kura_driver implementation over minato.

Where a query goes

Outside a transaction it goes to the pool: minato:query/4 borrows a connection, runs the statement and gives it back, all inside the call. Nothing is held and nothing is threaded through kura.

Inside a transaction it goes to the connection the transaction is running on, which kura_pool_minato holds and this module finds in the process dictionary. The reason is that a query written inside transaction/4's function is handed nothing, so the connection has to be found rather than passed.

Results and errors

minato answers #{command := atom(), num_rows := integer(), rows := [row()]}, which is the shape kura already reads, and a failing statement answers {error, {pgsql_error, Fields}} with code and constraint as binaries - which is what kura_repo_worker turns into a changeset error. Nothing in kura had to change for either.

numeric

numeric decodes to a float here, which is not minato's default: minato hands back the exact decimal text, because that is what the type is for. kura has always given callers a float for it, through pg_types, and a driver swap is the wrong moment to change what avg/1 returns for everybody.

A repo that wants the exact form says numeric_format => binary in its driver options, and should, for anything that is money.

Timeouts

timeout in the driver options is a deadline on the statement rather than on the read. minato cancels a statement that outlives it on the server and reads the cancellation through, so a slow query comes back as SQLSTATE 57014 and the connection goes back to the pool rather than being closed.