A synthetic continuous-profiling workload with ten nullable dictionary-string
labels. Fill rates range from 0% to 100%, with cardinalities from 4 to 800.
Consecutive samples repeat label sets, with lowercase values such as comm_0.
NULL labels do not match positive or negative string predicates.
Stacks are lists of location structs containing lists of line structs, with dictionary-encoded scalar children. Profile metadata uses dictionary strings, period and duration use Int64, and timestamps use nanoseconds in UTC.
The queries exercise:
- Narrow and wide time windows with profile metadata filters and stack projection.
- Raw equality and regular expressions on a label.
- Time-bucket sums, label grouping, distinct labels, mapping files and profile types.
- A positive conjunction of lowercase equality, substring, prefix, suffix and membership filters across labels, with stack projection.
- A negative conjunction of lowercase inequality, absent substring, prefix, suffix and nonmembership filters, with time-bucket sums.
- An equivalent positive ILIKE conjunction with the same stack projection. Escaped underscores match literal underscores rather than wildcard characters.
At scale factor 1, the generator writes one million rows in deterministic 10,000-row batches. Parquet uses Zstd level 3. Generic conversion writes standard or compact Vortex files, whose physical encodings are selected by the writer.
Parquet-versus-Vortex timings include representation differences. String filters can execute above the scan, so inspect physical plans separately from timings. The string values are lowercase ASCII. Lowercase filters exercise function evaluation but do not change these values or measure Unicode performance.
CI runs this suite from local NVMe as the PolarSignals Profiling PR comment.
It runs DataFusion over Vortex only, without a Parquet drift control.
vx-bench run polarsignals --engine datafusion --format vortex