Fix row policies silently ignored on Iceberg tables with PREWHERE enabled

[mkmkme]

The Iceberg read optimization (allow_experimental_iceberg_read_optimization) identifies constant columns from Iceberg metadata and removes them from the read request. When all requested columns become constant, it sets need_only_count = true, which tells the Parquet reader to skip all initialization — including preparePrewhere — and just return the raw row count from file metadata.

This completely bypasses row_level_filter (row policies) and prewhere_info, returning unfiltered row counts. The InterpreterSelectQuery relies on the storage to apply these filters when supportsPrewhere is true and does not add a fallback FilterStep to the query plan, so the filter is silently lost.

The fix prevents need_only_count from being set when an active row_level_filter or prewhere_info exists in the format filter info.

Fixes #1595

Changelog category (leave one):

• Bug Fix (user-visible misbehavior in an official stable release)

Changelog entry (a user-readable short description of the changes that goes to CHANGELOG.md):

Fix row policies silently ignored on Iceberg tables with PREWHERE enabled

Documentation entry for user-facing changes

...

CI/CD Options

Exclude tests:

• Fast test
• Integration Tests
• Stateless tests
• Stateful tests
• Performance tests
• All with ASAN
• All with TSAN
• All with MSAN
• All with UBSAN
• All with Coverage
• All with Aarch64
• All Regression
• Disable CI Cache

Regression jobs to run:

• Fast suites (mostly <1h)
• Aggregate Functions (2h)
• Alter (1.5h)
• Benchmark (30m)
• ClickHouse Keeper (1h)
• Iceberg (2h)
• LDAP (1h)
• Parquet (1.5h)
• RBAC (1.5h)
• SSL Server (1h)
• S3 (2h)
• S3 Export (2h)
• Swarms (30m)
• Tiered Storage (2h)

[mkmkme]

AI audit note: This review comment was generated by AI (claude-4.6-opus-high-thinking).

Audit update for PR #1597 (Fix row policies silently ignored on Iceberg tables with PREWHERE enabled):

Confirmed defects

No confirmed defects in reviewed scope.

The fix is correct and minimal. The analysis follows.

Bug being fixed (pre-existing): When allow_experimental_iceberg_read_optimization is enabled and Iceberg metadata identifies all requested columns as constant, the code unconditionally set need_only_count = true. In the ParquetV3 reader, this causes read() (at ParquetV3BlockInputFormat.cpp:105) to skip initializeIfNeeded() entirely — reading only file_metadata.num_rows from Parquet metadata and returning raw row counts. This completely bypasses preparePrewhere() (called via ReadManager), which is where row_level_filter (row policies) and prewhere_info are applied. Since InterpreterSelectQuery does not add a fallback FilterStep for these when supportsPrewhere() is true, the filters are silently lost.

Fix correctness: The added guard (!format_filter_info || (!format_filter_info->row_level_filter && !format_filter_info->prewhere_info)) prevents the need_only_count override when either filter exists. When need_only_count stays false:

• The Parquet reader initializes normally, reads filter columns via format_header, and applies prewhere/row-level filters during reading
• ExtractColumnsTransform then strips filter-only columns (producing correct row counts)
• generate() restores constant columns from constant_columns_with_values with the filtered row count

Null safety: The fix correctly short-circuits on !format_filter_info before dereferencing members.

Consistency with caller: The caller in StorageObjectStorage.cpp:448-452 already guards need_only_count with !read_from_format_info.prewhere_info && !read_from_format_info.row_level_filter. Both are populated from query_info, matching the format_filter_info fields checked by the fix.

filter_actions_dag not checked (correct by design): The fix intentionally does not guard against filter_actions_dag. This is correct because WHERE-clause pushdown via filter_actions_dag has a plan-level FilterStep fallback. Even if the format reader skips it, the plan re-evaluates the WHERE on the constant columns restored by generate(), producing correct results.

Coverage summary

• Scope reviewed: Single-file change in StorageObjectStorageSource::createReader (static overload, lines 573–964); callers in StorageObjectStorage::read and ReadFromObjectStorageStep::initializePipeline; downstream consumers in ParquetV3BlockInputFormat::read, ParquetBlockInputFormat, ReadManager::preparePrewhere; FormatFilterInfo struct definition.
• Categories passed: Null-pointer safety, guard condition logic, consistency with caller-level guard, constant-column restoration correctness, cache-path (num_rows_from_cache) correctness, thread-safety of shared_ptr parameter, C++ lifetime/RAII (no new objects or ownership changes).
• Categories failed: None.
• Assumptions/limits: No regression test is included in the PR; correctness of the fix relies on the plan-level FilterStep fallback for filter_actions_dag (verified by code inspection, not runtime).

[mkmkme]

I've also checked the regression test is passing with this change.

[ianton-ru]

LGTM