I have an ES7.2 index set up with a refresh_interval of 30s. Docs get inserted, a notification is sent to a SQS queue with a 31 sec delivery delay, and then a job reads from the queue and does a search against that index. The vast majority of the time, the new doc is searchable as expected. Sometimes though, it's not. When the task gets tried again in a little while, the doc is always there.
Am I just taking refresh_interval too literally? Maybe if the cluster is busy and the in-memory buffer is large it could take a few more seconds to get the new segment written and opened?