Multi-site search using Feeds and SearchAPI

[This is an old post that I wrote for System Seed's blog and meant to put on my own too but it fell off my radar until now. It's also about Drupal 7, but the general principle still applies.]

Handling clients with more than one site involves lots of decisions. And yet, it can sometimes seem like ultimately all that doesn't matter a hill of beans to the end-user, the site visitor. They won't care whether you use Domain module, multi-site, separate sites with common codebase, and so on. Because most people don't notice what's in their URL bar. They want ease of login, and ease of navigation. That translates into things such as the single sign-on that uses, and common menus and headers, and also site search: they don’t care that it’s actually sites search, plural, they just want to find stuff.

For the University of North Carolina, who have a network of sites running on a range of different platforms, a unified search system was a key way of giving visitors the experience of a cohesive whole. The hub site, an existing Drupal 7 installation, needed to provide search results from across the whole family of sites.

This presented a few challenges. Naturally, I turned to Apache Solr. Hitherto, I've always considered Solr to be some sort of black magic, from the way in which it requires its own separate server (http not good enough for you?) to the mysteries of its configuration (both Drupal modules that integrate with it require you to dump a bunch of configuration files into your Solr installation). But Solr excels at what it sets out to do, and the Drupal modules around it are now mature enough that things just work out of the box. Even better, Search API module allows you to plug in a different search back-end, so you can develop locally using Drupal's own database as your search provider, with the intention of plugging it all into Solr when you deploy to servers.

One possible setup would have been to have the various sites each send their data into Solr directly. However, with the Pantheon platform this didn't look to be possible: in order to achieve close integration between Drupal and Solr, Pantheon locks down your Solr instance.

That left talking to Solr via Drupal.

Search API lets you define different datasources for your search data, and comes with one for each entity type on your site. In a datasource handler class, you can define how the datasource gets a list of IDs of things to index, and how it gets the content. So writing a custom datasource was one possibility.

Enter the next problem: the external sites that needed to be indexed only exposed their content to us in one format: RSS. In theory, you could have a Search API datasource which pulls in data from an RSS feed. But then you need to write a SearchAPI datasource class which knows how to parse RSS and extract the fields from it.

That sounded like reinventing Feeds, so I turned to that to see what I could do with it. Feeds normally saves data into Drupal entities, but maybe (I thought) there was a way to have the data be passed into SearchAPI for indexing, by writing a custom Feeds plugin?

However, this revealed a funny problem of the sort that you don’t consider the existence of until you stumble on it: Feeds works on cron runs, pulling in data from a remote source and saving it into Drupal somehow. But SearchAPI also works on cron runs, pulling data in, usually entities. How do you get two processes to communicate when they both want to be the active participant?

With time pressing, I took the simple option: define a custom entity type for Feeds to put its data into, and SearchAPI to read its data from. (I could have just used a node type, but then there would have been an ongoing burden of needing to ensure that type was excluded from any kind of interaction with nodes.)

Essentially, this custom entity type acted like a bucket: Feeds dumps data in, SearchAPI picks data out. As solutions go, not the most massively elegant, at first glance. But if you think about it, if I had gone down the route of SearchAPI fetching from RSS directly, then re-indexing would have been a really lengthy process, and could have had consequences for the performance of the sites whose content was being slurped up. A sensible approach would then have been to implement some sort of caching on our server, either of the RSS feeds as files, or the processed RSS data. And suddenly our custom entity bucket system doesn’t look so inelegant after all: it’s basically a cache that both Feeds and SearchAPI can talk to easily.

There were a few pitalls. With Search API, our search index needed to work on two entity types (nodes and the custom bucket entities), and while Search API on Drupal 7 allows this, its multiple entity type datasource handler had a few issues to iron out or learn to live with. The good news though is that the Drupal 8 version of Search API has the concept of multi-entity type search indexes at its core, rather than as a side feature: every index can handle multiple entity types, and there’s no such thing as a datasource for a single entity type.

With Feeds, I found that not all the configuration is exportable to Features for easy deployment. Everything about parsing the RSS feed into entities can be exported, except the actual URL, which is a separate piece of setup and not exportable. So I had to add a hook_updateN() to take care of setting that up.

The end result though was a site search that seamlessly returns results from multiple sites, allowing users to work with a network of disparate sites built on different technologies as if they were all the same thing. Which is what they were probably thinking they were all along anyway.

