To be sure what you mean here: you plan to calculate a spatially global statistic, so you get a 1D timeseries, which you want to use for picking or dropping observations from another band/collection. E.g. global statistic is 42 on Feb 4 → use observations of Feb 4, global statistic is 7 on Feb 14 → drop Feb 14 observations, …?
And did that work, or what error do you get?
If it doesn’t work, have you already tried to resample your data down heavily before aggregating? If you just want quantiles (or statistics in general), downsampling is probably not going change the results much.
I don’t think we have short term implementation plans for “reduce_spatial”.
Also note that a generic reduce_spatial is far from trivial in a big EO data system, where you want to split the work in separate chunks. Simple reducers like sum, count and mean are easily doable because they trivially allow splitting the calculation. Quantiles (as requested here apparently) are a lot harder to do in a splittable way. Long story short: we will probably have to re-implement (if even possible) various reducer processes for usage in reduce_spatial, so it’s a big task to tackle.