On master, commit bf232a7 introduced the much welcomed support of dropping NaN/null fields. But it has done so by assembling line by line with function format_line().

This has degraded the performance significantly. See profiling result attached diagram of how format_line() is taking the majority of consumed time.

Writing a 86400-row dataframe that took only about <10 seconds with v5.0.0 now takes about 90 seconds in my test.