The chain of Enums are greedily creating a lot of intermediate data that you then end up dropping before the DF operations.

Consider using a lazy Stream to process each file up to the filter_with one at a time, and then join the results together at the end.

Ftp.local_paths_ls(:marketaxess)
|> Stream.take(3)
|> Stream.filter(fn x -> String.contains?(x, "mabond") end)
|> Stream.map(&read_csv/1)
|> Stream.filter(fn df -> DF.n_rows(df) > 0 end)
|> Enum.map(fn df ->
      DF.filter_with(df,
         &Explorer.Series.equal(&1["SETTLEMENTDATE"], 
           Explorer.Series.mode(df["SETTLEMENTDATE"])[0]))
   end)
|> DF.concat_rows()

As for your “messy” slicing (if it’s even necessary using the streams approach). DF.slice/2 can enumerate a lazy Elixir Range struct, so you could do:

|> Enum.map(fn df ->
      n_rows = DF.n_rows(df)
      range = 0..n_rows//skiprate # pretty sure DFs are 0-indexed

      df
      |> DF.filter_with(
         &Explorer.Series.equal(&1["SETTLEMENTDATE"], 
           Explorer.Series.mode(df["SETTLEMENTDATE"])[0])
         )
      |> DF.slice(range)
    end)
|> DF.concat_rows()