Suspected data loss (and/or duplicates) bug in BigQueyrServicesImpl #20891
Description
When this API yields errors specific to failed inserts for a row.
Rows are selected here for retrying, using the errorIndex which is returned from the error.
retryRows.add(rowsToPublish.get(errorIndex));
However, this errorIndex is not valid to index rowsToPublish. So it looks like the wrong rows are being selected to be retried.
Why can't you use errorIndex to index rowsToPublish?
because rowsToPublish contains all of the rows which were passed into insertAll.
These are then batched into a smaller list of "rows" , where multpile API calls are made to bigquery to insert the rows.
The errors returned actually refer to the list of rows passed into the call made to BigQuery, so they are only valid indices for "rows". Thus, they are not valid indices for "rowsToPublish".
Note: These lists have a different number of rows: rowsToPublish.size() > rows.size()
Imported from Jira BEAM-12139. Original Jira may contain additional context.
Reported by: ajamato@google.com.