Feat: allow overriding the dialects' normalization strategies (#2779)

Author: georgesittas

docs/guides/configuration.md

@@ -893,6 +893,27 @@ Example configuration:
 
 The default model kind is `VIEW` unless overridden with the `kind` key. For more information on model kinds, refer to [model concepts page](../concepts/models/model_kinds.md).
 
+##### Identifier resolution
+
+When a SQL engine receives a query such as `SELECT id FROM "some_table"`, it eventually needs to understand what database objects the identifiers `id` and `"some_table"` correspond to. This process is usually referred to as identifier (or name) resolution.
+
+Different SQL dialects implement different rules when resolving identifiers in queries. For example, certain identifiers may be treated as case-sensitive (e.g. if they're quoted), and a case-insensitive identifier is usually either lowercased or uppercased, before the engine actually looks up what object it corresponds to.
+
+SQLMesh analyzes model queries so that it can extract useful information from them, such as computing Column-Level Lineage. To facilitate this analysis, it _normalizes_ and _quotes_ all identifiers in those queries, [respecting each dialect's resolution rules](https://sqlglot.com/sqlglot/dialects/dialect.html#Dialect.normalize_identifier).
+
+The "normalization strategy", i.e. whether case-insensitive identifiers are lowercased or uppercased, is configurable per dialect. For example, to treat all identifiers as case-sensitive in a BigQuery project, one can do:
+
+=== "YAML"
+
+    ```yaml linenums="1"
+    model_defaults:
+      dialect: "bigquery,normalization_strategy=case_sensitive"
+    ```
+
+This may be useful in cases where the name casing needs to be preserved, since then SQLMesh won't be able to normalize them.
+
+See [here](https://sqlglot.com/sqlglot/dialects/dialect.html#NormalizationStrategy) to learn more about the supported normalization strategies.
+
 #### Model Kinds
 
 Model kinds are required in each model file's `MODEL` DDL statement. They may optionally be used to specify a default kind in the model defaults configuration key.

sqlmesh/core/context.py

@@ -49,7 +49,7 @@
 from types import MappingProxyType
 
 import pandas as pd
-from sqlglot import exp
+from sqlglot import Dialect, exp
 from sqlglot.lineage import GraphHTML
 
 from sqlmesh.core import analytics
@@ -325,6 +325,13 @@ def __init__(
 
         self.path, self.config = t.cast(t.Tuple[Path, C], next(iter(self.configs.items())))
 
+        # This allows overriding the default dialect's normalization strategy, so for example
+        # one can do `dialect="duckdb,normalization_strategy=lowercase"` and this will be
+        # applied to the DuckDB dialect globally
+        if "normalization_strategy" in str(self.config.dialect):
+            dialect = Dialect.get_or_raise(self.config.dialect)
+            type(dialect).NORMALIZATION_STRATEGY = dialect.normalization_strategy
+
         if self.config.disable_anonymized_analytics:
             analytics.disable_analytics()
 

tests/core/test_context.py

@@ -735,3 +735,20 @@ def test_disabled_model(copy_to_temp_path):
 
     assert (path[0] / "models" / "disabled.sql").exists()
     assert not context.get_model("sushi.disabled")
+
+
+def test_override_dialect_normalization_strategy():
+    config = Config(
+        model_defaults=ModelDefaultsConfig(dialect="duckdb,normalization_strategy=lowercase")
+    )
+
+    # This has the side-effect of mutating DuckDB globally to override its normalization strategy
+    Context(config=config)
+
+    from sqlglot.dialects import DuckDB
+    from sqlglot.dialects.dialect import NormalizationStrategy
+
+    assert DuckDB.NORMALIZATION_STRATEGY == NormalizationStrategy.LOWERCASE
+
+    # The above change is applied globally so we revert it to avoid breaking other tests
+    DuckDB.NORMALIZATION_STRATEGY = NormalizationStrategy.CASE_INSENSITIVE