The slop process will automatically attempt to repair writes that partially failed. The only cases where read repair is actually used would be when: a) the slop backlog is large and the read happens before the slops got processed, or b) because the node holding
the slops died irrecoverably, and those slops were lost forever.
In practice the inconsistencies should be pretty rare... If that is not the case, then perhaps there is a bigger issue preventing the above two mechanisms (slops and read repair) from working properly (very frequent node failures, as might be the case in certain
cloud configurations or certain defective batches of hardware? or perhaps you have a multi-datacenter deployment with very flaky links? etc.)