Database interface design - was how to build up the library.

[Andrei Alexandrescu]

On 10/7/11 9:11 AM, Steve Teale wrote:
> I'm thinking that for each database that gets covered there will
> need to be two modules, like:
>
> etc.c.mysql
> etc.mysqld
>
> The etc.c.xxx modules would be completely different between databases -
> just translations of the necessary header files. The xxxd modules
> would be as similar as possible consistent with allowing the strong
> points of individual database systems to show through, and the weak
> points of others to be avoided. I don't think all should be reduced
> to some LCD.

Well we may be hasty to go that way. The driver-based approach works 
well for other languages and APIs - why wouldn't it work for us? The 
differences across different DBMSs would be mostly in the SQL support, 
not the basic APIs.

> These modules should attempt to make a good range of capabilities
> available to the D programmer, but they not have to be all encompassing.
> Those users who want to do really fancy things can drop back to the
> low-level interface. They should probably have the following capabilities:
>
> 1) Establishing and closing database connections.

Sounds good. Since DB connections are a precious resource, there must be 
a RAII struct holding them. The functions below may be members of it.

> 2) The capability to execute literal SQL statements - execSQL()
> if you like. Some of these will generate result sets, of which more below.

Great.

> 3) The capability to create prepared statements with in and out
> parameters and association of the parameters with a source, and
> then to execute these. This breaks down into several components/
> capabilities, which could be labeled:
>
> 3a) createPreparedStatement() - marshal parameters, associate them
> with a sourceand have the server prepare the statement.
>
> 3b) execStatement() - for those SQL statements that don't have a
> result set.
>
> 3c) execStatementIncremental()/getNext() - execute the prepared statement,
> then fetch the result rows one at a time into some kind of result set.

Here's where the range interface might be used. We might simply have 
execStatement() that returns an input range. If the statement produced 
no rows, the range will be empty. No need for distinguishing 
execStatement() and execStatementIncremental().

> 3d) execStatementAll() - execute the prepared statement and get all
> the resulting rows into some kind of result set.

This is not a primitive, but instead a convenience function that should 
be used with caution and only for queries known to be small. I agree 
that it's good to have.

> 3e) (maybe) execScalar() - do the whole sequence prepare, execute,
> and get a single value result set placed into a D variable.

That's convenient for e.g. "SELECT COUNT(*) FROM table WHERE condition", 
which people run all the time.

> 3f) (maybe) execStoredProcedure() - another 'do the whole thing'
> capability TBD.

Well wouldn't that just execute a special SQL a la "CALL procedure"?

> It is when we come to the nature of the result sets that there is
> likely to be dissent. I favor arrays of structs, but we should
> probably do arrays of arrays of variants too for those situations
> where structures can't be sorted out at compile time. There needs
> to be some symmetry between what is used here, and what can be used
> as input to operations such as a prepared insert. It is of course
> vital that this part of each middle layer produce exactly the same
> kind of results. Otherwise the top layer could become difficult.

I'd like arrays of structs too but we must acknowledge that most 
databases people use will be large, in which case loading the whole 
thing eagerly takes a lot of RAM and potentially wastes time. So arrays 
are out except for e.g. convenience functions - but then we already have 
array() that converts an arbitrary input range into an array.

Now, regarding the "structs" part, I'd like that too when the schema is 
statically known. Two issues:

1. MFC had at a point a wizard that generated one struct per resultset. 
It was an absolute maintenance disaster and they recanted by offering 
dynamically-bound result sets. The lesson there is that defining a 
struct for each query won't likely play out, so we better use 
Tuple!(ColType1, ColType2, ...). A possible API would be:

auto db = std.database.connect("cdb.mysql");
auto rows = db.sql!(double, ulong)
     ("SELECT SUM(x), COUNT(x) FROM table");
// We know that rows is a range of Tuple!(double, ulong).
writeln("Sum is ", rows.front[0], " count is ", rows.front[1]);

Cool beans. I'd love to use such an API!

2. Statically-bound tuples work only when we know the query beforehand. 
We need to cater for resultsets of unknown column count and types. The 
question here is whether we traffic in untyped memory a la ubyte[] or 
some variant type. I favor Variant because it's efficient, accommodates 
any SQL type easily, and is convenient to use.

> On top of this set of two modules for each database, I envisage a
> higher-level module - etc.dbops - that provides a bunch of convenience
> templates for various common database operations, spanning the databases.
> Once the middle layer is in place, this top layer should be relatively
> easy to implement. It should be noted though that all these database
> wrappers will be monstrously difficult to test.
>
> I am at the point with MySQL where I can get the result of a plain
> old query into an array of a checked structure type. I have the
> prepared statement stuff, and know how the result will be created from
> a prepared query (the execStatementAll() case) - I just have to plow
> through a bit more binding and fetching.
>
> This is probably rather general and vague, but I would like to get
> comments so we can iterate toward a practical design.

Sounds good. What I'd suggest given the plethora of DB bindings already 
available is to form a sort of a "task force" and develop a shared 
vision towards a solid proposal. You may self-organize with community's 
help on this group and exchange ideas under e.g. the prefix 
"[std.database]" in the message titles.

What I can say for sure is that there will be an std.database.


Thanks for this initiative.

Andrei