Hi! As you can see there is some progress on ZBXNEXT-922. For now forecasting will be implemented by the means of trigger functions. We want them to be as robust and easy to use as possible. Therefore we stick with such common extrapolation techniques like linear regression, polynomials, exponential, logarithmic and power functions. I hope this is just a beginning and as soon as we get feedback on this feature we will probably proceed with something more complicated.
It is very interesting that you use Holt-Winters method. As I can see, there are three-four parameters in this model. This is probably too much for an ordinary user to tune. And we would not want to provide "default" parameters because it is very hard to find parameter set which would suit everyone. Wiki says about exponential smoothing:
Such techniques have broad application that is not intended to be strictly accurate or reliable for every situation.
And this is what concerns us the most.
Anyway, I would be glad to hear more from you about your vision of forecasting in Zabbix and about your experience with Holt-Winters.
Hello,
The difference between ZBXNEXT-922 and what I'm proposing is that 922 uses triggers identify capacity issues as they emerge, whereas this one gives an estimation of when that might be. These two complement each other more than compete. While it is ultimately helpful to know when you are about to reach a potential capacity issue, sometimes you also need to answer a question similar to this one: "If we add another server, based on our current growth of application X, how long until we would need to add another server?" While triggers handle most of capacity management, this proposal handles capacity planning.
Being able to see forecasts in graphs allows you to plan for capacity, rather than react to it. It's great that there is some forecasting being done via triggers, now I think we just need to put that into graphs as well. If the extrapolation techniques used for triggers would be suited for graphing as well, then that's fine. I chose Holt-Winters at the time because it was recommended for forecasting things with fluctuations, like CPU usage.
Hi,
ZBXNEXT-922 will allow you to create predictive triggers which can act and prevent problem from happening in an automated manner. Of course, same trigger functions can be used in calculated items and plotted on graphs (but this does not look too pretty at the moment). There are plans to transfer prediction maths to graphs (purely on the frontend side) when trigger functions are done and polished.
Fluctuations are unpredictable by nature. 
Sometimes Holt-Winters is able to guess them surprisingly well, but sometimes all can go totally wrong.
I'll point out that rrdtool implements Holt-Winters and is C code released under GPLv2 (or later), so the code can be directly copied into Zabbix.
for simple functions, calculating it on the fly as you access it can work, but for more complex things, it makes more sense to calculate it once and store the result the way you would any other data source.
looking at ZBXNEXT-992 it looks like it is an on-demand function to calculate the expected value now (or in the future with timeleft), but doesn't provide a way to have the history of the predictions short of running the calculation for each point on the graph as it's displayed.
Is there a way to have a predictive trigger create a data item that can then be graphed as you would any other item?
Any update with this feature?