Data Observability

Data Quality & Testing is the assertions you write. Data observability is the monitoring that runs in production: freshness SLAs, row-count anomalies, schema drift, column-level statistics — usually paired with Data Catalog & Lineage. For application observability (Sentry / OTel / Prometheus), see Observability & Errors and Prometheus Stack.

Open-source / free-tier

• ★ Elementary Data — runs over dbt artifacts; freshness, schema-change, anomaly, dbt-test results in one UI. Apache 2.0 OSS + paid Elementary Cloud. The default OSS pick if you live in dbt.
• ★ Datafold's data-diff — diff two tables / two warehouses / two dbt environments; MIT. The default for "did my refactor change the output?"
• Soda Core — assertions framework; lighter on ongoing monitoring than Elementary; Apache 2.0; see Data Quality & Testing. Hosted Soda Cloud adds anomaly + alerting.
• Re_Data — dbt-native anomaly detection; MIT; less active in 2026 than Elementary.
• Whylogs (WhyLabs) — log statistical profiles per dataset / model; Apache 2.0; pairs with WhyLabs Cloud.
• Pydeequ / Deequ (Amazon) — Spark-based; Apache 2.0; for Spark-heavy environments.

Hosted / paid

• ★ Monte Carlo — the category leader; paid; "data downtime" framing; small free trial. Pricey but the most polished.
• Bigeye — same niche; paid; ML-driven thresholds.
• Acceldata, Anomalo, Sifflet, Lightup, Validio — competitors in the data-observability SaaS space.
• Datafold Cloud — adds CI / column-level lineage on top of data-diff.
• Metaplane — modern data-observability SaaS; small free tier.

Lakehouse-native observability

• Apache Iceberg metadata tables (snapshots, files, manifests) — query directly for compaction / freshness signals.
• Delta Lake history + Databricks Unity Catalog observability — paid Databricks features.
• OpenLineage events → Marquez / DataHub / OpenMetadata for lineage-based anomaly signals; see Data Catalog & Lineage.

ML-specific drift / monitoring

• ★ Evidently AI — open-source ML monitoring; data + concept drift; Apache 2.0.
• NannyML — concept drift specifically; Apache 2.0.
• Whylogs / WhyLabs, Arize, Fiddler, Truera — broader ML observability; mostly paid.
• See AI/LLM and AI Evals for LLM-specific observability.

Patterns to adopt

• Freshness SLAs per asset. "This table must be < 1h stale" is the most actionable observability signal; alerts on it pay back fast.
• Row-count anomalies. Yesterday's count vs. 7-day-rolling-avg; cheap and catches a lot.
• Schema drift alerts. Column added / dropped / type changed → notify owner. Most catalogs and observability tools include this.
• Column-level statistics. Min/max, null %, cardinality — store nightly and diff. Surfaces "loaded the wrong file" silently-bad-data fast.
• Lineage-aware alerts. When the upstream orders_raw is stale, suppress downstream noise; Elementary / Monte Carlo do this.
• CI data-diff on every PR. Datafold's data-diff against staging vs. prod for changed dbt models; catches accidental SQL semantics changes.

Pick this if…

• Default OSS for dbt teams: Elementary.
• Default CI diffs: Datafold's data-diff.
• You want assertions + observability in one OSS tool: Soda Core + Soda Cloud.
• Pay for polish: Monte Carlo or Metaplane.
• ML pipelines: Evidently or NannyML.
• Lakehouse-native: Iceberg metadata + OpenLineage + a catalog.