IMMEDIATE ISSUES - look in to making query and command's _handle a shared_ptr, so that query and command classes can be copied. - rename _bind_index and _column_index to _next_* - change the transaction_guard interface so you can dereference it to get to its transaction and the transactions take care of not rolling back or committing when they already have done (as well as resetting in-progress queries). - turn on extended errcodes in open() and handle them in sqlite_error - query::prepare() isn't being called during construction (from basic_statement's constructor) - add columns() to row that returns a boost::tuple of various types so multple columns can be fetched at once (look in to using BOOST_PP_ITERATE macro) - add a row::column() that can take a column name. This is nexessary when doing a "SELECT *" and you don't know the column indicies. To implement this, the first time it is called, a column-name-to-index lookup would have to be built. This should be done in the query, not the row. This means that the row will have to know it's query (currently is copies its _handle) to be able to call column_index() on it. Is this a problem? - use sqlite3_db_mutex() to provide extended error information during sqlite_error construction. The general procedure would be to lock the db mutex, perform some sqlite3 command, check the error code, throw an sqlite_error (whilst obtaining extended error info) and then unlock the db mutex. Two options: - a macro would be simple - a templated safe-calling object (passing the comman's arg types as template params) may be overkill LONGER TERM IDEAS - make basic_statement and database keep a shared pointer to the database handle so the classes can be made copyable. The wrappers around the handle (implemented in sqlite::detail) can clean them up after use. This will also make the implementation of rows (to get round the forced non-dependency of rows on querys) a little easier to swallow. - A similar wrapper should be created for statement handles, making basic_statements, querys and commands copyable. Could weak_ptrs to these also be used in the database's list active querys? - expand sqlite_error - perhaps use boost::system_error (see boost/asio/error.hpp for an example of extending system_error) - see if we can #include "sqlite.h" in to a namespace. Pros: - we better encapsulate the library - we can reuse "sqlite3" as a namespace Cons: - makes access to real sqlite stuff awkward to sqlite3cc users, but does this matter? they can't access database._handle anyway! - potential incompatibility when linking to libraries that also link against sqlite