There does not seem to be an absolute rule, though "new DOI for a new version" seems most common. Speak to whoever's hosting your data in case they have a different preference.
The California Digital Library summarises various approaches here. Some require registering a new DOI for each new tranche of the data; some will reuse the DOI but use a new version number or date to help you identify the relevant bit; some distinguish between major and minor additions.
It's worth noting that the "generate a new DOI for a new version" approach assumes that you're adding, say, a new month's data to the set with each version. If it's being updated on a daily or hourly basis, this approach breaks down; you don't want to generate dozens of DOIs for only marginally different versions! Here, the "snapshot" approach recommended by the Digital Curation Centre is much more efficient - produce a static copy of the dataset as it currently stands, on an as-needed basis or at standard intervals, and cite that.
Update: the new STM Report on scholarly publishing, out today, notes that the "The RDA Data Citation Working Group is investigating possible technical solutions [for dynamic data]" without giving much more detail (p. 140); the most recent material to come out of that group seems to be this workshop report from last year.