A Dataflow DoFn fails on malformed input and you must reliably capture and reprocess bad records. What is the correct design change?

The correct approach is to use a try...catch block within your DoFn to gracefully handle malformed records. When an error occurs, the catch block should direct the erroneous record to a sideOutput. This creates a separate PCollection containing all bad records, which can then be written to a persistent store like Pub/Sub for later analysis or reprocessing. Filtering to skip errors (option 1) is not reliable as it discards data without capturing it for review. Extracting from logs (options 1 and 2) is inefficient and unreliable for programmatic reprocessing, as logs are primarily for monitoring, not data capture. Writing directly to Pub/Sub from within the DoFn (option 3) is generally discouraged in Dataflow as it can lead to performance issues and makes the pipeline less robust due to direct I/O operations from within a processing step. Side outputs are the idiomatic Dataflow way to handle divergent data paths.