Merge pull request #5934 from github/hmakholm/pr/monotonic-agg

QL language reference: add monotonic aggregate example

Author: hmakholm

docs/codeql/ql-language-reference/expressions.rst

@@ -488,6 +488,101 @@ value for each value generated by the ``<formula>``:
   value generated by the ``<formula>``. Here, the aggregation function is applied to each of the 
   resulting combinations.
 
+Example of monotonic aggregates
+-------------------------------
+
+Consider this query:
+
+.. code-block:: ql
+
+   string getPerson() { result = "Alice" or
+                        result = "Bob" or
+                        result = "Charles" or
+                        result = "Diane"
+                      }
+   string getFruit(string p) { p = "Alice"   and result = "Orange" or
+                               p = "Alice"   and result = "Apple" or
+                               p = "Bob"     and result = "Apple" or
+                               p = "Charles" and result = "Apple" or
+                               p = "Charles" and result = "Banana"
+                             }
+   int getPrice(string f) { f = "Apple"  and result = 100 or
+                            f = "Orange" and result = 100 or
+                            f = "Orange" and result =   1
+                          }
+
+   predicate nonmono(string p, int cost) {
+     p = getPerson() and cost = sum(string f | f = getFruit(p) | getPrice(f))
+   }
+
+   language[monotonicAggregates]
+   predicate mono(string p, int cost) {
+     p = getPerson() and cost = sum(string f | f = getFruit(p) | getPrice(f))
+   }
+
+   from string variant, string person, int cost
+   where variant = "default"  and nonmono(person, cost) or
+         variant = "monotonic" and mono(person, cost)
+   select variant, person, cost
+   order by variant, person
+
+The query produces these results:
+
++-----------+---------+------+
+|  variant  | person  | cost |
++-----------+---------+------+
+| default   | Alice   | 201  |
+| default   | Bob     | 100  |
+| default   | Charles | 100  |
+| default   | Diane   | 0    |
+| monotonic | Alice   | 101  |
+| monotonic | Alice   | 200  |
+| monotonic | Bob     | 100  |
+| monotonic | Diane   | 0    |
++-----------+---------+------+
+
+The two variants of the aggregate semantics differ in what happens
+when ``getPrice(f)`` has either multiple results or no results
+for a given ``f``.
+
+In this query, oranges are available at two different prices, and the
+default ``sum`` aggregate returns a single line where Alice buys an
+orange at a price of 100, another orange at a price of 1, and an apple
+at a price of 100, totalling 201. On the other hand, in the the
+*monotonic* semantics for ``sum``, Alice always buys one orange and
+one apple, and a line of output is produced for each *way* she can
+complete her shopping list.
+
+If there had been two different prices for apples too, the monotonic
+``sum`` would have produced *four* output lines for Alice.
+
+Charles wants to buy a banana, which is not for sale at all. In the
+default case, the sum produced for Charles includes the cost of the
+apple he *can* buy, but there's no line for Charles in the monontonic
+``sum`` output, because there *is no way* for Charles to buy one apple
+plus one banana.
+
+(Diane buys no fruit at all, and in both variants her total cost
+is 0. The ``strictsum`` aggregate would have excluded her from the
+results in both cases).
+
+In actual QL practice, it is quite rare to use monotonic aggregates
+with the *goal* of having multiple output lines, as in the "Alice"
+case of this example. The more significant point is the "Charles"
+case: As long as there's no price for bananas, no output is produced
+for him. This means that if we later do learn of a banana price, we
+don't need to *remove* any output tuple already produced. The
+importance of this is that the monotonic aggregate behavior works well
+with a fixpoint-based semantics for recursion, so it will be meaningul
+to let the ``getPrice`` predicate be mutually recursive with the count
+aggregate itself. (On the other hand, ``getFruit`` still cannot be
+allowed to be recursive, because adding another fruit to someone's
+shopping list would invalidate the total costs we already knew for
+them).
+
+This opportunity to use recursion is the main practical reason for
+requesting monotonic semantics of aggregates.
+
 Recursive monotonic aggregates
 ------------------------------