Finding the Chicago Taxi Trips Demo Prediction Output

So, one of the key new demos for the San Jose demo at O'Reilly's Architecture Conference will be the Chicago Taxi fare estimator demo. This is a Kubeflow bundling of a pre-existing TensorFlow demo that uses a dataset of Chicago taxi fares published in Google's Big Data.

The key question I sat down with at the start of the day was "how can I demonstrate some real time predictions at the end of this?" After going through all of the data and pages around this, the conclusion to which I've come is that the "deploy" step winds up setting up a TF Serving instance, which the "prediction" step winds up calling with a set of data. It then deposits the results of this training on a rok filesystem was cloned at the start of the pipeline.

So, the question becomes - how can one get access to that cloned file system?

The best answer I've come up with is to first get the name of the pod spun up to run TF Serving by looking at the logs of the Deploy step. The name you're looking for is highlighted below.

Now, run "vagrant ssh" in the same directory as your Vagrantfile to log into the Virtual Box running all of this. Run "kubectl get pods -n kubeflow" to get a list of all the Kubeflow pods. Look for a pod that is still running that starts with the boxed name in the image above -- this will be the pod that is running TensorFlow Serving for this pipeline. The appropriate line is boxed below.

You can do "kubectl exec" to this box to log in, as shown below. Notice that the first line of input pops up a little mangled. Don't worry about that - just hit Ctrl-C and proceed.

All the data used by the pipeline can be found under /mnt, but the prediction results specifically are in the prediction_results-00000-of-0001 file shown below. Adjust the exact path for the name of your pod, of course.