Controlling multiple sites with Drush 9

Drush 9 has removed dynamic site aliases. Site aliases are hardcoded in YAML files rather than declared in PHP. Sadly, that means that many tricks you could do with the declaration of the site aliases are no longer available.

The only grouping possible is based on the YAML filename. So for example, with the Acquia Cloud Site Factory site aliases generated by the 'blt recipes:aliases:init:acquia' command, you can run a command on the same site across different environments.

But what you can't do is run a command on all the sites in one environment.

One use case for this is checking whether a module is enabled on any sites, so you know that it's safe to remove it from the codebase.

Currently, this is quite a laborious process, as 'drush pm-list' needs to be run for each site.

With environment aliases, this would be a one liner:

drush @hypothetical-env-alias pm-list | ag some_module

('ag' is the very useful silver searcher unix command, which is almost the same as the also excellent 'ack' but faster, and both are much better than grep.)

While site aliases are fixed, they can be altered with Drush hooks. I considered that these might allow something to dynamically declare aliases, or a command option. There's an example of altering aliases with a hook in the Drush code.

In the meantime, a much simpler solution is to use xargs, which I have recently found is extremely useful in all sorts of situations. Because this allows you to run one command multiple times with a set of parameters, all you need to do is pass it a list of site aliases. Fortunately, the 'drush sa' command has lots of formatting options, and one of them gives us just what we need, a list of aliases with one on each line:

drush sa --format=list 

That gives us all the aliases, and we probably don't want that. So here's where ag first comes in to play, as we can filter the list, for example, to only run on live sites (I'm using my ACSF aliases here as an example):

drush sa --format=list| ag 01live

Now we have a filtered list of aliases, and we can feed that into xargs:

drush sa --format=list| ag 01live | xargs -I % drush % pm-list

Normally, xargs puts the input parameter at the end of its command, but here we want it inserted just after the 'drush' command. The -I parameter allows us to specify a placeholder where the input parameter goes, so:

xargs -I % drush % pm-list

says that we want the site name to go where the '%' is, and means that that xargs will run:

drush SITE-ALIAS pm-list

with each value it receives, in this case, each site alias.

Another thing we will do with xargs is set the -t parameter, which outputs each actual command it executes on STDERR. That acts as a heading in the output, so we can clearly see which site is outputting what.

Finally, we can use ag a second time to filter the module list down to just the module we want to find out about:

drush sa --format=list | ag live | xargs -t -I % drush % pml | ag some_module 

The nice thing about the -t parameter is that as it's STDERR, it's not affected by the final pipe to ag for filtering output. So the output will consist of the drush command for the site, followed by the filtered output.

And hey presto.

In conclusion: dynamic site aliases in Drush were nice, but the maintainers removed them (as far as I can gather) because they were a mess to implement, and removing them vastly simplified things. Doing the equivalent with xargs took a bit of figuring out, but once you know how to do it, it's actually a much more powerful way to work with multiple sites at once.

Unnatural file changes with git

When you check out a branch or commit with git, two things happen: git changes the files in the repository folder, and changes the file that tells it what is currently checked out. In git terms, it changes what HEAD points to. In practical terms, it updates the .git/HEAD text file to contain a different reference.

But these two operations can be done separate from one another, so that while the files correspond to one commit, git thinks that the HEAD commit is another one.

This of course puts your git repository in an unstable, even unnatural state. But there are useful reasons for doing this, and one of these comes up in the operation of dorgflow, my tool for working with git and patches.

Dorgflow creates a local branch for a issue, and creates a commit for each patch, so you end up with a nice sequential representation of the work that's been done so far.

But every patch that's uploaded to an issue is a patch against the master branch, so we can't just apply the patches one by one and make commits: the first patch will apply, and all the subsequent ones will fail.

What we need is for the files be in the same state as the master branch, so that the patch applies, but when we make a commit, it needs to be a new commit on the feature branch:

 * [Feature branch] Patch 1 <- Make patch 2 commit on top of this commit.
* [master branch] Latest master commit. <- Apply patch 2 to these files.

In git terms, we want the current files (the working tree, as git calls it) to be on the master branch, while git's HEAD is on the feature branch.

For my first attempt at getting this to work, I did the following:

  1. Put git on the feature branch.
  2. Check out the master branch's files, with git checkout master -- .

When you use git checkout command with a file or files specified, it doesn't move HEAD, but instead changes just those files to the given commit or branch. If you give '.' as the file to check out, then it does that for all of the repository's files. In effect, you get the files to look like the given commit, in our case the master branch.

