Antalya 26.3: Fix condition for using parquet metadata cache

src/Storages/ObjectStorage/StorageObjectStorageSource.cpp

@@ -884,7 +884,7 @@ StorageObjectStorageSource::ReaderHolder StorageObjectStorageSource::createReade
 
         InputFormatPtr input_format;
         if (context_->getSettingsRef()[Setting::use_parquet_metadata_cache] && use_native_reader_v3
-            && (object_info->getFileFormat().value_or(configuration->format) == "Parquet")
+            && (Poco::toLower(object_info->getFileFormat().value_or(configuration->format)) == "parquet")
             && !object_info->getObjectMetadata()->etag.empty())
         {
             const std::optional<RelativePathWithMetadata> object_with_metadata = object_info->relative_path_with_metadata;

tests/integration/test_storage_iceberg_with_spark/test_read_constant_columns_optimization.py

@@ -169,6 +169,18 @@ def execute_spark_query(query: str):
     for replica in started_cluster_iceberg_with_spark.instances.values():
         replica.query("SYSTEM FLUSH LOGS")
 
+    # Number of object-get requests per data file that are NOT served from caches
+    # after the warmup query above. The parquet metadata cache (enabled by default)
+    # caches the parquet footer keyed by the object's etag; the warmup query then
+    # populates it, so any subsequent read of the same file skips one object-get
+    # (the footer read). However, AzureObjectStorage::getObjectMetadata does NOT
+    # populate etag, so the cache guard `!etag.empty()` in
+    # StorageObjectStorageSource::createReader always fails for Azure, and the
+    # cache path is never taken there. As a result the multiplier is:
+    #   S3:    2 (footer served from cache, data-only gets remain)
+    #   Azure: 3 (cache never engaged, footer + data gets)
+    per_file_gets = 2 if storage_type == "s3" else 3
+
     def check_events(query_id, event, is_cluster, expected):
         res = instance.query(
             f"""
@@ -185,9 +197,10 @@ def check_events(query_id, event, is_cluster, expected):
             """)
         # Weird, but looks like ReadFileMetadata does not used local file cache in 26.1
         # metadata.json always downloaded in 26.1, once per query or subquery
-        # In 25.8 count was equal to expected, in 26.1 it is expected * 2 + 1 for Local case
-        # expected * 2 + 4 for Cluster case, because each subquery loads metadata.json
-        assert int(res) == expected * 2 + (4 if is_cluster else 1)
+        # In 25.8 count was equal to expected, in 26.1 it is expected * N + 1 for Local
+        # case and expected * N + 4 for Cluster case (each subquery loads metadata.json).
+        # N = per_file_gets (see comment above the function).
+        assert int(res) == expected * per_file_gets + (4 if is_cluster else 1)
 
     event = "S3GetObject" if storage_type == "s3" else "AzureGetObject"
 
@@ -212,4 +225,4 @@ def compare_selects(query):
 
     compare_selects(f"SELECT _path,* FROM {creation_expression} ORDER BY ALL")
     compare_selects(f"SELECT _path,* FROM {creation_expression} WHERE name_old='vasily' ORDER BY ALL")
-    compare_selects(f"SELECT _path,* FROM {creation_expression} WHERE ((tag + length(name_old)) % 2 = 1) ORDER BY ALL")
+    compare_selects(f"SELECT _path,* FROM {creation_expression} WHERE ((tag + length(name_old)) % 2 = 1) ORDER BY ALL")