At the moment (7.4) the fulltext is extracted by a job using the default pool and saved into the db with another job (fulltext udpater) using a dedicated pool with a single thread to serialize db access.
The serialization is needed for some database backend that does not support concurrent update on a field with a fulltext index.
When creating large amount of documents this can be a bottleneck, the updater work has a lower throughput and the queue size increase until it generates GC storm.
When the fulltext index is only done with Elasticsearch using nuxeo.vcs.fulltext.search.disabled=true. The extraction process can be optimized using a single job instead of 2, also these worker should have a dedicated pool so it can be easily tuned.
Indexing flow is optimized by using ES bulk indexing features, when fulltext indexation is disabled at the repository level.