Details

New Feature

Status: Resolved

Major

Resolution: Implemented

None
Description
The FLIP design doc can be found at https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=184615783.
The existing Flink ML library allows users to compose an Estimator/Transformer from a pipeline (i.e. linear sequence) of Estimator/Transformer, and each Estimator/Transformer has one input and one output.
The following usecases are not supported yet. And we would like to address these usecases with the changes proposed in this FLIP.
1) Express an Estimator/Transformer that has multiple inputs/outputs.
For example, some graph embedding algorithms (e.g., MetaPath2Vec) need to take two tables as inputs. These two tables represent nodes labels and edges of the graph respectively. This logic can be expressed as an Estimator with 2 input tables.
And some workflow may need to split 1 table into 2 tables, and use these tables for training and validation respectively. This logic can be expressed by a Transformer with 1 input table and 2 output tables.
2) Express a generic machine learning computation logic that does not have the "transformation" semantic.
We believe most machine learning engineers associate the name "Transformer" with the "transformation" semantic, where the a record in the output typically corresponds to one record in the input. Thus it is counterintuitive to use Transformer to encode aggregation logic, where a record in the output corresponds to an arbitrary number of records in the input.
Therefore we need to have a class with a name different from "Transformer" to encode generic multiinput multioutput computation logic.
3) Online learning where a longrunning Model instance needs to be continuously updated by the latest model data generated by another longrunning Estimator instance.
In this scenario, we need to allow the Estimator to be run on a different machine than the Model, so that the Estimator could consume sufficient computation resource in a cluster while the Model could be deployed on edge devices.
4) Provide APIs to allow Estimator/Model to be efficiently saved/loaded even if state (e.g. model data) of Estimator/Model is more than 10s of GBs.
The existing PipelineStage::toJson basically requires developer of Estimator/Model to serialize all model data into an inmemory string, which could be very inefficient (or practically impossible) if the model data is very large (e.g 10s of GBs).
In addition to addressing the above usecases, this FLIP also proposes a few more changes to simplify the class hierarchy and improve API usability. The existing Flink ML library has the following usability issues:
5) fit/transform API requires users to explicitly provide the TableEnvironment, where the TableEnvironment could be retrieved from the Table instance given to the fit/transform.
6) A Pipeline is currently both a Transformer and an Estimator. The experience of using Pipeline is inconsistent from the experience of using Estimator (with the needFit API).