YALineo
EN
Install

Read Unity Catalog lineage and measure the impact of a change

Unity Catalog records the reads and writes of each table when it can infer them. Before you change or drop a table, that lineage tells what depends on it. Here is where to read it, what it misses, and what YALineo does.

Updated on 6 November 2026 · Abdouramane Youssouf

Where Databricks shows lineage

Unity Catalog records the lineage of queries run on Databricks, down to the column, across all the workspaces of the same metastore. Databricks lists impact analysis among its uses: before you change or drop a table, find the tables, jobs and dashboards that depend on it.

In Catalog Explorer, a table's Lineage tab lists the related tables. See Lineage Graph opens the graph on one level, which each node expands. Jobs then Downstream gives the jobs that consume the table, Dashboards its dashboards, and a click on a column shows the related columns. The time range shown is one year by default; nothing captured before 1 September 2024 is available.

The permissions you need

A masked node is still a dependency: someone with the rights can tell what lies behind it.

The lineage system tables

Two read-only system tables give lineage in SQL: system.access.table_lineage, one row per read or write of a table or a path, and system.access.column_lineage, the same per column. By default, only users who are both account admin and metastore admin can read them; for others, an admin grants USE CATALOG, USE SCHEMA and SELECT.

The columns that matter for impact:

These tables keep a rolling 395 days by default. Beyond that, Catalog Explorer and the lineage API remain.

The direct downstream of a table

What read main.silver.orders over 90 days, and what those reads wrote:

SELECT coalesce(target_table_full_name, target_path) AS downstream,
       target_type,
       entity_metadata.job_info.job_id AS job_id,
       entity_metadata.notebook_id AS notebook_id,
       entity_metadata.dashboard_id AS dashboard_id,
       entity_metadata.sql_query_id AS sql_query_id,
       count(*) AS lineage_rows,
       max(event_date) AS last_day
FROM system.access.table_lineage
WHERE source_table_full_name = 'main.silver.orders'
  AND event_date >= date_sub(current_date(), 90)
  AND coalesce(direct_access, true)
GROUP BY ALL
ORDER BY last_day DESC

A row with a downstream is a table, a view or a path fed by your table. A row without one is a read only. The condition on direct_access keeps direct readers; remove it to also see those that go through a view. For jobs, notebooks and dashboards, lineage only gives IDs. Table owners are in system.information_schema.tables (column table_owner), which only returns the tables you can see.

Why a recursive walk needs care

A change travels from table to table down to the dashboard. WITH RECURSIVE (Databricks SQL, Databricks Runtime 17.0 and above) follows that chain, with three traps:

  1. Cycles. Two tables can feed each other. The documentation's example keeps the path walked in an array and stops when the next step is already in it.
  2. Depth. MAX RECURSION LEVEL is 100 by default, and going past it raises an error instead of cutting the result. Bound the level yourself.
  3. Several paths. A table reached by three paths comes back three times, and the result is limited to 1,000,000 rows. Start from distinct pairs, then keep the nearest level of each table.

What lineage does not see

According to Databricks, these tables are a subset of all read and write events: a record is only emitted when lineage can be inferred. Among the documented limits:

An empty downstream therefore does not prove that nothing depends on the table.

What breaks if I change or drop this table?

  1. Downstream tables and views, direct first, then along the whole chain.
  2. Readers: dashboards, queries, notebooks, jobs. A drop or a rename breaks direct readers at once.
  3. Columns, for a rename or a type change, in column_lineage.
  4. Owners, to notify beforehand.
  5. What lineage does not see. Before a drop, keep a copy and a waiting period.

What YALineo does

YALineo reads these system tables in your browser, with your Databricks permissions, read only, with no YALineo server.

The map turns each data domain into a colored line and each schema into a station (tutorial). The focus isolates a station (tutorial), and a followed table shows its upstream and downstream, nearest first (tutorial).

The dependencies of an asset first give the direct neighbors, counted like the Lineage tab, then the whole chain (tutorial). A link seen only through a view is not direct. A table the inventory no longer lists (dropped, or unreadable with your rights) leaves the map, and the links that went through it are marked "bridged". The link window is 365 days by default (the rule).

Impact analysis starts from a table, a view or a job, and a scenario: change, drop or rename, late refresh, job stopped. It reads the whole downstream, without the map's filters, and counts each asset once, at its shortest distance. Severity follows a published points rule, for example 3 points for a dashboard, 2 for the gold layer, 2 for 100 reads or more, and 2 more for a direct reader of a dropped table. An asset is critical from 5 points, high at 3 or 4, medium at 1 or 2, low at 0 (every reason).

The analysis also gives the monthly refresh of the touched tables: their refresh cost over the FinOps period, brought to an average month of 365/12 days, each table counted once. That amount tells what these tables take each month; weigh it against an outage. Then come the owners, each with a message ready to copy, and the "To do, in order" list (tutorial).

For a column, YALineo reads column_lineage on demand. Each asset gets a single answer: "Observed relation", "Table dependency to check" (no column link known, as for a dashboard) or "Coverage too thin" (column links of the table reach the asset, none from this one). The details are in the method.

Frequently asked questions

Why do YALineo and Catalog Explorer not count the same downstream?

Direct neighbors are counted the same way. The gaps come from the link window, the catalogs chosen, bridged links and reads through a view, counted for the view (the method).

Does lineage see a Power BI or Tableau report?

Not on its own. A row with no Databricks object can come from a JDBC query, without saying which tool ran it. In YALineo, a table's sheet counts those reads apart.

Do I need access to the system tables?

For the queries of this guide, yes. Without it, Catalog Explorer shows lineage with BROWSE, and YALineo's analyst mode draws the map with the lineage API, without reads or costs (tutorial).

Sources

Documentation read on 6 November 2026.