Blog

Dynamically changing Views table joins

By joachim, Fri, 10/28/2011 - 08:52

I've recently had cause to make Views make joins to tables in peculiar ways. Here's some notes on the peculiar things I did with the views_join class to accomplish that.

First of all I'll briefly recap how we define a table to Views. Each item in the $data array returned to hook_view_data() represents all the information about a table. Each key in the array is a field on that table (well, or pseudofield), except for the 'table' key which has the basic data about our table, like this:

$data['my_table'] = array(
  // This defines how the table joins back to different bases.
  'table' => array(
    // How to join back to the base table 'crm_party'.
    'crm_party' => array(
      'left_field' => 'pid',
      'field' => 'pid',
    ),
  ),
);
// Now we can add field definitions on this table.

That's the simplest case. It says, 'to join back to {crm_party}, join on the column 'pid' on both tables'. (Note I will say 'column' when I am speaking of the database, and 'field' for Views, though that can mean both a field that you add to the view, and a field on the table that provides filters, sorts, or arguments.)

So adding a field on this table will cause Views to add this join clause to the query:

... JOIN my_table ON crm_party.pid = my_table.pid

We can easily join where the columns have different names, by giving different values for 'field' and 'left_field' in the table definition.

If the join requires conditions, that's where the 'extra' clause comes in, like this:

// How to join back to the base table 'crm_party'.
'crm_party' => array(
  'left_field' => 'pid',
  'field' => 'pid',
  'extra' => 'foo = 42',
),

This now gives us:

... JOIN my_table ON crm_party.pid = my_table.pid AND foo = 42

The 'extra' can also take an array, in which case each item is an array containing field, operator, and value. (If it seems a bit like the Database API, but not quite, that's because all this was introduced in Views 2 on Drupal 6).

// How to join back to the base table 'crm_party'.
'crm_party' => array(
  'left_field' => 'pid',
  'field' => 'pid',
  'extra' => array(
    // The 'extra' array is numeric, hence has no keys. This always looks odd to me!
    array(
      'field' => 'foo',
      'value' => 42,
      'numeric' => TRUE,
    ),
  ),
),

So far, this is all covered in the Advanced Help documentation contained within Views. But for our relationship handler from CRM Parties to attached entities, we needed a condition on the join depending on values selected in the UI. So the 'extra', defined in hook_views_alter(), won't do, as it's not changeable. Or is it?

When a relationship handler is adding itself to the query, the query hasn't been fully built yet. Rather, Views has a views_plugin_query_default object which will eventually be used to make a DatabaseAPI SelectQuery. This means we can actually reach into the table queue and change the definition for any table to the left of us, like this:

// Our relationship handler's query method:
function query() {
  // Call our parent query method to set up all our tables and joins.
  parent::query();

  if ($this->options['main']) {
    // This is a little weird.
    // We don't add an 'extra' (ie a further join condition) on our
    // relationship join, but rather on the join that got us here from
    // the {crm_party} table.
    // This means reaching into the query object's table queue and fiddling
    // with the join object.
    // Setting a join handler for the join definition is not useful, as that
    // would have no knowledge of the user option set in this relationship
    // handler.
    // @todo: It might however be cleaner to set one anyway and give it
    // a method to add the extra rather than hack the object directly...
    $table = $this->table;
    $base_join = $this->query->table_queue[$table]['join'];
    $base_join->extra = array(array('field' => 'main', 'value' => TRUE));
  }

When I wrote those comments in the code last week, I'd found that using a custom join handler isn't useful, because that has no knowledge of the relationship handler's data. However, this week I found myself working on a different case where I did need to find a way to do just that.

This week's problem was how to filter out Drupal Commerce products that are in the current cart, or more generally in any order (and the current cart's order ID can be supplied with a default argument plugin).

It seems a reasonable enough thing to ask of Views, but it's actually pretty complex, as what's in a cart or order is not products but line items, each of which refers to a product with a reference field.

After several failed attempts, I managed to write a query that produces the correct result, but it requires joining to a subquery which itself has the order ID within it (see the issue for gory details).

