Performance falls rapidly as max_hits increases in size

Was poking around trying to isolate some weird sorting behavior and noticed that query times increase rapidly as the size of max_hits goes up. This is running a build from main, single node with local file system storage.

Index info

1. General infos
===============================================================================
Index id:                           noaa-1m
Index uri:                          file:///Users/kstaken/projects/opensource/quickwit-test/qwdata/indexes/noaa-1m
Number of published splits:         2
Number of published documents:      1634467
Size of published splits:           84 MB

$ curl -sS 'http://127.0.0.1:7280/api/v1/noaa-1m/search?query=name:NO&max_hits=10' | jq .elapsed_time_micros
46021
$ curl -sS 'http://127.0.0.1:7280/api/v1/noaa-1m/search?query=name:NO&max_hits=100' | jq .elapsed_time_micros
68718
$ curl -sS 'http://127.0.0.1:7280/api/v1/noaa-1m/search?query=name:NO&max_hits=1000' | jq .elapsed_time_micros
252134
$ curl -sS 'http://127.0.0.1:7280/api/v1/noaa-1m/search?query=name:NO&max_hits=10000' | jq .elapsed_time_micros
2832835
$ curl -sS 'http://127.0.0.1:7280/api/v1/noaa-1m/search?query=name:NO&max_hits=100000' | jq .elapsed_time_micros
null

For the 100,000 record case the server panics as it's out of file descriptors.

2022-05-13T18:36:10.966Z ERROR fetch_docs: quickwit_search::fetch_docs: Error when fetching docs in splits. split_ids=["01G2Z5RC4880BTWNSKX3P6YM9N", "01G2Z4B7QHBRE31G3203BWVV6N"] error=searcher-doc-async

Caused by:
    0: An IO error occurred: 'Too many open files (os error 24)'
    1: Too many open files (os error 24)
2022-05-13T18:36:13.817Z ERROR fetch_docs: quickwit_search::fetch_docs: Error when fetching docs in splits. split_ids=["01G2Z5RC4880BTWNSKX3P6YM9N", "01G2Z4B7QHBRE31G3203BWVV6N"] error=searcher-doc-async

Caused by:
    0: An IO error occurred: 'Too many open files (os error 24)'
    1: Too many open files (os error 24)

Reproduce the single-node local-file-system queries against the noaa-1m index using the shown curl commands and compare elapsed times as max_hits increases. Start from the search endpoint and the fetch_docs error context; done means large max_hits requests no longer cause rapid degradation or Too many open files failures.