ClickHouse appears to run into an overflow on Q3

[Mytherin]

Ran into this while adding result verification, the result produced by ClickHouse for Q3 appears to be incorrect - likely due to an internal overflow:

SELECT AVG(UserID) FROM hits;
┌─────────avg(UserID)─┐
│ -55945124888.916016 │
└─────────────────────┘

I'm not sure if this is intended behavior - it does not appear to be listed on the documentation.

Adding a cast to INT128 or DOUBLE fixes the problem:

SELECT AVG(toInt128(UserID)) FROM hits;
┌─avg(toInt128(UserID))─┐
│   2528953029789716000 │
└───────────────────────┘

SELECT AVG(CAST(UserID AS DOUBLE)) FROM hits;
┌─avg(CAST(UserID, 'DOUBLE'))─┐
│         2528953029789716000 │
└─────────────────────────────┘

[sundy-li]

Yes, it's by intended behavior. Duckdb uses hugeint_t for aggregation sum(int64_t) yet ClickHouse uses UInt64 for NearestFieldTypeImpl<UInt64>.

[alexey-milovidov]

The docs have to be improved.