This is what we want, but it has one crucial flaw. Suppose patch 1 added a new file, foo.php. When we check out master's files, foo.php is not changed, because it's not on master. There's nothing on master to overwrite it. The git checkout master -- . command doesn't actually say 'make all the files look like master', it says 'check out all the files from master'. foo.php isn't on master, and so it's simply left alone.

Suppose now that patch 2 also adds the foo.php file, which it most likely will, since in most cases, a newer patch incorporates all the work of the previous one.

Applying patch 2 to the files in their current state will fail, because patch 2 tries to create the file foo.php, but it's already there.

So this approach is no good. I was stuck with this problem with dorgflow for months, until I had a brainwave: instead of staying on the feature branch and changing the files to look like master, why not check out the master branch, but tell git that the HEAD is at the feature branch?

That solves the problem of the foo.php file: when you check out master, the foo.php that's at the patch 1 commit vanishes, because it doesn't exist on master. So applying patch 2 will be fine.

The only remaining question was how to you tell git it's on a different branch without doing a checkout of files? Turns out this is a simple plumbing command: git symbolic-ref HEAD refs/heads/BRANCH. This is just telling git to change the reference that HEAD points to, and that's how git stores the fact that a particular branch is the current one.

This was a simple change to make in dorgflow (it was actually more work to update the tests so the mocks had the correct method call expectations!).

This means that dorgflow accordingly now handles applying a sequence of issue patches that add files now works in the latest release.

Getting more than you bargained for: removing a Drupal module with Composer

It's no secret that I find Composer a very troublesome piece of software to work with.

I have issues with Composer on two fronts. First, its output is extremely user-unfriendly, such as the long lists of impenetrable statements about dependencies that it produces when it tells you why it can't make a change you request. Second, many Composer commands have unwanted side-effects, and these work against the practice that changes to your codebase should be as simple as possible for the sake of developer sanity, testing, and user acceptance.

I recently discovered that removing packages is one such task where Composer has ideas of its own. A command such as remove drupal/foo will take it on itself to also update some apparently unrelated packages, meaning that you either have to manage the deployment of these updates as part of your uninstallation of a module, or roll up your sleeves and hack into the mess Composer has made of your codebase.

Guess which option I went for.

Step 1: Remove the module you actually want to remove

Let's suppose we want to remove the Drupal module 'foo' from the codebase because we're no longer using it:

$ composer remove drupal/foo

This will have two side effects, one of which you might want, and one of which you definitely don't.

Side effect 1: dependent packages are removed

This is fine, in theory. You probably don't need the modules that are dependencies of foo. Except... Composer knows about dependencies declared in composer.json, which for Drupal modules might be different from the dependencies declared in module info.yml files (if maintainers haven't been careful to ensure they match). UPDATE: I've been informed in comments that's packaging process ensures these are kept in sync. So that's one less thing to worry about!

Furthermore, Composer doesn't know about Drupal configuration dependencies. You could have the situation where you installed module Foo, which had a dependency on Bar, so you installed that too. But then you found Bar was quite useful in itself, and you've created content and configuration on your site that depends on Bar. Ideally, at that point, you should have declared Bar explicitly in your project's root composer.json, but most likely, you haven't.

So at this point, you should go through Composer's output of what it's removed, and check your site doesn't have any of the Drupal modules enabled.

I recommend taking the list of Drupal modules that Composer has just told you it's removed in addition to the requested one, and checking its status on your live site:

$ drush pml | ag MODULE

If you find that any modules are still enabled, then revert the changes you've just made with the remove command, and declare the modules in your root composer.json, copying the declaration from the composer.json file of the module you are removing. Then start step 1 again.

Side effect 2: unrelated packages are updated

This is undesirable basically because any package update is something that has to be evaluated and tested before it's deployed. Having that happen as part of a package removal turns what should be a straight-forward task into something complex and unpredictable. It's forcing the developer to handle two operations that should be separate as one.

(It turns out that the maintainers of Composer don't even consider this to be a problem, and as I have unfortunately come to expect, the issue on github is a fine example of bad maintainership (for the nadir, see the issue on the use of JSON as a format for the main composer file) -- dismissing the problems that users explain they have, claiming the problems are by design, and so on.)

