sqlgg: SQL Guided (code) Generator

Problem

Writing database layer code is usually tedious and error-prone, due to the mix of different languages. SQL queries constructed dynamically need to bind external data (from application), and the resulting rowset must be decomposed into application native data. Data crossing these application-to-database boundaries is what causes troubles. One can factor out all common database communication code, hide the database under some application-specific abstraction, but one always needs to manually specify correspondence between SQL query binding slots (or resulting rowset columns) and code variables. This mapping should be updated manually every time SQL query is modified.

Solution

SQL query parser and code generator which ensures that application code and database queries are in sync. It analyzes SQL query and determines the set of input parameters (values for INSERT, run-time substitution parameters) and the set of resulting columns (for SELECT). Then it generates the code in host language, matching query input and output to function parameters and return values with corresponding native data types. So basically you provide an SQL query and generator creates a function which takes the set of typed parameters as required to fill slots in a query. Generated code binds provided parameters into query and executes it. SELECT statements additionally return the collection of structures with fields representing columns of resulting rowset (or pass those structures to callback-function). The most fruitful consequence of such approach is that the host language compiler will itself check that functions generated from SQL queries will be called correctly (i.e. all parameters bound with correct types). So if you modify the query and forget to update the code -- the compiler will point on erroneous parts.

Complete example

Suppose we have the database of some money relationships for the group of people. We use the following SQL tables and queries:

-- @create_person
CREATE TABLE person (id INTEGER PRIMARY KEY AUTOINCREMENT,name TEXT,surname TEXT);
-- @add_person
INSERT INTO person (name,surname) VALUES;

-- @create_money
CREATE TABLE money (src INTEGER, dst INTEGER, amount INTEGER);
-- @add_money
INSERT INTO money VALUES;

-- @calc_total
SELECT name || ' ' || surname AS fullname, SUM(amount) as total FROM person JOIN money ON src = id GROUP BY id;
-- @list_donors
SELECT DISTINCT surname FROM person JOIN money ON src = id AND dst = (SELECT id FROM person WHERE surname LIKE @surname) LIMIT @limit;

Generate the binding C++ code (boilerplate omitted, only function prototypes shown):

// DO NOT EDIT MANUALLY

// generated by sqlgg 0.2.0 (735e2815)

#pragma once

template <class Traits>
struct sqlgg
{
    // CREATE TABLE person (id INTEGER PRIMARY KEY AUTOINCREMENT,name TEXT,surname TEXT)
    static bool create_person(typename Traits::connection db);

    // INSERT INTO person (name,surname) VALUES
    static bool add_person(typename Traits::connection db, typename Traits::Text const& name,   typename Traits::Text const& surname);

    // CREATE TABLE money (src INTEGER, dst INTEGER, amount INTEGER)
    static bool create_money(typename Traits::connection db);

    // INSERT INTO money VALUES
    static bool add_money(typename Traits::connection db, typename Traits::Int const& src, typename Traits::Int const& dst, typename Traits::Int const& amount);

    // SELECT name || ' ' || surname AS fullname, SUM(amount) as total FROM person JOIN money ON src = id GROUP BY id
    struct data_4
    {
        typename Traits::Text fullname;
        typename Traits::Int total;
    }; // struct data_4

    template<class T>
    static bool calc_total(typename Traits::connection db, T& result);

    // SELECT DISTINCT surname FROM person JOIN money ON src = id AND dst = (SELECT id FROM person WHERE surname LIKE @surname)
    struct data_5
    {
        typename Traits::Text surname;
    }; // struct data_5

    template<class T>
    static bool list_donors(typename Traits::connection db, T& result, typename Traits::Text const& surname, typename Traits::Int const& limit);

}; // struct sqlgg

Things to note above:

