org.elasticsearch.xpack.logsdb.LogsdbIndexModeSettingsProvider#newIndexHasSyntheticSourceUsage is one of the big contributors to the many shards benchmarks regressions we're seeing.
The reason for that is that it does a redundant mapping merge exactly like the one we do anyway. We need to find a more efficient solution here, this is severely impacting at least the productivity of working with the many-shards benchmarks but also has production impact on master node stability during mass index rollover events and the like.

See below profile (note this is 30% of CPU overall, including JIT and GC which are quite heavy here):