The first hurdle with this is easy to overcome: there's nothing wrong about telling Views about a table that doesn't exist. This is often done with aliased tables, but in fact it can be totally fictional provided we also provide our own join handler which understands what to do to the query. That can be anything, as long as the SELECT fields we also add make sense. Hence it's fine to do this in hook_views_data():

// Fake table for the 'product is in order' argument, made from a subquery.
$data['commerce_product_commerce_line_item'] = array(
  'table' => array(
    'group' => 'Commerce Product',
    'join' => array(
      // Join to the commerce_product base.
      'commerce_product' => array(
        'left_field' => 'entity_id',
        'field' => 'line_item_id',
        'handler' => 'views_join_commerce_product_line_item',
      ),
    ),
  ),
);

Our custom join class now has to add the subquery to the view, but it also needs the argument value to do this.

The way I worked around this was to override the ensure_my_table() method in the argument handler. Normally, this calls $this->query->ensure_table() which then creates the join, but ensure_table() can take a join parameter to work with. The overridden version of ensure_my_table() creates the join object, and sets the argument value on it:

function ensure_my_table() {
  // Pre-empt views_plugin_query_default::ensure_table() by setting our join up now.
  // Argh, hack this in for now. This may mean relationships using this break?
  $relationship = 'commerce_product';

  // Get a join object for our table.
  // This is of class views_join_commerce_product_line_item, which takes
  // care of joining to a subquery rather than a table.
  $join = $this->query->get_join_data($this->table, $this->query->relationships[$relationship]['base']);

  // We add the argument value to the join handler as it needs to use it
  // within its subquery.
  $join->argument = $this->argument;

This means that in the build_join() method for our custom join handler, views_join_commerce_product_line_item, we can rely on the argument value that the views has received:

function build_join($select_query, $table, $view_query) {
  // (snip...) build a SelectQuery object for the subquery
  // Set the condition based on the argument value.
  $subquery->condition('cli.order_id', $this->argument);
  // (snip...)
  // Add the join on the subquery.
  $select_query->addJoin($this->type, $subquery, $table['alias'], $condition);

The views_join class's build_join() method is where the Views system of building a query is translated into a DatabaseAPI SelectQuery object. Here we build up our own query (and we don't need Views-safe aliases, as it's a completely internal, non-correlated subquery), and pass it in as a join which uses it as a subquery.

It remains only for the argument handler's query() method to add the conditions for its field, using the alias and field names we gave for the subquery.

In conclusion, the data structure Views understands may appear to be a fixed, declared thing, but with a little bit of tweaking the way tables are joined in a Views query can be affected by both site configuration and user input.

It's Amazing What You Find: Crusty Bits of the Menu System

By joachim, Sun, 10/17/2010 - 16:13

I've been poking in the innards of the menu system the last few days. This is due to yet another client wanting to do something that goes completely against the grain of Drupal.

In this case, it's the way Drupal only shows you menu links you have access to. This to me seems perfectly reasonable good usability: why show you something you can't use? On the other hand, there is a long-standing feature in Comment module that shows anonymous users a 'Login or register to post comments' link on nodes, so 'incitements to action' or whatever the social media buzzword is do exist in Drupal.

So the challenge was to show a link that the anonymous user can't use, send them to login, and back to the link they wanted in the first place. The last part is just some hook_form_alter() work with form redirection (though it did allow me to discover that everything drupal_get_form() is passed is available to the alter hooks, just in a funny place). The access to the menu item is done by intercepting in hook_menu_item_alter() to save a twin of the menu item, and a checkbox in the menu edit form to trigger this. Even registering the path is easy: a custom menu access callback which takes as access arguments those of the original item plus its access callback and negates whatever the original item would return for access. The part I'm (so far) stuck on is getting hook_menu_alter() to know about what the admin user has done in hook_menu_item_alter(): the next job will probably involve creating a truly ugly query that uses %LIKE% to grab menu items based on their options array.

But that's not what I came here to tell you about today.

In my prodding around of the menu system, I found that menu router items have a 'block_callback' property, with its own database field and everything. Now get this: only one item in the entire {menu_router} table on D6 has this filled, and to boot, it has absolutely no effect. (It's the admin theme page, by the way.) This property is something to do with the way that the root admin page is made out of things that are sort of blocks, but not quite. If the property exists on the router item, then the callback is called to add (!!) to the content. (The admin theme page of course plays no part in the root admin page.)

