DuckDB.NET.Data 1.5.6.1

• net10.0

  • Apache.Arrow (>= 23.0.0)
  • DuckDB.NET.Bindings (>= 1.5.6.1)

• net8.0

  • Apache.Arrow (>= 23.0.0)
  • DuckDB.NET.Bindings (>= 1.5.6.1)

- Built on DuckDB v1.5.6

New features:
- Read a TIME column as TimeSpan
- Read FLOAT as double, DOUBLE as float, and both as decimal, in columns, list items, maps and structs
- Read HUGEINT and UHUGEINT columns as DuckDBHugeInt and DuckDBUHugeInt with GetFieldValue and GetProviderSpecificValue
- The mapped appender now accepts DuckDBDateOnly and DuckDBTimeOnly properties, and scalar functions accept them as parameters and results
- BigInteger values can now be bound and appended to UHUGEINT columns as unsigned values, including through the mapped appender
- A table function over an array now uses the length of the array as its cardinality hint when none is given

Performance (measured against 1.5.6 on .NET 10 with a million rows; the gains are in moving values between DuckDB and .NET, so bulk appends, reading large results and queries that call C# functions benefit most, while queries dominated by DuckDB's own work change little):
- Reading numbers and booleans through GetInt32, GetFieldValue<T> and the other typed getters 3.5 to 6 times faster, an integer column read as double 12 times faster
- Reading short strings up to 1.35 times faster
- Reading dates, times and timestamps 3 to 12 times faster, appending a DateTime 14 to 26 times faster
- UUID, DECIMAL and HUGEINT reads 1.7 to 7 times faster
- Typed scalar functions 5 to 8 times faster, 1.6 times with a string argument
- Table functions 2.5 to 6 times faster
- With projection pushdown, the columns a query does not read are no longer computed: a query that reads one column of a four-column table function is about 6 times faster, and the gain grows with the number of unread columns
- LIST, STRUCT and MAP columns read into .NET types up to 6 times faster
- Appending BigInteger, lists, strings, C# enums and through the mapped appender 1.3 to 4 times faster; numbers 1.2 to 1.4 times faster; decimals of up to 18 digits 1.3 to 2 times faster depending on the CPU, wider decimals more than twice as fast

Memory (per million rows, against 1.5.6):
- Appending no longer allocates for each value. 1.5.6 allocated about 750 bytes for every DateTime (2.9 GB for four TIMESTAMP columns), 756 MB for a record of mixed types in the mapped appender, 211 MB for lists of int?, 188 MB for C# enums and up to 315 MB for BigInteger; each is now about 0.1 MB
- Scalar functions with nullable arguments or results, and table functions over structs or rows with nullable columns, no longer allocate for each row: 21 to 47 MB in 1.5.6, now about 0.1 MB
- An integer read as double no longer allocates (375 MB for eight million values in 1.5.6), a DECIMAL wider than 18 digits is read without a BigInteger (107 MB, now none), and a large HUGEINT read as BigInteger allocates a third as much
- LIST, STRUCT and MAP reads allocate up to 60% less, and GetValues 9 to 17% less

Fixes:
- A Read() after the end of a streaming result, and the first Read() of an empty streaming result, no longer throw (#358)
- Reading an ENUM column as a type other than string no longer reads past the stored value
- Reading a BIT column as an unsupported type throws InvalidCastException instead of ending the process with a stack overflow
- Negative intervals can be read as TimeSpan, bound and appended
- A TIMETZ read as DateTimeOffset no longer throws for a time earlier in the day than its positive offset
- -2^127 is accepted as a HUGEINT value, and the HUGEINT limits no longer fail under cultures with another minus sign
- A decimal parameter is sent with the exact number of its digits; 1000 was sent as DECIMAL(3,0)
- A list parameter that DuckDB cannot build is refused instead of crashing the process
- The exception of a scalar function that throws in WHERE or HAVING is now the InnerException of the error
- A table function projection that calls a constructor (a record or a tuple) is refused with a clear message instead of a NullReferenceException
- The mapped appender disposes its appender when validation fails
- DuckDBClientFactory.CanCreateCommandBuilder and CanCreateDataAdapter return false instead of throwing

Behavior changes:
- Appending a decimal too wide for its column throws instead of storing a value the column cannot hold
- A TimeSpan parameter where DuckDB cannot infer the type is bound as an INTERVAL instead of as text; declare DbType.Time to bind it as a time
- Looking up a missing parameter by name through DbParameterCollection throws IndexOutOfRangeException, as the typed indexer does
- A TIMETZ read as DateTimeOffset is placed on 0001-01-02 instead of 0001-01-01
- A BigInteger parameter where DuckDB cannot infer the type is bound as HUGEINT if it fits, as UHUGEINT up to 2^128 - 1 and as BIGNUM beyond, instead of throwing for values outside the HUGEINT range