MLflow Supported Model Flavors and Model Serving: JohnSnowLabs vs SparkNLP

[rickyschools]

Hi all,

My team has been using MLflow for some time to test, evaluate, and serve our models. Really appreciate the work done here by the maintainers and the community.

Recently, we have some fine-tuned language models that leverage SparkNLP ML Pipelines for batch inference in Job Clusters in Databricks. We're overhauling our ML serving pipelines/system are trying to figure out how to put those models into an MLflow deployment (with our end target to leverage Databricks Real-Time model serving).

The MLflow docs have cited experimental support of johnsnowlabs models since ~2.4 , but this seems to only be the enterprise version. Was there a strategic reason/decision to not support OS SparkNLP? It's not a blocker for us to move forward, but something we noticed as we've been experimenting with the model flavor. If there wasn't a strategic reason to not support it, have you heard others struggling or asking for this feature? Happy to help contribute here, just wondering if the juice is worth the squeeze.

https://mlflow.org/docs/2.4.2/models.html#john-snow-labs-johnsnowlabs-experimental
https://mlflow.org/docs/2.12.1/models.html#john-snow-labs-johnsnowlabs-experimental
https://sparknlp.org/docs/en/serving_spark_nlp_via_api_databricks_mlflow

Thanks in advance!

RS