Performance improvements for SPARQL’s regex in Tracker

The original SPARQL regex support of Tracker is using a custom SQLite function. But of course back when we wrote it we didn’t yet think much about optimizing. As a result, we were using g_regex_match_simple which of course recompiles the regular expression each time.

Today Jürg and me found out about sqlite3_get_auxdata and sqlite3_set_auxdata which allows us to cache a compiled value for a specific custom SQLite function for the duration of the query.

This is much better:

static void
function_sparql_regex (sqlite3_context *context,
                       int              argc,
                       sqlite3_value   *argv[])
{
  gboolean ret;
  const gchar *text, *pattern, *flags;
  GRegexCompileFlags regex_flags;
  GRegex *regex;

  if (argc != 3) {
    sqlite3_result_error (context, "Invalid argument count", -1);
    return;
  }

  regex = sqlite3_get_auxdata (context, 1);
  text = sqlite3_value_text (argv[0]);
  flags = sqlite3_value_text (argv[2]);
  if (regex == NULL) {
    gchar *err_str;
    GError *error = NULL;
    pattern = sqlite3_value_text (argv[1]);
    regex_flags = 0;
    while (*flags) {
      switch (*flags) {
      case 's': regex_flags |= G_REGEX_DOTALL; break;
      case 'm': regex_flags |= G_REGEX_MULTILINE; break;
      case 'i': regex_flags |= G_REGEX_CASELESS; break;
      case 'x': regex_flags |= G_REGEX_EXTENDED; break;
      default:
        err_str = g_strdup_printf ("Invalid SPARQL regex flag '%c'", *flags);
        sqlite3_result_error (context, err_str, -1);
        g_free (err_str);
        return;
      }
      flags++;
    }
    regex = g_regex_new (pattern, regex_flags, 0, &error);
    if (error) {
      sqlite3_result_error (context, error->message, error->code);
      g_clear_error (&error);
      return;
    }
    sqlite3_set_auxdata (context, 1, regex, (void (*) (void*)) g_regex_unref);
  }
  ret = g_regex_match (regex, text, 0, NULL);
  sqlite3_result_int (context, ret);
  return;
}

Before (this was a test on a huge amount of resources):

$ time tracker-sparql -q "select ?u { ?u a rdfs:Resource . FILTER (regex(?u, '^titl', 'i')) }"
real	0m3.337s
user	0m0.004s
sys	0m0.008s

After:

$ time tracker-sparql -q "select ?u { ?u a rdfs:Resource . FILTER (regex(?u, '^titl', 'i')) }"
real	0m1.887s
user	0m0.008s
sys	0m0.008s

This will hit Tracker’s master today or tomorrow.