So to revert this, you need to pick apart the changes Composer has made, and reverse some of them.

Before you go any further, commit everything that Composer changed with the remove command. In my preferred method of operation, that means all the files, including the modules folder and the vendor folder. I know that Composer recommends you don't do that, but frankly I think trusting Composer not to damage your codebase on a whim is folly: you need to be able to back out of any mess it may make.

Step 2: Repair composer.lock

The composer.lock file is the record of how the packages currently are, so to undo some of the changes Composer made, we undo some of the changes made to this file, then get Composer to update based on the lock.

First, restore version of composer.lock to how it was before you started:

$ git checkout HEAD^ composer.lock

Unstage it. I prefer a GUI for git staging and unstaging operations, but on the command line it's:

$ git reset composer.lock

Your composer lock file now looks as it did before you started.

Use either git add -p or your favourite git GUI to pick out the right bits. Understanding which bits are the 'right bits' takes a bit of mental gymnastics: overall, we want to keep the changes in the last commit that removed packages completely, but we want to discard the changes that upgrade packages.

But here we've got a reverted diff. So in terms of what we have here, we want to discard changes that re-add a package, and stage and commit the changes that downgrade packages.

When you're done staging you should have:

  • the change to the content hash should be unstaged.
  • chunks that are a whole package should be unstaged
  • chunks that change version should be staged (be sure to get all the bits that relate to a package)

Then commit what is staged, and discard the rest.

Then do a git diff of composer.lock against your starting point: you should see only complete package removals.

Step 3: Restore packages with unrelated changes

Finally, do:

$ composer update --lock

This will restore the packages that Composer updated against your will in step 1 to their original state.

If you are committing Composer-managed packages to your repository, commit them now.

As a final sanity check, do a git diff against your starting point, like this:

$ git diff --name-status master

You should see mostly deleted files. To verify there's nothing that shouldn't be there in the changed files, do:

$ git diff --name-status master | ag '^[^D]'

You should see only composer.json, composer.lock, and the autoloader's files.

PS. If I am wrong and there IS a way to get Compose to remove a package without side-effects, please tell me.

I feel I have exhausted all the options of the remove command:

  • --no-update only changes composer.json, and makes no changes to package files at all. I'm not sure what the point of this is.
  • --no-update-with-dependencies only removes the one package, and doesn't remove any dependencies that are not required anywhere else. This leaves you having to pick through composer.json files yourself and remove dependencies individually, and completely obviates the purpose of a package manager!

Why is something as simple as a package removal turned into a complex operation by Composer? Honestly, I'm baffled. I've tried reasoning with the maintainers, and it's a brick wall.

PPS. Since writing this post, I’ve made Composer Manifest a small Composer plugin which makes it easier to see what Composer has decided to change behind your back. Every time you do a Composer update, install, or remove, it writes a YAML file that lists all the installed packages with their versions. Committing that to your repository means you have an easy way to see exactly what's been changed and when.


Drupal Code Builder: the analytical, adaptive code generator

It's now just over 10 years since webchick handed maintainership of the Module Builder project over to me. I can still remember nervously approaching her at the end of a day at DrupalCon 2008, outside the building where people were chatting, and asking her if she'd had time to look at the patches I'd posted for it. I was at my first DrupalCon; she'd just been named as the core co-maintainer for the work about to start on Drupal 7. No, she hadn't, she said, and she instantly came to the conclusion that she'd never have the time to look at them, and so got her laptop out of her rucksack and there and then edited the project node to make me the maintainer.

Given the longevity of the project, I am often surprised when I talk to people that they don't know what its crucial advantage is over the other Drupal code generating tools that now also exist. Drupal Code Builder, as it's now called, is fundamentally different from Drupal Console's Generator command, and the Drupal Code Generator library that’s included in Drush 9.

I feel that perhaps this requires a buzzword, to make it more noticeable and memorable.

So here is one: Drupal Code Builder is an analytical code generator.

I'm going to add a second one: the Drupal Code Builder Drush commands and Module Builder (which still exists, though is now just module that provides a Drupal-based UI for the Drupal Code Builder library) are adaptive.

What do I mean by those?

When you first install Drupal Code Builder, it has no information about Drupal hooks, or plugin types, or services. There is no data in the codebase on how to generate a hook_form_alter() implementation, or a Block plugin, or how to inject the entity_type.manager service.

