Permit skipping metadata for MariaDB

[rusher]

Since MariaDB 10.6 (with https://jira.mariadb.org/browse/MDEV-19237) binary result-set can skip sending metadata when they haven't changed.
This avoid network data and parsing client side.

Benchmarks results using DDL:

CREATE TABLE test100 (i1 int,i2 int,...,i100 int);
INSERT INTO test100 value (1,2,...,100);

test methods:

    [Benchmark]
    public async Task<int> readOneRow100FieldsAsync()
    {
        var total = 0;
        using (var cmd = Connection.CreateCommand())
        {
            cmd.CommandText = "select * FROM test100";
            using (var reader = await cmd.ExecuteReaderAsync())
            {
                while (await reader.ReadAsync())
                    for (var i = 0; i < 100; i++)
                        total += reader.GetInt32(i);
            }
        }
        return total;
    }

    [Benchmark]
    public async Task<int> readOneRow100FieldsPrepareAsync()
    {
        var total = 0;
        using (var cmd = Connection.CreateCommand())
        {
            cmd.CommandText = "select * FROM test100";
            cmd.Prepare();
            using (var reader = await cmd.ExecuteReaderAsync())
            {
                while (await reader.ReadAsync())
                    for (var i = 0; i < 100; i++)
                        total += reader.GetInt32(i);
            }
        }
        return total;
    }

before PR:

|                          Method |        Library |      Mean |    Error |
|-------------------------------- |--------------- |----------:|---------:|
|        readOneRow100FieldsAsync | MySqlConnector | 104.47 us | 1.468 us |
| readOneRow100FieldsPrepareAsync | MySqlConnector |  92.49 us | 1.154 us |

After PR:

|                          Method |        Library |      Mean |    Error |
|-------------------------------- |--------------- |----------:|---------:|
|        readOneRow100FieldsAsync | MySqlConnector | 103.99 us | 0.914 us |
| readOneRow100FieldsPrepareAsync | MySqlConnector |  71.50 us | 1.402 us |

[bgrainger]

Thanks! I will take a look when I'm back from vacation.

I assume this may be related to #446 although MariaDB uses a different flag value to enable the capability? (Or maybe it's a completely different on-the-wire way to avoid sending metadata?)

[rusher]

In terms of the protocol, this corresponds to the mysql change you indicate.
However, it is not exactly the same thing:

The MySQL implementation is a very specific optimization for some customers that cache application-side metadata, pushing the metadata to the connector metadata for each request.

For MySQL, there is a CLIENT_OPTIONAL_RESULTSET_METADATA capability that enables this feature, then when @@session.resultset_metadata is "NONE", then all result set metadata will not be sent. Upon receiving the query result, the application provides metadata to the connector to parse the rows.

MariaDB's implementation is more general: when the client sets the MARIADB_CLIENT_CACHE_METADATA capability, the metadata MAY not be sent: running a COM_QUERY will still have the metadata. The modification concerns only the prepared command (COM_STMT_EXECUTE). Once prepared, the connector already knows the metadata, so on the second run, when MARIADB_CLIENT_CACHE_METADATA is set, the server will not return the metadata to the COM_STMT_EXECUTE result set if the metadata has not changed since the previous run. So to sum up, the first run result will have metadata, the second one won't.

[bgrainger]

I believe the MariaDbCacheMetadata value is defined incorrectly and needs to be fixed.

[bgrainger]

Note to self: this struct is starting to accumulate a lot of mutable public fields; redesign?

[bgrainger]

Note to self: Can we move this outside the if statement (and remove the corresponding code from the else branch)?

[rusher]

I think best would be to get rid of the ColumnTypes variable, saving some calculation that are useful for metadata only, so that may not be used at all. => save ConnectionSettings and just rely on ColumnDefinitions

[bgrainger]

Agreed: 5f72a2f.

[bgrainger]

Note to self: Most other Payload objects are classes, but would it be worthwhile making this (and others?) a readonly struct?

[bgrainger]

3e746d8