Hi Filip We (alibaba & tencent) are doing the apache iceberg row-level update/deletes POC, syncing the change log (such as row-level binlog) into iceberg data lake is the classic case we are trying to implement (another classic case would be one streaming job or batch job with one or more update SQL sentences). we currently prefer to use <Lazy with NRI> solution discussed in update/delete doc [1] if the change log table defined the unique key, because we can append the unique keys into differential files while if <lazy with SRI> solution we need to read the row identifiers for each change log entry which would limit the ingest throughput a lot (bad experiences for data lake I think) . Of course maybe there're some tables which don't have their unique key, I considered that the iceberg users can also choose to use the <lazy with SRI> although their throughput are limited but can work (the compute engine spark/flink may need to shuffle the change logs into batches for each iceberg partition so that we can get the rowId in batch and write the differential files...it's a draft :-) ). In general, we may focus on the cases which table have a unique key.
[1]. https://docs.google.com/document/d/1Pk34C3diOfVCRc-sfxfhXZfzvxwum1Odo-6Jj9mwK38/edit?usp=sharing
