernetes. You can verify that they are up and running via:

$ kubectl get pods -n kubeflow | grep "issue"
issue-summarization-issue-summarization-77b7975cd9-8gh69   2/2       Running   0          10m
issue-summarization-issue-summarization-77b7975cd9-v4t74   2/2       Running   0          10m

Now let's try out an inference request. First, port forward the inferences services so we can access them:

$ kubectl port-forward $(kubectl get pods -n ${NAMESPACE} -l service=ambassador -o jsonpath='{.items[0].metadata.name}') -n ${NAMESPACE} 8080:80

Then from another terminal (logged into the workshop instance), send a request:

$ curl -X POST -H 'Content-Type: application/json' -d '{"data":{"ndarray":[["issue overview add a new property to disable detection of image stream files those ended with -is.yml from target directory. expected behaviour by default cube should ot process image stream files if user does not set it. current behaviour cube always try to execute -is.yml files which can cause some problems in most of cases, for example if you are using kuberentes instead of openshift or if you use together fabric8 maven plugin with cube"]]}}' http://localhost:8080/seldon/issue-summarization/api/v0.1/predictions
{
  "meta": {
    "puid": "8i536c5upsnc5e04fiusoq6se4",
    "tags": {
    },
    "routing": {
    }
  },
  "data": {
    "names": ["t:0"],
    "ndarray": [["make with with and with and"]]
  }
}

Woohoo! You have set up a complete end-to-end data pipeline with KubeFlow + Pachyderm + Seldon, from raw data to versioned/exported model to model inference!

9. Update the model

All this has been well and good, but there are a bunch of ways to orchestrate the training of a model. Where this approach has HUGE benefits is related to updating and maintaining your model over time. Building a model once is easy, but setting up a system to manage the training data, pre-processing, training, model versioning, and export over time can be overwhelming.

The Pachyderm pipeline that we have deployed takes a unified view of (versioned) data and processing (pipelines). As such, Pachyderm knows when your results aren't up to data with the latest versions of your data or processing. When you update data in one of your repositories or update a pipeline, Pachyderm will automatically update your results. In this case, that means updating your model version and building a new Docker image for model export and serving.

Let's try this out. Try updating your training data (e.g., take out a couple of rows), and then overwriting your previously committed data:

$ cd /home/pachrat/oreilly-ai-k8s-tutorial
$ wget https://nyc3.digitaloceanspaces.com/workshop-data/github_issues_medium.csv
$ vim github_issues_medium.csv
$ pachctl put-file raw_data master -o -f github_issues_medium.csv

(Remember, all this data is versioned, so we could go back at any point to another version of the raw data to revert or review changes)

Right away, you will notice that Pachyderm kicks off the necessary jobs to pre-process and train/build a new model:

$ pachctl list-job
ID                               OUTPUT COMMIT                                STARTED           DURATION   RESTART PROGRESS  DL       UL       STATE
2fb745f718564a81a22e24ab32cc262e train/695367f8c4df4266b46f85482d073926       8 seconds ago     -          0       0 + 0 / 1 0B       0B       running
0174ba12447b486e8e63ddf5949426c4 pre_process/bd2ceaa9e1ff4643bc68061ef7271df1 8 seconds ago     7 seconds  0       1 + 0 / 1 106.8KiB 200.6KiB success
dd0e24ea423341dd91ccf3fca82716a0 split/a50b3f41f4bd4aeba3fa8926a402490f       8 seconds ago     2 seconds  0       1 + 0 / 1 2.56MiB  106.8KiB success
b496e8c500df43d58d667d113ee8aee7 export/98c3f65fcd3d4f41bebff0e83452f7db      About an hour ago 15 minutes 0       1 + 0 / 1 27.98MiB 0B       success
618dc8220d3b433b95dad3e1680f778d build/190f5440fed749d4aec87761f40e9bc6       About an hour ago 53 seconds 0       1 + 0 / 1 27.95MiB 27.98MiB success
435e6243b4374ba6888d598431200d5d train/a096f70a6ffc43b883285d26dccf4a29       2 hours ago       2 minutes  0       1 + 0 / 1 201KiB   27.75MiB success
226fec9ba9014032a30660b1b1caff38 pre_process/357232558c0a47eeb9f21ac949c06bb3 2 hours ago       2 minutes  0       1 + 0 / 1 104.2KiB 201KiB   success
70eb099852b543ce9818bf11dbe68c76 split/99468d94ac51433e8968362b1015fb2c       2 hours ago       2 minutes  0       1 + 0 / 1 2.561MiB 104.2KiB success

