You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository was archived by the owner on Jul 28, 2026. It is now read-only.
Arrow and polars default to a NaN is NULL style behavior, while cudf doesn't for aggregations. In the normal reductions (min/max -- maybe also sum/mean, but not sure), we currently account for this inside the task.
Since arrow does this, we should also account for this inside group-by aggregations.
In practice a nans_to_nulls function may be useful, but so long we mix arrow and libcudf it seems likely easier to solve this at the lower rather than higher level with pre-processing.
Similar behavior is that in some places NaN handling in either polars or cudf or arrow might be that NaNs are considered equal to NULLs or not.
Aligning these (similar to for reductions/groupbys) is probably impossible without nans_to_nulls (or work-arounds in other places).
Arrow and polars default to a NaN is NULL style behavior, while cudf doesn't for aggregations. In the normal reductions (min/max -- maybe also sum/mean, but not sure), we currently account for this inside the task.
Since arrow does this, we should also account for this inside group-by aggregations.
In practice a
nans_to_nullsfunction may be useful, but so long we mix arrow and libcudf it seems likely easier to solve this at the lower rather than higher level with pre-processing.Similar behavior is that in some places NaN handling in either polars or cudf or arrow might be that NaNs are considered equal to NULLs or not.
Aligning these (similar to for reductions/groupbys) is probably impossible without
nans_to_nulls(or work-arounds in other places).