# Column-level lineage answers questions that table-level lineage cannot, but it costs more to build and has gaps

GUIDE, read 10 Nov 2026

**The short answer.** Table lineage shows which tables feed which; column lineage shows which columns feed which. All 16 tools in the Upstream matrix document both, yet the gaps differ: dbt reflects select statements only, Atlan excludes JOIN, FILTER and GROUP BY, and Databricks shows source-to-target links without transformation expressions found.

## Table lineage tells you which tables feed which

Table-level lineage is a graph of tables and views. An edge says that one table was built from another. Snowflake's lineage, per its docs, tracks data movement such as CTAS, INSERT and MERGE statements and object dependencies such as a view on a table.

That is enough for a first question: if this table is wrong, what sits upstream? It is not enough for the second question, which is the one an analytics engineer asks after tracing a wrong number: which column in the upstream table did the bad value come from, and what happened to it on the way?

A wide table can feed fifty downstream tables, and only three of them may read the column you are about to change. Table-level lineage reports all fifty. The result is an impact list that is too long to trust, so people stop reading it.

## Column lineage tells you which columns feed which

Column-level lineage records, for each output column, the input columns used to produce it. OpenLineage defines this as a columnLineage dataset facet. Euno's docs say it tracks relationships at column level, not just table or model level. dbt's Catalog says it gives end-to-end column-level lineage for each column of a model, source or snapshot.

Some tools add a third layer: the transformation. OpenLineage's facet carries a per-column transformation type and subtype, such as DIRECT with subtype TRANSFORMATION or IDENTITY, plus a masking flag. That describes the kind of transformation, not the SQL expression. dbt's column evolution lens goes further, showing when a column is transformed versus passed through or renamed, and previewing the formula.

All 16 tools in the matrix document table-level and column-level lineage. The differences are in how the column edges are built and what they leave out.

## Column lineage is derived in four different ways

How a tool derives column edges decides where it is wrong. Four methods show up in the evidence.

- **Parsing SQL.** DataHub says its parser, built on sqlglot, generates column lineage for most of its integrations, and claims 97 to 99 percent accuracy in its own benchmarks. Atlan parses SQL for sources such as Snowflake, BigQuery, Redshift and dbt. SYNQ parses SQL and flags lineage as uncertain when parsing fails.
- **Reading query history.** Coalesce Catalog builds column-to-column lineage from observed SQL in the warehouse, and says truncated or unsupported SQL can limit column detail. Snowflake builds lineage from its own query history, which began in November 2024.
- **Analyzing source code.** Foundational parses transformation code rather than query logs.
- **Capturing at runtime.** Databricks Unity Catalog captures lineage automatically for queries run on Databricks, down to the column level.

## Parsing can fail, and good tools say so

A parser-based column edge is only as good as the SQL it can read. dbt says its column lineage in Catalog is built from SQL parsing of models and can be incomplete when parsing fails. SYNQ flags lineage as uncertain when parsing or dependency extraction fails. Coalesce Catalog says truncated or unsupported SQL can limit column detail. Atlan lists a lineage mechanism per source type, and says source-to-warehouse hops outside supported pairs need a CSV, name matching, OpenLineage or its SDK.

Runtime capture avoids some of this, because the engine reports what it executed. It has its own limit. Databricks documents column links from source to target, and we found no documentation that Unity Catalog stores the transformation expression of a column (none found on 10 Nov 2026). You learn that column A fed column B. You do not learn from that page whether B is A plus tax.

Treat a missing edge and a wrong edge as different failures. A flag such as SYNQ's uncertain marker tells you to check by hand. A silent gap tells you nothing.

## Joins and filters are where column lineage thins out

A column used only in a JOIN or WHERE clause does not appear in the output, so a parser that follows the SELECT list never sees it. If someone changes that column, the output rows change, and the lineage graph reports no impact. Vendors handle this differently, and several state the limit plainly.

| Tool | Documented treatment of joins and filters | Source type |
| --- | --- | --- |
| dbt | Column lineage reflects select statements only, not joins and filters | dbt docs |
| Atlan | Column lineage covers direct and conditional transformations; JOIN, FILTER and GROUP BY are excluded | Atlan docs |
| Snowflake | Lineage excludes tables used only for joins or filters | Snowflake docs |
| Monte Carlo | Shows SELECT-clause field lineage, and WHERE-clause relationships as field-to-table; some JOIN+USING cases and wildcard tables are not covered | Monte Carlo docs |

## Some sources stay at table level

Atlan says column-level lineage covers relational and SQL sources, and that native object stores and message queues are table-level only. Collibra says it does not generate column-level lineage for tables created by SQL statements unless the SQL is supplied through a shared storage connection. Snowflake says column lineage is not supported for semantic views, and that lineage is not available for objects in shared databases or the SNOWFLAKE database.

