INDY-1251: change math of spikes detection.

[sergey-shilov]

Changed math of node request and cluster throughtput spikes detection.
The main changes:

• Now smothing constant (alpha) is really constant as declares exponential
  moving average (EMA), alpha = 2 / (N + 1), where N is a minimal count of
  passed periods needed for comparison of expected value and real value, a.g.
  learning period. For now N = 15, so alpha = 0.125.
• Rised min activity from 2 to 10, a.g. we start to compare expected and real
  values if we handle more than 10 request per time period, that now is set
  to 60 secs (10 requests per minute).
• Rised borders coefficient to extend the lower and higher borders when observed
  value is treated as normal relative to expected value, previous border coefficient = 3:
  • lower border: expected_val / 3
  • higher border: expected_val * 3
    new borders coefficient = 10:
  • lower border: expected_val / 10
  • higher border: expected_val * 10
    Also implemented weighted borders coefficient. It means adaptation of
    borders coeffitient to larger values using formula:
    weighted_borders_coef = borders_coef / log(expected_val)
    Such approach allows to lower borders for larger values. This
    functionality is optional and enabled by default.

Signed-off-by: Sergey Shilov [email protected]