What's the status of TensorFlow data service?

TensorFlow data service seems to be a promising API to separate data processing from model training.

However, it’s still in the experimental namespace. The last commit seems to have been made 6 months ago.

So it’s not clear if this is a popular API in production environments. Is it likely for this API to be provided as part of the main TensorFlow API (moved out of experimental)? Or is this something that doesn’t have a lot of successful adoption and might be phased out?

Thank you.

@ganeshparameswaran,

Welcome to the Tensorflow Forum!

Yes, it is still in the experimental namespace and development is going on.

For the tf.data service workers, a new 905aa22 commit was merged yesterday to eliminate explicit prefetch. Please refer to the master branch for the most recent updates.

Thank you!

@chunduriv Thanks for sharing the latest update on TF data service!

I have a quick follow-up question: if we decide to adopt TF data service in our model training pipeline, considering it’s still in the experimental namespace, what’s the level of support we can expect from TF team when we find any issues/bugs? What’s the best communication channel for this experimental feature?

Thank you very much!

@z5sh,

You can post issues or bugs related to the TF data service on Github, where Tensorflow engineers are active.

Thank you!