Before you can generate code, you need to run a command or submit an admin form which will analyse your entire codebase, and store the resulting data on hooks, plugin types, services, and other Drupal structures (the list of what is analysed keeps growing.).

The consequence of this is huge: Drupal Code Builder is not limited to core hooks, or to the hooks that it's been programmed for. Drupal Code Builder knows all the hooks in your codebase: from core, from contrib module, even from any of your own custom code that invents hooks.

Now repeat that statement for plugin types, for services, and so on.

And that analysis process can, and should, be repeated when you download a new module, because then any Drupal structures (hooks, plugin types, etc) that are defined in new modules gets added to the list of what Drupal Code Builder can generate.

This means that you can use Drupal Code Builder to do something like this:

  1. Generate a new plugin type. This creates a plugin manager service class and declaration, an annotation class that defines the plugin annotation, and an interface and base class for the plugins.
  2. Re-run DCB's analysis.
  3. Generate a plugin of the type you've just made.

When I say that this technique of analysis makes Drupal Coder Builder more powerful than other code generators, I'm not even blowing my own trumpet: the module I was handed in 2008 did the same. Back then in the Drupal 6 era is needed to connect to's CVS repository to download the api.php files that document a module's hooks (which is why the Drush command for updating the code analysis prior to Drush 9 is called mb-download). For Drupal 7, the api.php files were moved into the core Drupal codebase, so from that point the process would scan the site's codebase directly. For Drupal 8, I added lots more code analysis, of plugin types and services, but still, the original idea remains: Drupal Code Builder knows how to detect Drupal structures, so it can know more structures than can be hardcoded.

This analysis can get pretty complex. While for hooks, it's just a case of running a regex on api.php files, for detecting information about plugin types, all sorts of techniques are used, such as searching the class parentage of all the plugins of a type to try to guess whether there is a plugin base class. It’s not always perfect, and it could always use more refinement, but it’s a more powerful approach than hardcoding a necessarily finite number of cases. (If only plugin types were declared in some way, or if the documentation for them were systematic in the way hook documentation is, this would all be so much simpler, and much more importantly, it would be more accurate!)

My second buzzword, that UIs for Drupal Code Builder are adaptive, describes the way that neither of the Drush command nor the Drupal module need to know what Drupal Code Builder can build. They merely know the API for describing properties, and how to present them to the user, to then pass the input values to Drupal Code Builder to get the generated code.

This is analogous to the way that Form API doesn’t know anything about a particular form, just about the different kinds of form elements.

This isn’t as exciting or indeed relevant to the end-user, but it does mean that development on Drupal Code Builder can move quickly, as adding new things to generate doesn’t require coordinated releases of software packages for the UI. In fact, I think nearly every new release of Drupal Code Builder has featured some sort of new code generating functionality, while keeping the API the same.

For example, this used to be the UI for adding a route on Drupal and Drush:

Screenshot showing Drush UI

Screenshot showing Module Builder UI

A later release of Drupal Code Builder turned the UI into this, without the Drush command or Module Builder needing their code to be updated:

Screenshot showing Drush UI

Screenshot showing Module Builder UI

Similarly, an even later version of Drupal Code Builder added code analysis to detect all top-level admin menu items, and so the admin settings form generation now lets you pick where to place the form in the menu.

It would be nice to think that a couple of buzzwords could gain Drupal Code Builder more attention and more users, but I fear that Drupal Console’s generator has got rather more traction in Drupal 8, despite its huge limitation. It’s disappointing that Drush has now added a third code generator to the mix, to even further dilute the ecosystem, and that it’s just as limited by hardcoding.

So where do we go from here?

Well, people get attached to UIs, but they don’t care much about what powers them, especially in this day and age of Composer, where adding dependencies no longer creates an imposition on the end-user.

So I suggest the following: the code analysis portion of Drupal Code Builder could be extracted to a new package. It doesn’t depend on anything on the code generation side, so that should be fairly simple. It provides an API that supports analysis being run in batches, which could maybe do with being spruced up, but it’s good enough for a 1.0.0 version for now.

Then, any code generating system would be able to use it. Both Console and Drush could replace their hardcoded lists of hooks and plugins with analytical data, while keeping the commands they expose to the end-user unchanged.

I’ll be at DrupalEurope in Darmstadt, where I’ll be running a BoF to discuss where we go next with code generation on Drupal.


Subscribe to Joachim's Drupal blog