$ pachctl list-job
ID                               OUTPUT COMMIT                                STARTED       DURATION   RESTART PROGRESS  DL       UL       STATE
3319d503a06548cdb1e7ecb5da052527 export/2670af46ba4e40ff878f438f6b22c287      4 minutes ago -          0       0 + 0 / 1 0B       0B       running
18228a20d9ee47848d7a4949fd25689d build/caded47edb564039a6e318641b6eee0a       4 minutes ago 29 seconds 0       1 + 0 / 1 27.85MiB 27.88MiB success
2fb745f718564a81a22e24ab32cc262e train/695367f8c4df4266b46f85482d073926       4 minutes ago 28 seconds 0       1 + 0 / 1 200.6KiB 27.65MiB success
0174ba12447b486e8e63ddf5949426c4 pre_process/bd2ceaa9e1ff4643bc68061ef7271df1 4 minutes ago 7 seconds  0       1 + 0 / 1 106.8KiB 200.6KiB success
dd0e24ea423341dd91ccf3fca82716a0 split/a50b3f41f4bd4aeba3fa8926a402490f       4 minutes ago 2 seconds  0       1 + 0 / 1 2.56MiB  106.8KiB success
b496e8c500df43d58d667d113ee8aee7 export/98c3f65fcd3d4f41bebff0e83452f7db      2 hours ago   15 minutes 0       1 + 0 / 1 27.98MiB 0B       success
618dc8220d3b433b95dad3e1680f778d build/190f5440fed749d4aec87761f40e9bc6       2 hours ago   53 seconds 0       1 + 0 / 1 27.95MiB 27.98MiB success
435e6243b4374ba6888d598431200d5d train/a096f70a6ffc43b883285d26dccf4a29       2 hours ago   2 minutes  0       1 + 0 / 1 201KiB   27.75MiB success
226fec9ba9014032a30660b1b1caff38 pre_process/357232558c0a47eeb9f21ac949c06bb3 2 hours ago   2 minutes  0       1 + 0 / 1 104.2KiB 201KiB   success
70eb099852b543ce9818bf11dbe68c76 split/99468d94ac51433e8968362b1015fb2c       2 hours ago   2 minutes  0       1 + 0 / 1 2.561MiB 104.2KiB success

Once all the jobs (split, pre-process, train, build, and export) finish, you will have an branch new model image ready for serving, but you will also be able to tie this, or another, model image back to the exact jobs, data, and code that produced it! One way you can access this information is with Pachyderm's built in commit inspection API, which returns "data provenance" (i.e., all the data and processing specs that lead to a specific commit of data):

$ pachctl inspect-commit train a096f70a6ffc43b883285d26dccf4a29
Commit: train/a096f70a6ffc43b883285d26dccf4a29
Started: 2 hours ago
Finished: 2 hours ago
Size: 27.75MiB
Provenance:  __spec__/495a01e3df3a4318bb90b46d307d4009  __spec__/3526704d403041cd8514eccccd0539fe  __spec__/69bb10e11e9e45d2ae559176b40c0c95  pre_process/357232558c0a47eeb9f21ac949c06bb3  raw_data/cdd1ab3ced9048d49f63459433554761  split/99468d94ac51433e8968362b1015fb2c

Resources:

Slides from the class can be found here.

KubeFlow:

Pachyderm:

Seldon:

About

Materials for the "AI on Kubernetes" tutorial at O'Reilly AI SF 2018

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published