Is this issue cropping up during the stream (real time) enrichment process or the batch enrichment process?
Overall your setup looks good but in the stream enrich process it is possible for this to occur if an event referencing that schema is sent before the schema is available on S3. The enrich process will record this as a cache miss so future events will go to bad as long as that reference is in the cache. The quick solution here is to manually evict the cache. There’s some more information on the cache and eviction in this thread.