Speed up CSV writer

Further improvements:

- simply moving "require 'fastercsv'" outside of csv_writer reduces the time to 2.9s. (I guess that 'require' is very expensive when rubygems is around)

- if you change build_table_body as follows:

    def build_table_body
      fcsv = csv_writer
      data.each { |row| fcsv << row }
    end

then it still takes about 2.9s. However this improvement is now independent of the possibly controversial render_data_by_row patch.

- there is another issue. In my renderer, if I replace

      data.detail.as(:csv, :io=>output)

with

      FCSV(output) { |fcsv|
        fcsv << data.detail.column_names
        data.detail.each { |row| fcsv << row }
      }

then the time drops to 1.1 seconds. In fact, even if I make build_table_body an empty method, data.detail.as(...) still takes 1.6s, without generating any output other than the column headings!

Profiling shows that the issue is Renderer taking a fullblown copy (or two) of the table:

  %   cumulative   self              self     total
 time   seconds   seconds    calls  ms/call  ms/call  name
 13.33    13.90     13.90    26312     0.53     4.93  Array#each
 11.66    26.06     12.16    26302     0.46     4.97  Ruport::Data::Table#recordize
 11.05    37.59     11.53    92064     0.13     1.44  Kernel.dup
 10.63    48.68     11.09    92082     0.12     0.21  Kernel.send
  7.81    56.83      8.15    13151     0.62     1.12  Ruport::Data::Record#initialize
  6.00    63.09      6.26    26303     0.24     1.40  Ruport::Data::Table#record_class
  5.31    68.63      5.54    13151     0.42     3.13  Ruport::Data::Table#normalize_hash
  4.28    73.09      4.46    92061     0.05     0.05  Module#===
...
  0.00   104.31      0.00        2     0.00 52100.00  Ruport::Renderer#data=
..
  0.00   104.31      0.00        1     0.00 104200.00  Ruport::Data::Table#initialize_copy
..
  0.00   104.31      0.00        2     0.00 52165.00  Ruport::Renderer#render
..
  0.00   104.31      0.00        1     0.00 104300.00  Ruport::Renderer::Hooks.as

This can be eliminated simply by changing

  def data=(val)
    formatter.data = val.dup
  end

to

  def data=(val)
    formatter.data = val
  end

I'm not sure if a deep copy is warranted here, when simply telling the formatter about the source object. If it is there may be a cheaper way to do it, e.g. Marshal.load(Marshal.dump(..))

With these changes my report now takes 1.0-1.1s to write the CSV table to disk, more than 30x faster than when I started.

patch applied: r1255

Remember, the bug tracker is not a mailing list! :)

Please see further discussion at:

Replying to sandal:

patch applied: r1255 Remember, the bug tracker is not a mailing list! :) Please see further discussion at:

http://groups.google.com/group/ruport-dev/browse_thread/thread/88d48fc664e597f8