`supportsDelete` in `DataLakeConfiguration` throws `LOGICAL_ERROR` on freshly-attached data lake table with uninitialized metadata

[clickgapai]

Found via ClickGap automated review. Please close or comment if this is incorrect or needs adjustment.

Describe what's wrong

DELETE FROM table WHERE ... on a freshly-attached Iceberg/DeltaLake/Hudi table with corrupted or unloadable metadata throws LOGICAL_ERROR: Metadata is not initialized instead of a user-facing exception

Root cause: DataLakeConfiguration.h:133-137: supportsDelete() calls assertInitialized() instead of lazyInitializeIfNeeded(). Unlike the four methods fixed by this PR (optimize, alter, checkMutationIsPossible, checkAlterIsPossible), supportsDelete() was not updated to use lazy initialization.

Why we believe this is a bug: InterpreterDeleteQuery::execute() (line 92) calls table->supportsDelete() → StorageObjectStorage::supportsDelete() → DataLakeConfiguration::supportsDelete() (line 133-137) which calls assertInitialized(). The PR fixes checkMutationIsPossible but not supportsDelete() which is called BEFORE checkMutationIsPossible in the DELETE FROM path.

Affected locations:

• src/Storages/ObjectStorage/DataLakes/DataLakeConfiguration.h:133 — supportsDelete() calls assertInitialized()

Impact: DELETE FROM queries on freshly-attached data lake tables with uninitialized metadata throw LOGICAL_ERROR, which crashes debug/sanitizer builds and produces confusing error messages in release builds.

Does it reproduce on most recent release?

Yes — confirmed on current master (commit 1732358b6ce).

How to reproduce

Requires IcebergLocal support and user_files_path access (cannot run in Fiddle).

-- Setup
CREATE TABLE test_datalake (c0 Int32) 
ENGINE = IcebergLocal('/var/lib/clickhouse/user_files/test_datalake/')
SETTINGS iceberg_metadata_compression_method = 'deflate';

-- Corrupt metadata by using incompatible compression level
INSERT INTO test_datalake VALUES (2)
SETTINGS allow_insert_into_iceberg=1, output_format_compression_level=11;
-- This INSERT fails with INCORRECT_DATA (expected)

-- DETACH/ATTACH to get uninitialized metadata state
DETACH TABLE test_datalake SYNC;
ATTACH TABLE test_datalake;

-- BUG: This throws LOGICAL_ERROR instead of user-facing exception
DELETE FROM test_datalake WHERE c0 = 1 
SETTINGS allow_insert_into_iceberg=1;

-- Cleanup
DROP TABLE IF EXISTS test_datalake SYNC;

Expected behavior

INSERT failed as expected
DELETE FROM did not crash with LOGICAL_ERROR
1

Error message and/or stacktrace

INSERT failed as expected
BUG: DELETE FROM crashed with LOGICAL_ERROR
1

Additional context

Open risks:

• supportsParallelInsert() at line 139-143 has the same issue but the code path to reach it without initialization may be protected by earlier update() calls
• mutate() at line 145-154 has assertInitialized() but is called after checkMutationIsPossible which now initializes metadata

Suggested fix: Change supportsDelete() to take ObjectStoragePtr and ContextPtr parameters and call lazyInitializeIfNeeded() instead of assertInitialized(), matching the pattern used for the other fixed methods. Also need to update StorageObjectStorage::supportsDelete() to pass the required parameters.

Analysis details: Confidence HIGH | Severity P1 | Testability: STATELESS_SQL

Found during automated review of PR #104738.

---

ClickGapAI · Confidence: HIGH · Severity: P1 · Finding: h_pr104738_001

[PedroTadim]

@clickgapai the Fiddle doesn't have queries. Can you give a reproducible?

[clickgapai]

Reproducer (requires IcebergLocal + user files path):

-- Setup
CREATE TABLE test_datalake (c0 Int32) 
ENGINE = IcebergLocal('/var/lib/clickhouse/user_files/test_iceberg/')
SETTINGS iceberg_metadata_compression_method = 'deflate';

-- Trigger the bug path: detach + attach leaves metadata uninitialized
DETACH TABLE test_datalake SYNC;
ATTACH TABLE test_datalake;

-- DELETE calls supportsDelete() which calls assertInitialized() instead of lazyInitializeIfNeeded()
SET allow_insert_into_iceberg = 1;
DELETE FROM test_datalake WHERE c0 = 1;
-- Expected: normal error or success
-- Actual: Code: 49. DB::Exception: Logical error: 'Metadata is not initialized'

The issue: DataLakeConfiguration.h:133 supportsDelete() calls assertInitialized() while the four methods fixed in PR #104738 (optimize, alter, checkMutationIsPossible, checkAlterIsPossible) now call lazyInitializeIfNeeded(). This asymmetry causes DELETE FROM to fail on freshly-attached tables.

[zlareb1]

@clickgapai for tests which can't run on local you shouldn't give fiddle link in issue description.

[clickgapai]

You're right — removed the Fiddle link from the issue description. IcebergLocal requires local file paths which Fiddle doesn't support. Will avoid this for data lake / local-path tests going forward.

[zlareb1]

@clickgapai also write proper reproducer in "How to reproduce" section of the issue. Just writing text isn't the correct way.

[clickgapai]

Updated — the "How to reproduce" section now has executable SQL commands instead of the text description.

[fm4v]

@groeneai fix the issue, it's seems to be broken after #104738

[groeneai]

@fm4v thanks for the ping. The supportsDelete gap is exactly what PR #104917 fixes, and it has been approved by @alexey-milovidov since 2026-05-17 (awaiting merge).

PR: #104917 (Closes #104891)

What it does: wraps DataLakeConfiguration::supportsDelete() and supportsParallelInsert() at the StorageObjectStorage layer to call lazyInitializeIfNeeded(object_storage, CurrentThread::tryGetQueryContext()) before delegating. This matches the existing pattern of supportedPrewhereColumns() and getColumnSizes() in the same file. The earlier PR #104738 fixed the same shape for optimize / alter / checkMutationIsPossible / checkAlterIsPossible but missed supportsDelete / supportsParallelInsert (which run before checkMutationIsPossible in InterpreterDeleteQuery::execute), which is why your reproducer still hits the LOGICAL_ERROR.

Tests:

• 04238_iceberg_supports_delete_metadata_not_initialized_104891.sh covers 6 direct-query shapes after DETACH/ATTACH with corrupted metadata (DELETE, TRUNCATE, UPDATE, INSERT, ALTER DELETE, ALTER UPDATE).
• 04240_iceberg_supports_parallel_insert_mv_104891.sh covers the MV-dependent-sink path via InsertDependenciesBuilder::supportsParallelInsert.

Verified FAIL->PASS locally on debug build: without the fix the server SIGABRTs on DELETE FROM with Logical error: 'Metadata is not initialized' (matches your reproducer); with the fix the server stays alive and returns INCORRECT_DATA 117 on the corrupted metadata.

The only currently-blocking CI signal is the CH Inc sync test_e2e_catalogs/...[onelake] failures, which are an unrelated OneLake backend issue (clickhouse-private#51399) and do not touch this PR's code paths.