The index builds architecture guide includes a subsection called "Temporary Table for Key Generation Errors" which states "on both primaries and secondaries, the records associated with key generation errors are skipped and recorded in a temporary table, the skipped record table." Per this comment, this is not entirely true. Primaries do not suppress errors other than duplicate key errors, and instead abort immediately, propagating the error to the user.

Update the architecture guide to clarify that primary nodes abort index builds immediately after key generation errors and thus do not re-generate or re-insert records using the SkippedRecordsTracker.

The motivation for this ticket was the conversation here.