These are documented limits, not bugs. The practical reading is that a column lineage claim needs a scope: which sources, which statement types, which clauses.

## What each tool documents, level by level

The table below is built from the matrix block in each tool's evidence file as of 10 Nov 2026. A value of yes means the tool's own pages document it. Partial means documented with a stated limit or for a subset of systems. None found means no page read documented it.

| Tool | Table | Column | Transformation logic | BI field | Metric |
| --- | --- | --- | --- | --- | --- |
| Alation | yes | yes | partial | partial | none found |
| Atlan | yes | yes | yes | partial | none found |
| Coalesce Catalog | yes | yes | partial | partial | partial |
| Collibra | yes | yes | yes | partial | partial |
| Databricks | yes | yes | none found | partial | none found |
| DataHub | yes | yes | yes | yes | none found |
| dbt | yes | yes | yes | partial | partial |
| Euno | yes | yes | partial | yes | partial |
| Foundational | yes | yes | yes | yes | none found |
| Manta (IBM) | yes | yes | yes | partial | none found |
| Monte Carlo | yes | yes | yes | partial | none found |
| OpenLineage | yes | yes | partial | none found | none found |
| Secoda | yes | yes | partial | partial | partial |
| Select Star | yes | yes | partial | yes | partial |
| Snowflake | yes | yes | partial | partial | partial |
| SYNQ | yes | yes | yes | partial | none found |

## When table-level lineage is enough

Table lineage is enough for ownership routing, freshness alerts and a quick scan of what a dropped table breaks. Column lineage earns its cost when you rename or retype a column, when a field in a dashboard is wrong, or when you must show where a sensitive column went.

Cost matters too. The tools that capture lineage from query logs need the queries to have run, so a model that has never executed has no edges yet. Coalesce Catalog's graph uses the last 30 days of SQL. Foundational says it can give column lineage for assets never queried, because it reads code instead of logs. Which of these matters depends on whether your estate has many rarely-run tables.

If your own work is mostly the second kind, test the column edges on your own SQL before you buy. Pick a model with a join, a window function and a CASE expression. See which of those edges the tool draws. [The method page](/method/) describes how Upstream scores depth, and [the matrix](/matrix/) shows the same grid for all 16 tools.

## Sources

- Snowflake: docs.snowflake.com/en/user-guide/ui-snowsight-lineage.md: https://docs.snowflake.com/en/user-guide/ui-snowsight-lineage.md (read 10 Nov 2026)
- OpenLineage: openlineage.io/docs/spec/facets/dataset-facets/column_lineage_facet: https://openlineage.io/docs/spec/facets/dataset-facets/column_lineage_facet (read 10 Nov 2026)
- Euno: docs.euno.ai/getting-started/core-concepts.md: https://docs.euno.ai/getting-started/core-concepts.md (read 10 Nov 2026)
- dbt: docs.getdbt.com/docs/explore/column-level-lineage: https://docs.getdbt.com/docs/explore/column-level-lineage (read 10 Nov 2026)
- SYNQ: docs.synq.io/lineage/column-level-lineage.md: https://docs.synq.io/lineage/column-level-lineage.md (read 10 Nov 2026)
- DataHub: docs.datahub.com/docs/lineage/sql_parsing: https://docs.datahub.com/docs/lineage/sql_parsing (read 10 Nov 2026)
- Atlan: docs.atlan.com/product/capabilities/lineage/best-practices/capture-and-troublesh: https://docs.atlan.com/product/capabilities/lineage/best-practices/capture-and-troubleshoot-lineage (read 10 Nov 2026)
- Collibra: productresources.collibra.com/docs/collibra/latest/Content/CollibraDataLineage/r: https://productresources.collibra.com/docs/collibra/latest/Content/CollibraDataLineage/ref_technical-lineage-supported-data-sources.htm (read 10 Nov 2026)
- Coalesce Catalog: docs.coalesce.io/docs/catalog/navigate/lineage.md: https://docs.coalesce.io/docs/catalog/navigate/lineage.md (read 10 Nov 2026)
- Foundational: docs.foundational.io/en/articles/14046538-explore-data-lineage.md: https://docs.foundational.io/en/articles/14046538-explore-data-lineage.md (read 10 Nov 2026)
- Monte Carlo: docs.getmontecarlo.com/docs/field-lineage: https://docs.getmontecarlo.com/docs/field-lineage (read 10 Nov 2026)
- Databricks: docs.databricks.com/aws/en/data-governance/unity-catalog/data-lineage: https://docs.databricks.com/aws/en/data-governance/unity-catalog/data-lineage (read 10 Nov 2026)

Last reviewed 10 Nov 2026. Canonical: https://datalineagecompare.com/guides/table-vs-column-lineage/
