Replace in_tail with a faster implementation.

[ganmacs]

Which issue(s) this PR fixes:
no

What this PR does / why we need it:

slice! is very slow. so I replaced in_tail implementation with slice + freeze.
This change makes FIFO of in_tail about 100x faster in ruby 2.6's simple benchmark.

benchmark script & result: https://gist.github.com/ganmacs/3f517c79881f56547cde305ae7a86b35

Docs Changes:

no

Release Note:

Replace in_tail with faster implementation.

[repeatedly]

Hmm... why slice! is so slow?
slice! is inplace update so memory usage is lower than slice + freeze?

[repeatedly]

How about reducing method call by passing @lines?

# call side
@fifo.parse_lines(@lines)

# implementation
def parse_lines(lines)
  idx = @buffer.index(@eol)

  until idx.nil?
    @buffer.freeze
    rbuf = @buffer.slice(0, idx + 1)
    @buffer = @buffer.slice(idx + 1, @buffer.size)
    idx = @buffer.index(@eol)
    lines << convert(rbuf)
  end
end

[ganmacs]

According to the profiling, slice + freeze version allocate less memory than slice! one.

slice! seems to create a new string object per each call and copy whole value given as an argument.

[ganmacs]

How about reducing method call by passing @lines?

looks good. I confirmed this way is faster.
0b17c04

[repeatedly]

Good.

[repeatedly]

Note: fluentd uses slice! in several code.

% grep -rne 'slice!' lib/fluent/                                                                   (git)-[master]
lib/fluent//plugin_helper/server.rb:753:                @_handler_write_buffer.slice!(0, written_bytes)
lib/fluent//compat/socket_util.rb:85:          @buffer.slice!(0, pos) if pos > 0
lib/fluent//plugin/in_tcp.rb:86:          conn.buffer.slice!(0, pos) if pos > 0
lib/fluent//plugin/in_syslog.rb:195:          buffer.slice!(0, pos) if pos > 0
lib/fluent//plugin/buffer/memory_chunk.rb:51:          @chunk.slice!(@chunk_bytes, @adding_bytes)

[ganmacs]

I will replace them in my spare time.