Bloom filters were already disabled (still indexed, but not used/loaded) by
default as of 1.4.0, so that shouldn't explain the change, unless in your
1.4.5 test you had enabled them?
Those segment counts look healthy (merges are keeping up), and that's good
that there's no indexing nor store throttling.
How many concurrent threads are indexing, and how many docs per bulk
If you change to refresh_interval: 10s for both 1.4.5 and 1.7.1 does it
change the result? This may increase effective concurrency of the test
since a separate thread is writing segments.
If you drop index.number_of_replicas to 0, does that change the result?
We track indexing performance nightly at
https://benchmarks.elastic.co/index.html ... it's somewhat noisy, and it
tracks master (not 1.x releases), but performance has not changed that much
for the "fast" settings over time.