Quality icons

The 58 quality icons in Data Engineering, described

Annotatorannotator

An annotator — the person the labels come from, a human labeler.

Data anomalyanomaly-data

A data anomaly — the row that does not belong, an outlier or a wrong value.

Bucketbucket-data

Bucket — every value routed into its bin so the data can be grouped.

Canonicalisecanonicalize

Canonicalise — all the spellings become the one standard spelling.

Class balanceclass-balance

Class balance — as many examples of one class as of the other.

Constraintconstraint-db

A database constraint — what the database refuses to store, not-null or unique.

Data augmentdata-augment

Data augmentation — more rows derived from the rows you already have.

Data contractdata-contract

A data contract — the schema both producer and consumer agreed on.

Data contract breakdata-contract-break

A document split by a crack — a data contract broken by an upstream change.

Data qualitydata-quality

Data quality — checks that records are complete, valid and what they claim to be.

Data validationdata-validation

Data validation — records checked against the shape and rules they must satisfy.

Dataset carddataset-card

A dataset card — what this data is, where it came from and how to use it, on one page.

Dataset splitdataset-split

A dataset split — most rows to learn from, some held back to be judged on.

Dataset versiondataset-version

A dataset version — the data exactly as it was, kept as a snapshot in the lineage.

Fuzzy dedupededupe-fuzzy

Fuzzy dedupe — nearly the same is treated as the same and merged.

Deduplicatededuplicate

Deduplicate — two of the same become one, distinct and unique rows.

Distribution shiftdistribution-shift

Distribution shift — the world moved and the data followed, so the curve is not where it was.

Duplicateduplicate

A duplicate — the same row twice, a repeated record that should be one.

Entity resolveentity-resolve

Entity resolution — two spellings, one person, records matched to one identity.

Expectationexpectation

An expectation — the rule the data has to keep, a contract asserted on every load.

Featurefeature

A feature — the column of input a model actually reads as a signal.

Feature crossfeature-cross

A feature cross — two columns multiplied into a third to capture their interaction.

Feature driftfeature-drift

Feature drift — a column that is no longer what it was in March, its distribution wandered.

Freshness checkfreshness-check

A freshness check — is this still current, recently verified and not stale?

Gold labelgold-label

A gold label — the answer the graders agreed on, verified ground truth.

Group bygroupby

A bracket gathering three rows together — grouping rows by a shared key.

Holdoutholdout

A holdout — rows fenced off that the model never gets to see until evaluation.

Labellabel

A label — the class or tag a row is taught under, an annotation.

Label queuelabel-queue

A label queue — rows waiting to be annotated with their names.

Label reviewlabel-review

Label review — somebody looks twice at the name a row was given, QA on tags.

Lineage nodelineage-node

A lineage node — this table as one step in the chain that produced it.

Noise injectionnoise-inject

Noise injection — a little grit added on purpose to make a model robust.

Normalisenormalize-data

Normalise data — everything scaled into the same band or range.

Nullnull

Null — there is nothing here, an empty or missing value in the field.

Null checknull-check

A null check — is anything missing, empty or blank in this record?

Null fillnull-fill

A dot filling the gap in a row — replacing missing values with a default.

One-hotone-hot

One-hot encoding — a vector of all zeros with a single one marking the category.

Outlier removaloutlier-remove

Outlier removal — the extreme value that was never really data, dropped.

Perturbperturb

Perturb — the same row shaken slightly with jitter or noise to test robustness.

Pivotpivot

Rows turning into columns — pivoting a table from long to wide.

Qualityquality

Quality — good enough to trust, graded against a standard and checked.

Quality gatequality-gate

A quality gate — data does not pass into the next stage until it checks out.

Record linkrecord-link

Record link — these two rows are the same story, paired across sources.

Retention policyretention-policy

A retention policy — how long data is kept before it is deleted, a TTL.

Row count checkrow-count-check

Row count check — as many rows as there were supposed to be, verified.

Sample rowssample-rows

Sample rows — a few rows picked out to look at as a preview of the data.

Schemaschema

A schema — the shape the data has to have, its tables, fields and types.

Schema diffschema-diff

A schema diff — the same table one column later, compared for changes.

Schema pinschema-pin

Schema pin — the columns held still on purpose, a frozen contract.

Stalenessstaleness

Staleness — how long since this was last true, the age of outdated data.

Stratified samplestratified-sample

A stratified sample — one draw from every layer so each stratum is fairly represented.

Synthetic flagsynthetic-flag

A synthetic flag — made-up rows carry a marker saying they were generated.

Synthetic rowsynthetic-row

A synthetic row — a record nobody collected, generated to fill a gap.

Target leaktarget-leak

Target leakage — the answer key fell into the training set and the model cheated.

Unpivotunpivot

Columns turning into rows — unpivoting a table from wide to long.

Volume anomalyvolume-anomaly

One bar spiking far above its neighbours — a sudden anomaly in data volume.

Weak labelweak-label

A weak label — a heuristic guess wearing a name tag, noisy but useful.

Winsorizewinsorize

A clip line cutting across the tallest bar — capping extreme values at a threshold.

In code, each is one import — import { Annotator } from "@iconmind/react/icons/annotator" — and the same name in Vue, Svelte, Solid, Preact, React Native, Astro, Blade and Flutter.

Narrow it by tag