Hi Peter, I think what you referred is typical amendment process where partial or all results need to modified. I think it is definitely interesting topic! Here is my two cents
In ideal world, reference data source can ingest updated used values as events and join with buffered events in windows . (it’s a bit counter intuitive, but think there is a magic function where we ingest all reference data as stream instead of doing on demand rpc) Unfortunately, in lots of use cases, it seems hard to know exactly how reference data source used and dump reference data costs too much. So replay pipeline might be cheapest way to get things done in general. In some cases, results are partitioned and bounded. It makes possible to recomputed within bounded windows, that may requires a bit work to customize window which hold longer than watermark pass its endtime. I remember there was a Jira talk about retraction. In other cases, results are derived from long history which makes not rationale to keep. A side pipeline capture those events with late arriving event handling might interact with external storage and amend results. Thanks, Chen From: Peter Lappo Sent: Sunday, September 10, 2017 3:00 PM To: user@flink.apache.org Subject: ETL with changing reference data hi, We are building an ETL style application in Flink that consumes records from a file or a message bus as a DataStream. We are transforming records using SQL and UDFs. The UDF loads reference data in the open method and currently the data loaded remains in memory until the job is cancelled. The eval method of the UDF is used to do the actual transformation on a particular field. So of course reference data changes and data will need to reprocessed. Lets assume we can identify and resubmit records for reprocessing what is the best design that * keeps the Flink job running * reloads the changed reference data so that records are reprocessed in a deterministic fashion Two options spring to mind 1) send a control record to the stream that reloads reference data or part of it and ensure resubmitted records are processed after the reload message 2) use a separate thread to poll the reference data source and reload any changed data which will of course suffer from race conditions Or is there a better way of solving this type of problem with Flink? Thanks Peter