@Aklapper I've updated the desc

Unfortunately I do not think this is very feasible. I already inquired upstream about this and was told that it's complicated...

Let me see if I can find the upstream task.

See also T190342 and T224082.

I believe that swapping this to a ProjectFunctionDatasource would "just work", provided Global search is served by Ferret. On this install, Global search is served by ElasticSearch instead of Ferret, so swapping the tokenizer would cause us to start submitting none(), not(...), any(...), etc., queries to ElasticSearch as string literals. All of these queries will fail so the net effect would be that using any tokenizer function would never match any results.

Supporting these queries against ElasticSearch would require implementing the "EdgeLogic" behavior from PhabricatorCursorPagedPolicyAwareQuery in terms that ElasticSearch can understand. Some of these constraints are very complicated -- for example, if you search for "Tags: A, B" the search really means "find tasks tagged with A, and/or any child of A, and B, and/or any child of B, but they must be tagged with at least one project in the A set and at least one project in the B set". In MySQL, we express this like this (the "ancestor" edge logic operator):

case PhabricatorQueryConstraint::OPERATOR_ANCESTOR:
  // This is tricky. We have a query which specifies multiple
  // projects, each of which may have an arbitrarily large number
  // of descendants.

  // Suppose the projects are "Engineering" and "Operations", and
  // "Engineering" has subprojects X, Y and Z.

  // We first use `FIELD(dst, X, Y, Z)` to produce a 0 if a row
  // is not part of Engineering at all, or some number other than
  // 0 if it is.

  // Then we use `IF(..., idx, NULL)` to convert the 0 to a NULL and
  // any other value to an index (say, 1) for the ancestor.

  // We build these up for every ancestor, then use `COALESCE(...)`
  // to select the non-null one, giving us an ancestor which this
  // row is a member of.

  // From there, we use `COUNT(DISTINCT(...))` to make sure that
  // each result row is a member of all ancestors.
  if (count($list) > 1) {
    $idx = 1;
    $parts = array();
    foreach ($list as $constraint) {
      $parts[] = qsprintf(
        $conn,
        'IF(FIELD(%T.dst, %Ls) != 0, %d, NULL)',
        $alias,
        (array)$constraint->getValue(),
        $idx++);
    }
    $parts = qsprintf($conn, '%LQ', $parts);

    $select[] = qsprintf(
      $conn,
      'COUNT(DISTINCT(COALESCE(%Q))) %T',
      $parts,
      $this->buildEdgeLogicTableAliasAncestor($alias));
  }
  break;

I'm not experienced enough with ElasticSearch to know if this is easy/hard/impossible to express with Elastic syntax.

Ideally, the simplest approach is to turn off Elastic search and cross our fingers; if Ferret works fine and we don't actually need the Elastic engine, we can eventually deprecate/remove Elastic support and this becomes a 1-line change.

If we need to retain Elastic support, implementing all of these functions against the Elastic engine is likely quite involved. (Another possibility is to pull this part of the query out of Elastic and execute it in the application instead, but that creates a different set of also-hard problems.)

@epriestley: I'd like to try switching off elastic and see how it goes for a while. If we don't get a bunch of complaints then it's a big win in terms of code maintenance going forward.