1. Syntax. All queries are written in plain SQL. Every statement should be terminated with semicolon. There is no need to write (?,?) after VALUES in INSERT statement. The annotation [sqlgg] name=function_name (or simply @function_name) before the query specifies the name of the generated function (this annotation optional but very convenient).
2. The generated code is parametrized by database-specific class Traits. It specifies the correspondence between SQL and native types, provides types for database connection and implements actual code to execute statements. Traits should be implemented once for every specific database API.
3. add_money() function takes three data parameters -- the values to INSERT into table (note the names and types).
4. calc_total() returns data via result parameter. Under the scenes it will bind columns of resulting rowset to the fields of T::value_type, which should provide fields fullname of type Traits::Text and total of type Traits::Int (otherwise it will fail to compile). For convenience a structure satisfying the requirements for output type is generated alongside the function, data_4 in this particular case, so std::vector<data_4> for result is fine.
5. The types of parameters for list_donors were inferred correctly (limit is Int and surname is Text. SQL is not a statically-typed language so the inferred types are based on some reasonable assumptions.
6. Statements of arbitrary depth across many tables should be supported.
7. Statements are checked for correctness as far as generator is concerned, so it will detect syntax errors, non-existent columns in expressions, mismatched columns in compound statements, ambiguous column names etc.

Then manually-written C++ code boils down to (without error-checking):

#include "../sqlite3_helper.hpp" // sqlite3 traits
#include "demo_cxx_gen.hpp" // generated
#include <iostream>
#include <vector>

using namespace std;

typedef sqlgg<sqlite3_traits> gen;
typedef long long int64;

int main()
{
  sqlite3* db = NULL;
  sqlite3_open(":memory:", &db);

  // create tables
  gen::create_person(db);
  gen::create_money(db);

  // add all person records
  gen::add_person(db,"John","Black");
  int64 john = sqlite3_last_insert_rowid(db);
  gen::add_person(db,"Ivan","Petrov");
  int64 ivan = sqlite3_last_insert_rowid(db);
  gen::add_person(db,"Sancho","Alvares");
  int64 sancho = sqlite3_last_insert_rowid(db);

  // add money relations
  gen::add_money(db,john,ivan,200);
  gen::add_money(db,john,sancho,100);
  gen::add_money(db,john,sancho,250);
  gen::add_money(db,sancho,ivan,300);

  // summarize by person
  typedef vector<gen::data_4> collection;
  collection all;
  gen::calc_total(db,all);

  // output
  cout << "Total transfers:" << endl;
  for (collection::const_iterator i = all.begin(), end = all.end(); i != end; ++i)
  {
     cout << i->fullname << " = " << i->total << endl;
  }

  // list donors
  typedef vector<gen::data_5> people;
  people p;
  gen::list_donors(db,p,"petrov",100);

  cout << "Donors:" << endl;
  for (people::const_iterator i = p.begin(), end = p.end(); i != end; ++i)
  {
    cout << i->surname << endl;
  }

  // properly close database
  sqlite3_close(db);

  return 0;
}

The code is straightforward and free of SQL-specific details. More importantly it is statically checked by the compiler -- suppose we change the database schema and add one more field to money table (e.g. timestamp of transfer). Compilation rightfully fails:

demo_cxx.cpp: In function `int main()':
demo_cxx.cpp:29: error: no matching function for call to
`sqlgg<sqlite3_traits>::add_money(sqlite3*&, int64&, int64&, int)'
demo_cxx_gen.hpp:75: note: candidates are: static bool sqlgg<Traits>::add_money(typename
Traits::connection, const typename Traits::Int&, const typename Traits::Int&, const typename
Traits::Int&, const typename Traits::Int&) [with Traits = sqlite3_traits]

Details

The main idea is that the generator should take care only of semantic binding between SQL and code sides, being as unobtrusive as possible. So the choice of the specific database and API is a programmer's choice. Similarly, queries to the database are expressed in plain SQL, so that the generator can be easily plugged in any existing project -- just move all SQL statements used in the code to separate file and feed it to generator.

Distinguishing feature of sqlgg is that it starts off with actual SQL queries, not object models or SQL table descriptions. Ideally, it generates {meta,template,abstract,generic} code which is later specialized for actual database/environment (if the target language provides such ability).

TODO

• reuse prepared statements
• query parameter placeholders are specific to database API, decide what to do
• choose better names for some common cases (WHERE id = ? etc)
• fix line numbers in error output
• enhance error reporting
• type-inferrer is too primitive
• detect statements on single tables and group the corresponding generated code together
• check function names for uniqueness
• support/test other SQL engines
• generate code for more languages
• read SQL spec
• distinguish predicates and expressions
• type check expressions

2009-05-25