On Fri, 29 Dec 2006, Kamesh Jayachandran wrote:
> >In this particular case, your change logs should actually have been a
> >little different, since you introduced a behavioral difference in the
> >get_merge_info() function.
> Do you mean 'select 1 from mergeinfo' as a behavioural change?
> >+ Our queries will never cause SQL error, so we can assume SQLITE_ERROR
> >to be
> >You hope. What if a well-meaning sysadmin modified our table
> >structure, or fiddled around in some other fashion? I wouldn't make
> >such an absolute statement.
> sqlite operations to the db is limited to
> 'subversion/libsvn_fs_util/merge-info-sqlite-index.c' where as of today
> I see any no place we drop tables. And I could not envisage a situation
> where the constructed query could go wrong(May be I am wrong here, would
> need to see the queries much closely.)
I'm referring to unexpected differences, rather than to differences
that we can predict based on Subversion's code.
> Anyways CREATE_MERGEINFO_SCHEMA is used only in 2 places(delete from
> mergeinfo_changed where revision=some_rev and select 1 from mergeinfo).
> In both such places we never get SQLITE_ERROR except when those table
> does not exist.
> >+#define CREATE_MERGEINFO_SCHEMA(x, db) do
> >"err" would be a more meaningful name than "x".
> I feel retval could be more meaningful.
That sounds good.
> >@@ -242,7 +257,7 @@
> > deletestring = apr_psprintf(pool,
> > "delete from mergeinfo_changed where
> > revision = %ld;",
> > new_rev);
> >- SVN_ERR(util_sqlite_exec(db, deletestring, NULL, NULL));
> >+ CREATE_MERGEINFO_SCHEMA(sqlite3_exec(db, deletestring, NULL, NULL,
> >NULL), db);
> >I guess we don't need this SQL statement to succeed, if the schema
> >hasn't been created yet.
> Yes we don't need this to succeed in normal cases(Where we proper schema
> in place).
> When schema not in place I feel this is the right candidate to check for
> schema existence as it does not need any re-execution.
> >@@ -502,6 +517,9 @@
> > sqlite3_trace (db, sqlite_tracer, db);
> > #endif
> >+ CREATE_MERGEINFO_SCHEMA(sqlite3_exec(db, "select 1 from mergeinfo",
> >+ NULL, NULL, NULL),
> >+ db);
> >-1 on issuing a "select 1" on this often-invoked code path.
> How often?
> 1)svn url1 url2
> 2)svn merge
> In both these cases we are running it only once(This query should be no
> cost query)
(At least) 'merge', 'copy', and 'move' should call get_merge_info().
While I have not measured this code, and I concur that a "select 1"
*should* be inconsequential, performance-wise, my concern about an
"often-invoked code path" is more that this is happening on the
repository-side, meaning that in a highly-concurrent environment this
effectively no-op set of instructions could be getting run quite a
I'd prefer that instead of using a no-op statement like "select 1",
that we do what we typically do across the Subversion code base, and
attempt the actual operation we want to perform. If that operation
fails, detect a missing schema, create the schema, and re-try the
original SQL. This does somewhat throw a wrench in your macro-based
implementation, but I'm sure you'll come up with an adjusted
implementation which fits the bill.
Received on Fri Dec 29 21:06:47 2006
- application/pgp-signature attachment: stored