double ParameterizedFilter::EvaluateConditionNonJoinWeight(Descriptor &d, bool for_or)
{
// Interpretation of weight:
// an approximation of logarithm of answer size
// (in case of for_or: an approximation of (full_table - answer) size)
// 0 -> time is very short (constant).
// high weight -> schedule this query to be executed later
double eval = 0.0;
_uint64 no_distinct, no_distinct2;
_uint64 answer_size;
if(d.IsTrue() || d.IsFalse())
eval = 0; // constant time
else if(d.IsType_AttrValOrAttrValVal()) { // typical condition: attr=val
if(!d.encoded) {
return log(1 + double(d.attr.vc->NoTuples())) + 5; // +5 as a penalty for complex expression
}
SingleColumn* col = static_cast<SingleColumn*>(d.attr.vc);
answer_size = col->ApproxAnswerSize(d);
if(for_or)
answer_size = d.attr.vc->NoTuples() - answer_size;
_int64 no_in_values = 1;
if(d.op == O_IN || d.op == O_NOT_IN) {
MultiValColumn* iscol = static_cast<MultiValColumn*>(d.val1.vc);
no_in_values = iscol->NoValues(NULL);
}
eval = log(1 + double(answer_size)); // approximate size of the result
if(no_in_values > 1)
eval += log(double(no_in_values)) * 0.5; // INs are potentially slower (many comparisons needed)
if(col->Type().IsString() && !col->Type().IsLookup())
eval += 0.5; // strings are slower
if(col->Type().IsFloat())
eval += 0.1; // floats are slower
if(d.op == O_LIKE || d.op == O_NOT_LIKE)
eval += 0.2; // these operators need more work
} else if(d.IsType_AttrAttr()) { // attr=attr on the same table
_uint64 no_obj = d.attr.vc->NoTuples(); // changed to uint64 to prevent negative logarithm for NULL_VALUE_64
if(!d.encoded)
return log(1 + double(2 * no_obj)) + 5; // +5 as a penalty for complex expression
else if(d.op == O_EQ) {
no_distinct = d.attr.vc->GetApproxDistVals(false);
if(no_distinct == 0)
no_distinct = 1;
no_distinct2 = d.val1.vc->GetApproxDistVals(false);
if(no_distinct2 == 0)
no_distinct2 = 1;
if(no_distinct2 > no_distinct)
no_distinct = no_distinct2; // find the attribute with smaller abstract classes
if(for_or)
eval = log(1 + (no_obj - double(no_obj) / no_distinct));
else
eval = log(1 + double(no_obj) / no_distinct); // size of the smaller abstract class
} else {
eval = log(1 + double(no_obj) / 2); // other operators filter potentially a half of objects
}
eval += 1; // add to compensate opening two packs
} else if(d.IsType_OrTree() && !d.IsType_Join()) {
eval = d.tree->root->EvaluateConditionWeight(this, for_or);
} else { // expressions and other types, incl. joins (to be calculated separately)
if(d.IsType_IBExpression())
return log(1 + double(d.attr.vc->NoTuples())) + 2; // +2 as a penalty for IB complex expression
eval = 99999;
}
return eval;
}