So we had here a completely useless database field, probably left over from Drupal 5 or even earlier. By the way, in a standard Drupal 7 install, the database field is completely unused. I made a dummy patch to get the issue queue testbot confirm that we never get to this particular piece of code, and the superfluous field has now been removed. It's amazing what you find!

By the way, if anyone fancies some bikeshedding, I still don't have a name better than 'menu_login' for this module. To your paintbrushes!

The Oxford Comma

By joachim, Wed, 06/30/2010 - 15:59

Here's a little function I wrote today because I needed to be able to turn a list of between one and three items into a string like 'apples, oranges, and pears', 'apples and pears', or just 'stairs'.

I figured I might as well handle everything in one place, and throw in the option to have an 'or' instead of an 'and'. There may be occasions you don't want the Oxford comma, but I can't think of any.

/**
* Grammatically fun helper to make a list of things in a sentence, ie
* turn an array into a string 'a, b, and c'.
*
* @param $list
*  An array of words or items to join.
* @param $type
*  The text to use between the last two items. Defaults to 'and'.
* @param $oxford
*  Change this from default and you are a philistine.
*/
function oxford_comma_list($list, $type = 'and', $oxford = TRUE) {
  $final_join = " $type ";
  if ($oxford && count($list) > 2) {
    $final_join = ',' . $final_join;
  }
  $final = array_splice($list, -2, 2); 
  $final_string = implode($final_join, $final);
  array_push($list, $final_string);
  return implode(', ', $list);
}

Now the real question: how would you make this translatable?

Using Constants For Permission Names: WHY?

By joachim, Wed, 06/30/2010 - 14:19

I keep seeing this sort of thing in so many modules:

define("MY_MODULE_PERM_ACCESS_WIDGETS", 'access widgets');
// Names changed to protect the guilty ;)

Am I missing something, or is this utterly pointless?

The only advantage I see is that you can change the permission string later on. But you actually can't change a permission string once you've made a release without an almighty amount of work in a hook_update_N()[*] to migrate users' existing permissions, so what is the point of using a constant apart from just creating shouty caps noise?

[*] Is there a helper function in core yet for migrating permission names during updates? There really should be.

Once & Only Once: The Conversion To Drupal 7's FieldAPI

By joachim, Tue, 04/06/2010 - 14:06

I'm not a fan of repeating myself, or of doing the same work twice. So when I first got a look at FieldAPI back at DrupalCon Paris, my thought after 'Wow this is going to change everything,' was 'Every module that converts its custom data storage to this is going to be doing the same work, over and over'.

It seemed to me that we have a lot of modules that add things to stuff. The examples that spring to mind include Taxonomy image (add an image to a taxonomy term), Comment upload (add an uploaded file to a comment), User terms (you get the picture), and the daddy of them all, Image (which now we must call Image Oldskool, due to there being a shiny new Image module in Core).

What all these have in common is that on Drupal 7 their things-to-stuffness would be perfectly served by FieldAPI. Users get more flexibility, an expanding universe of formatters and widgets; module maintainers can write less code and in some cases even hang up their hats on a particular project as it can live entirely in configuration space. Everyone wins. But how to get there?

This matter has been bubbling at the back of my mind ever since Paris. On the one hand I could sort of see how this could be all done with a single framework: you tell it what sort of entity you are dealing with (node, comment, term), which bundle (article, page, which vocab), what your fields are, and how to load the old data. The framework creates the fields for you, then runs a batch to load each object, moves values around in clever ways that you don't need to worry about, and then saves the object. Banzai! Your data is now saved in a field.

On the other hand, this was clearly crack. Of the highest order.

And yet it works. It quite possibly still is crack, and some extra pairs of eyes to de-crackify it would be very welcome. But I can confirm that I've run data conversions on both User terms and oldskool Image nodes, and got term fields on my users and beautiful image fields on my nodes.

There remains some work to be done: I'd like to make conversion to file and image fields a bit more straightforward and create a nice way to specify how to populate the alt and title fields for images. And all this is largely academic and has only run on D7 databases with dummy D6 tables imported, since core's 6 to 7 upgrade process is not currently working. So I could use a hand. And if you have a things-to-stuff module, get in touch and give it a whirl.