At Spotify we have many time series and the challenge is to forecast them all considering the model tuning. In this talk I'm going to present how to distribute Prophet or other Python models to decrease runtime from days to hours.
Spotify has many time series to be forecasted such as streams forecast, monthly active users, ads inventory and consumption, etc.
Prophet library has been very effective in capturing most of the time series requirements, however tuning is necessary as the default set of parameters don't perform well on our more noisy datasets with many confounding factors.
Spotify serves in 78 markets globally, as these countries get smaller the timelines become noisier and further demand tuning of parameters for reasonable results. To add to these more global factors constant product design changes impact these time series significantly and make it hard to be predictable.
Hence scaling the solution is paramount in such Data Science projects, as of today there is no off the shelf solution to scaling the solution using an existing python library, so we made a stack for it.
In this talk we will discuss: The stack of this solution Tuning at scale Solutions benefits Solutions limitations