Skip to main content

0009 — The full rolling-hour series is built from RV

Status: accepted (2026.11.0)

Context​

full_rolling_series extends the forecast_rolling_1h attribute of Peak hourly precipitation next 120 min by the 12 hours that started before the release. RS used to feed the sensor with the option on, decoding all 25 members, so RS downloaded its whole archive every 5 minutes instead of reading three members by range. Measured on the 576 RS archives on OpenData from 11 to 13 Oct 2026, that was 4.6 MB a release against 0.9 MB, about 1.3 GB a day (0.66 GB on the dry day, 2.1 GB on the wet one), and it doubled an install's traffic, from about 0.93 to 2.0 GB a day.

Earlier live checks had shown that RS's forecast hours are RV's 5-minute steps added up. On 13 and 14 Oct 2026, every RS member of two releases (one wet, one moderate) matched, cell for cell on the whole grid, the sum of the twelve RV steps in its hour: the RV analyses (_000) of the releases up to T, then RV T's forecast. That includes the members whose hour started before T. A missing step counts as 0, and RS has no data exactly where all twelve steps have none.

Decision​

RV builds the whole series. It keeps the analysis of each release it adopts for an hour, in memory, and with the option on adds the 12 earlier windows from them. RS always reads its three members by range. A live test (test_live_rs_is_summed_rv) checks the identity daily.

The 12 points appear only when all 11 earlier analyses are known. Missed ones are not fetched, so after a restart, or a release that was never adopted, the series has the 13 future points for up to an hour.

Alternatives​

  • Fetch the missed analyses in a background round, as the rain-events tracker does: a complete series right after a restart, for about 2 MB and the ordering care such a round needs. Not worth it for an attribute that is not recorded.
  • Keep RS for the series, reading only members _000–_055 (the first 12, about half the archive) by range: halves the cost instead of removing it.
  • Drop RS altogether and build Precipitation last 60 min from RV's analyses too, about 260 MB a day less. Not done: it changes the source of released sensors, and would make them depend on RV and on an hour of history.

Consequences​

With the option on, an install downloads what it does with the option off, and RS no longer decodes 25 members a release. The option no longer changes the peak's value, window or future points: they come from RV either way. An install that had the option on read them from RS before, so in the rare cells where DWD's RS and summed RV differ (see the later measurement below) its peak may move slightly with this change. If DWD stops publishing RS as summed RV, the live test fails; the series' first point and Precipitation last 60 min would then disagree.

Later measurement (17 Oct 2026)​

The identity is not exact on every release. Twelve releases 4 h apart from 15 to 17 Oct matched in every cell while dry. On the wet morning of 17 Oct, 48 consecutive releases of RS lead 0 matched summed RV exactly in 98.2 to 100 % of the wet cells; at most 0.44 % differed by more than 0.01 mm and 0.11 % by more than 0.05 mm, the largest by 0.54 mm. A step that differs stays in the sums for the hour it covers. The forecast leads deviate the same way, less often. RS and RV never disagreed about which cells have data.

The decision stands: the deviations are DWD's own, in a percent or two of the wet cells, on an attribute that is not recorded. The live test checks the identity by its share rather than exactly (the same cells have data, and at least 90 % of wet cells match), so DWD's deviations do not fail it and a systematic change, which would put most wet cells off, still does.