Nishi Compare

Data Engineering & Analytics

Nishi's sovereign, integer-exact analytics stack — measured against dbt, Spark, DuckDB, ClickHouse, Arrow, pandas/Polars, Superset, DataSketches and OpenLineage. Every score is graded by nx_dataeng_sota_census with a liar-kill (each axis opens a real organ); this page is generated from that state, never hand-asserted.

616meet-SOTA / 1000
10Have
13Partial
5Gap
2Exceed

30 axes · 0 ungrounded (liar-kill armed). Started this arc at 333 — the area flagged "toy level"; +283 across 8 gated organs.

The honest read: real primitives existed but sat unwired. A gated analytics engine was built on them — point at columns (in-memory or from the sovereign store), get exact or approximate aggregates, a plain-English profile, the relationships between them, trends over time, and query them. Integer-exact means bit-reproducible aggregates — a determinism edge float tools cannot match.

The gated analytics stack

OrganGateCapability
nx_dataframe9/9exact aggregation: count / sum / min / max / mean / variance / stddev / quantile / filter / GROUP BY / join
nx_dataframe_approx5/5HLL distinct + t-digest quantile at scale, measured vs exact truth (HLL 6 permil error)
nx_analyst_data6/6profile any column + plain-English read (categorical / id / constant / skew / outliers)
nx_analyst_multi8/8Pearson correlation + strength class + key-driver finder
nx_analyst_report5/5"analyze this dataset" front door — composes the whole stack into one report
nx_timeseries6/6rolling mean + least-squares trend + growth over time
nx_analyst_store4/4runs the whole stack on real stored data — seg_store scan → extract → analyze
nx_query7/7query executor: SELECT count/sum/avg/min/max(col) WHERE col op val GROUP BY col, integer-exact

Census axes vs SOTA

AxisStatusGrounding
Aggregation front doorHAVEcount/sum/avg/min/max/percentile over a column, integer-exact (nx_dataframe)
Descriptive stats (general)HAVEmean/var/stddev/quantile over arbitrary columns
Approx analytics at scaleHAVEHLL distinct + t-digest, measured vs exact (killed the unmeasured-claim debt)
Multivariate correlationHAVEPearson + key-driver; discovered price~5*size-2*age from data alone
General data analystHAVEprofile + find patterns + explain a whole dataset (nx_analyst_report)
Stored-data analyticsHAVEthe analyst runs on real seg_store records (round-trip proven)
Query enginePARTIALSELECT agg WHERE GROUP BY executor, integer-exact (nx_query 7/7); string-syntax parser = finer gap
Timeseries analyticsPARTIALrolling/trend/growth done; resample + seasonality = finer gap
Quantile sketchPARTIALt-digest tail-tight (p99 10 permil) but median coarse; interpolation follow-on
Columnar formatPARTIALnx_seg_columnar RLE+dict primitive; unwired to a query layer
Data lineagePARTIALnx_store_lineage: organ->store PRODUCER/CONSUMER edges (OpenLineage-class)
Data contractsPARTIALnx_seg_schema: Avro-class schema + backward/forward evolution-safety
Transform DAG (dbt-class)GAPmodels + data tests + docs + semantic layer
Vectorized / distributed execGAPbatch/vectorized + multi-node (blocked-adjacent: the 1-core threading gap)
BI dashboardGAPquery-driven interactive dashboards (needs the UI arc)
Arrow interchangeGAPzero-copy columnar interchange format

Sovereign exceeds

AxisStatusWhy beyond mainstream
Integer-exact analyticsEXCEED100%-integer aggregates are bit-reproducible; DuckDB/pandas float sums are order-dependent, ours are not
Own-stack, zero depsEXCEEDthe whole primitive stack (HLL++/t-digest/columnar/store) is sovereign — no Arrow/DataSketches/JVM

Honest remaining gaps

Five GAPs — a transform-DAG, vectorized + distributed execution (blocked on the 1-core threading gap), a BI dashboard, and Arrow interchange. The next make-it-real step is scaling the stored-data path past the 256-per-key store cap. Every score here is measured and re-runnable.

Generated by nx_dataeng_compare_page composing the sovereign S-class kit — TLS 1.3, no third-party web server, no external fonts or assets. Back to the compare hub