Answer
Each request has a fixed cost on the cluster: a network round trip, a thread, a write to the transaction log and a refresh to schedule. One bulk request of a few hundred documents pays that cost once, so indexing takes a small share of the nodes and leaves the rest for queries. The price is a little delay: a document waits until its batch is full or a short timer fires.
Why
Flush on whichever comes first, a size or a time limit, so a quiet stream still reaches the index within a second or two. A bulk request reports a result per document, so retry only the ones that failed.