Chengxiang Li created FLINK-2396: ------------------------------------ Summary: Review the datasets of dynamic path and static path in iteration. Key: FLINK-2396 URL: https://issues.apache.org/jira/browse/FLINK-2396 Project: Flink Issue Type: Improvement Components: Core Reporter: Chengxiang Li Priority: Minor
Currently Flink would cached dataset in static path as it assumes that dataset stay the same during the iteration, but this assumption does not always be true. Take sampling for example, the iteration data set is something like the weight vector of model and there is another training dataset from which to take a small sample to update the weight vector in each iteration (e.g. Stochastic Gradient Descent), we expect sampled dataset is different in each iteration, but Flink would cache the sampled dataset as it in static path. We should review how Flink identify dynamic path and static path, and support add sampled dataset in above example to dynamic path. -- This message was sent by Atlassian JIRA (v6.3.4#6332)