Styxis
Company
Knowledge

Check observation presence when a singleton GPU datetime mean is NULL

Use the identity that one valid instant has that same mean. Distinguish operation support from CPU-fallback requirements.

Observation presence and mean presence

For one valid observation, the mean should be that observation under the approved instant-mean definition. Do not record a NULL aggregate as original absence and silently remove an existing observation. Hand off valid-observation count together with aggregate presence. Separately extracted valid-observation and actual-mean presence form the Truthound comparison inputs. [1, 2, 3]

The follow-up narrows the fallback issue

Polars #25616 reports a singleton constant datetime mean retained on CPU but returned as NULL on GPU. A maintainer follow-up frames the task as correcting CPU fallback for an unsupported GPU mean. Check the actual installed support combination rather than assuming current behavior from the report. [1]

Retain valid original instants

Retain valid original instants per group and actual mean presence. Hold NULL aggregates where valid observations exist while testing a verified engine alternative. Extract valid counts, timezones, and units from the source, checking actual instant agreement separately from presence. [1, 3]

A presence reference at 1970-01-12

Illustrative group G1 contains the single valid instant 1970-01-12T00:00:00. Its mean must exist as that instant, independently requiring presence true. Actual true passes; a NULL represented by false fails. After passing presence, still compare the actual mean instant with the original. [1, 3]

A presence flag is not the mean value

ColumnPairValidator in Truthound Python SDK 3.3.2 compares actual_mean_present and required_mean_present prepared by the operator from raw observations and actual aggregation. Units, timezones, and execution are supplied externally. Core does not run GPU datetime aggregation, and presence alone does not establish the correct mean instant. [1, 2, 3]

CPU and epoch alternatives

Compare a verified CPU path or explicit epoch aggregation with unit and range checks. Existing regressions for the same engine and dtype can reduce another audit’s value. A fallback must retain the defined timestamp meaning rather than merely produce a nonmissing number. [1, 3]

Application conditions

Extract valid original instants, counts, units, timezones, and actual mean presence separately.

Alternatives

Compare verified CPU or explicit epoch-aggregation paths alongside unit and range tests.

Limitations

  • Presence does not establish the correct mean instant or support for GPU aggregation.

Evidence

Event checked
Subject checked
Modified