Data Lineage

This is the most crucial step in Big Data debugging. The captured lineage is combined and processed to obtain the data flow of the pipeline. The data flow helps the data scientist or a developer to look deeply into the actors and their transformations. This step allows the data scientist to figure out the part of the algorithm that is generating the unexpected output. A big data pipeline can go wrong in two broad ways. The first is a presence of a suspicious actor in the data-flow. The second being the existence of outliers in the data.

The first case can be debugged by tracing the data-flow. By using lineage and data-flow information together a data scientist can figure out how the inputs are converted into outputs. During the process actors that behave unexpectedly can be caught. Either these actors can be removed from the data flow or they can be augmented by new actors to change the data-flow. The improved data-flow can be replayed to test the validity of it. Debugging faulty actors include recursively performing coarse-grain replay on actors in the data-flow, which can be expensive in resources for long dataflows. Another approach is to manually inspect lineage logs to find anomalies, which can be tedious and time-consuming across several stages of a data-flow. Furthermore, these approaches work only when the data scientist can discover bad outputs. To debug analytics without known bad outputs, the data scientist need to analyze the data-flow for suspicious behavior in general. However, often, a user may not know the expected normal behavior and cannot specify predicates. This section describes a debugging methodology for retrospectively analyzing lineage to identify faulty actors in a multi-stage data-flow. We believe that sudden changes in an actor’s behavior, such as its average selectivity, processing rate or output size, is characteristic of an anomaly. Lineage can reflect such changes in actor behavior over time and across different actor instances. Thus, mining lineage to identify such changes can be useful in debugging faulty actors in a data-flow.

The second problem i.e. the existence of outliers can also be identified by running the data-flow step wise and looking at the transformed outputs. The data scientist finds a subset of outputs that are not in accordance to the rest of outputs. The inputs which are causing these bad outputs are the outliers in the data. This problem can be solved by removing the set of outliers from the data and replaying the entire data-flow. It can also be solved by modifying the machine learning algorithm by adding, removing or moving actors in the data-flow. The changes in the data-flow are successful if the replayed data-flow does not produce bad outputs.