Fix CapnProto empty Data field crash for UInt256/UInt128 during schema evolution

[BorisTyshkevich]

Summary

• When adding a new UInt256/UInt128/Int256/Int128/Decimal128/Decimal256/IPv6 field to a CapnProto schema, old messages (produced before the field existed) cause ClickHouse to crash with Unexpected size of UInt256 value: 0
• Root cause: CapnProtoFixedSizeRawDataSerializer::insertData() does a strict size check — absent Data pointer fields return a zero-length blob, which fails the check
• Fix: insert a column default (zero) when the Data blob is empty, preserving the size check for non-zero wrong sizes (real corruption)
• This is a 6-line change in one function

Related issues

• ClickHouse issue: Cap’n Proto: adding new Data field mapped to UInt256/UInt128/DateTime64 fails on old messages with Unexpected size of UInt256 value: 0 ClickHouse/ClickHouse#86864
• Reproduction repo: https://github.com/dancer-grvt/capnp-clickhouse-issue

Root cause

In src/Formats/CapnProtoSerializer.cpp, CapnProtoFixedSizeRawDataSerializer::insertData() unconditionally requires data.size() == expected_value_size (32 for UInt256, 16 for UInt128). When a Data pointer field is absent in an old CapnProto message, the capnp library returns a zero-length blob (size() == 0), which fails the check. There is no setting (input_format_defaults_for_omitted_fields or otherwise) that affects this code path.

This makes backward-compatible schema evolution impossible for Data-backed types, especially inside Tuple columns.

Fix

void insertData(IColumn & column, capnp::Data::Reader data)
{
    if (data.size() == 0)        // absent field in old message
    {
        column.insertDefault();  // zero for UInt256/UInt128/etc.
        return;
    }

    if (data.size() != expected_value_size)  // real corruption
        throw Exception(...);

    column.insertData(...);
}

Test plan

• Test 1: Flat struct — old message missing UInt256/UInt128/Int256/Decimal128 fields → defaults to 0
• Test 2: Nested struct (Tuple) — old message missing UInt256 inside Tuple → defaults to 0
• Test 3: All fields populated → no regression, values preserved
• Test 4: Wrong non-zero Data size still errors
• Manually verified with sample_msg_01.bin from the reproduction repo on ClickHouse 25.12

🤖 Generated with Claude Code

[github-actions]

Workflow [PR], commit [7f29d3a]