i#102625 avoid fetching same row twice in different queries

We do a "SELECT * FROM table" just to fetch the primary key columns;
so reuse the same XResultSet to fetch all columns.
Else, we immediately issue a "SELECT * FROM table WHERE
primary_key=current_value" to read the other columns, which is
wasteful and particularly silly.

Commit 1ae17f5b03cc14844fb600ca3573a96deb37ab3b already tried
to do that, but was essentially reverted piecewise because
it caused fdo#47520, fdo#48345, fdo#50372.

Commit c08067d6da94743d53217cbc26cffae00a22dc3a thought it did that,
but actually reverted commit 1ae17f5b03cc14844fb600ca3573a96deb37ab3b.

This implementation fetches the whole current row and caches it in memory;
only one row is cached: when the current row changes, the cache contains
the new current row.

This could be problematic (wrt to memory consumption) if the current
row is big (e.g. with BLOBs) and nobody is interested in the data
anyway (as would often be the case with BLOBs). Note that because of
our "SELECT *", the driver most probably has it in memory already
anyway, so we don't make the situation that much worse.

This could be incrementally improved with a heuristic of not
preemptively caching binary data (and also not LONGVARCHAR / TEXT /
MEMO / ...); a getFOO on these columns would issue a specific "SELECT
column FROM table WHERE primary_key=current_value" each time.

The *real* complete fix to all these issues would be to not do "SELECT
*" at all. Use "SELECT pkey_col1, pkey_col2, ..." when we are only
interested in the key columns. As to data, somehow figure out which
columns were ar interested in and "SELECT" only these (and maybe only
those with "small datatype"?). Interesting columns could be determined
by our caller (creator) as an argument to our constructor, or some
heuristic (no binary data, no "big" unbound data).
Also be extra smart and use *(m_aKeyIter) when getFOO is called
on a column included in it (and don't include it in any subsequent
SELECT).

However, there are several pitfalls.

One is buggy drivers that give use column names of columns that we
cannot fetch :-| Using "SELECT *" works around that because the driver
there *obviously* gives us only fetchable columns in the result.

Another one is the very restrictive nature of some database access
technologies. Take for example ODBC:

- Data can be fetched only *once* (with the SQLGetData interface;
bound columns offer a way around that, but that's viable only for
constant-length data, not variable-length data).

This could be addressed by an intelligent & lazy cache.

- Data must be fetched in increasing order of column number
(again, this is about SQLGetData).

This is a harder issue. The current solution has the nice advantage
of completely isolating the rest of LibO from these restrictions.

I don't currently see how to cleanly avoid (potentially
unnecessarily) caching column 4 if we are asked for column 3 then
column 5, just in case we are asked for column 4 later on, unless
we issue a specific "SELECT column4" later. But the latter would be
quite expensive in terms of app-to-database roudtripe times :-( and
thus creates another performance issue.

Change-Id: I999b3f8f0b8a215acb390ffefc839235346e8353

0 files changed, 0 insertions, 0 deletions