Snowflake-Labs/dbt_constraints 1.0.10

dbt Constraints v1.0.10

Repository: Snowflake-Labs/dbt_constraints

Tag: 1.0.10

Published: 2026-09-30T23:38:15Z

Prerelease: no

Release notes: A bug fix release. Two fixes change what the package does at runtime, so read the upgrade notes below.

No constraints were created at all on 1.0.8

Two clauses in create_constraints_by_type read meta.materialized with a literal default of "other". Almost no node sets that property, so both clauses read "other", which is in neither tuple they test, and both were therefore always true.

One of them set verify_permissions for every model, so the ownership pre-check ran before every constraint. That check does not reliably report OWNERSHIP for a table that the same run just replaced, so the package skipped the constraint. If you upgraded to 1.0.8 and your constraints stopped appearing, this is the cause.

The other clause excludes views, ephemeral models and dynamic tables. It sits in an or inside an inclusive filter, so an always-true operand made the whole exclusion dead and those nodes passed.

meta.materialized is an override, not an independent property. A custom materialization uses it to declare the kind of object it builds. The package now reads it first and falls back to config.materialized, through a new effective_materialization macro that both call sites share. Fixes #131.

Tests that set the where config produced invalid SQL

dbt replaces the model with (select * from x where ...) dbt_subquery when a test sets the where config, and that text already carries an alias. The test SQL read from {{model}} pk_test, so the query held two aliases and the database rejected it with a syntax error on the package's own alias. Primary key, unique key and foreign key tests were all affected.

The test SQL now selects from the model through a subselect, so the test supplies the only alias at its own level. The parent relation in the foreign key test is unchanged, because the where config never rewrites it. Fixes #100, a regression from 1.0.6, and #119.

Constraints on sources

dbt_constraints_sources_enabled never worked. The package resolved test dependencies against graph.nodes, but dbt keeps sources in graph.sources, so every source dependency resolved to nothing and no DDL was issued. Dependencies now resolve from graph.nodes and fall back to graph.sources. A test on a source has no attached_node, so the package accepts a test that depends on exactly one node and reads the table name from identifier.

If you already set `dbt_constraints_sources_enabled: true`, the package starts creating constraints on your source tables after you upgrade. Confirm the dbt role should alter those tables. Set the variable to false to turn source constraints off.

Deadlocks on databases that enforce constraints

Postgres and Oracle deadlocked under a parallel run. Dropping a constraint with CASCADE also locks every table whose foreign key depends on it, and drop_relation did that on each worker thread, so two threads could take the same locks in opposite order. The previous mitigation was itself one of the deadlock participants. Ordering cannot fix it, because a foreign key does not imply a ref(), so dbt has no ordering guarantee between the two models a foreign key joins.

The fix is timing. A new on-run-start hook calls release_constraints_for_rebuild(), which drops the constraints on the relations the run will drop or replace, from a single session, before the worker threads start. Same DDL against the same constraints. Seeds are always included, because a seed without a full refresh is TRUNCATEd, which needs the constraints gone just as much as a drop does.

This adds an `on-run-start` hook to the package. Adapters opt in by implementing release_constraints_for_rebuild; the default does nothing, so Snowflake, Redshift, BigQuery and Vertica are unaffected. Set dbt_constraints_release_before_build: false to drop constraints during the build as before.

Verified at 64 threads: postgres and oracle both clean, no ORA-00060.

Faster constraint generation on large projects

test_selected ran once per test, and each call looped over every node in the graph to find the unique and primary keys behind a selected model's foreign keys. That is O(n squared) and it was slow on large projects. The foreign key model list is now built once per create_constraints_by_type call and reused, so test_selected does a dict lookup. Contributed by @RichardGoerlerKS in #128.

Other bug fixes

• #85 Foreign key creation raised UndefinedError when the parent table was missing.

The guard's else-branch read fk_table_relation.is_table on a None relation and pk_model.is_table on a manifest node that has no such attribute.

• #115 Constraints were silently skipped on versioned models. The identifier was built

from node.alias or node.name, giving my_model where the database has my_model_v1, so adapter.get_relation() returned nothing. The package now prefers node.relation_name when node.version is set.

• #117 lookup_table_privileges had no dispatch wrapper, which broke projects that

call dbt_constraints.create_constraints() from their own on-run-end hook, a pattern from before 0.6.3.

• #107 Generated constraint names only stripped double quotes, so a column containing

parentheses, spaces, dots or dashes produced invalid DDL such as ADD CONSTRAINT TBL_OBJECTID(SMTH)_PK. A new sanitize_constraint_name() helper uppercases, strips quotes, and replaces every character outside [A-Z0-9_$] with an underscore. It runs before the postgres and oracle length truncation, so the truncated identifier is always valid.

• #110 Restored the RELY/NORELY flip on constraints that already exist on Snowflake.
• #120 Corrected macro argument types in macros.yml for dbt Fusion compatibility.
• #130 dbt 1.12 could not parse the package. macros.yml used the argument type

boolean, which is not in the accepted grammar, and five macros documented fewer arguments than they declare and in a different order. dbt 1.12 validates all of that, and a project with warn_error_options set to error: all failed at parse time. Types are now bool, and verify_permissions, lookup_cache and rely_clause are documented in declaration order.

Documentation

The README listed the supported databases in two places and both were wrong. BigQuery was missing from the...