rjbs forgot what he was saying

not logged in (root) | by date | tagcloud | help | login

an overview of the metabase

by rjbs, created 2009-04-05 23:05
last modified 2009-04-06 11:15
tagged with: @markup:md journal metabase perl programming

At both the Oslo and Birmingham QA Hackathons, my big project was the CPAN Metabase. I'm hoping that in 2010, I'll be working on something new, but only because I like variety -- I really like the product we've got and I believe it can be a great tool for a lot of problems. Here's a simple overview of what it is and how we plan to use it.

Metabase is a system for storing and searching structured annotations about resources that you don't (necessarily) control. Its first job is to store the next generation of CPAN Testers reports.

Whenever you upload a distribution to the CPAN, a bunch of other servers download it and try to install its prereqs and run its test suite. When they're done, these servers (or sometimes end users) submit reports that say things like "all tests successful" or "tests failed." They also include some other information that humans can look at to decide what really happened. This has been an extremely useful system for me, as it makes it very easy to see, in effect, well-filed bug reports that required no human effort to generate. (There is a constant debate about whether these reports are useful. To me, they are. YMMV.)

Rather than just send in a dump of what happened when installation was attemped, though, we wanted to get a bunch of structured data like:

All these facts about a test run are distinct items that can be submitted together as a report about an installation attempt. We break down this sort of submission into a few parts:

  1. the content -- a list of these facts
  2. the resource -- the distribution file from the CPAN that was being tested
  3. the metadata -- data we can summarize for indexing and searching

As more and more reports in this format are submitted, we can easily say "show me all new reports for my dist RJBS/URI-cpan-1.003.tar.gz with the metadata test-result set to FAIL."

There's more to it, though. The searchable metadata is drawn from the content, the resource, and from the submitting user. This means we can say "show me all reports for any dist file where I was the uploader, where result was FAIL, where the OS was VMS, and where the perl version was at least 5.008."

Then we can extract just the TAP from the report and load it into our sweet GUI visualizer, or we could just look at a summary of users submitting those reports, or do lots of other things with the data.

Now, imagine that one of the facts reported by testers is "What prerequisites did I think your module had?" This is a really interesting question when your distribution dynamically determines prereqs based on the installer. It's also useful if you find out that different install tools are computing things differently. We might give this fact a name of CPAN-Metabase-Dist-Prereqs.

A great consequence of the structuring of our test reports is that the units of contained data can be re-used. That means that if CPANTS also wants to submit these facts, it can. So could other pieces of software that look at CPAN distributions and guess what they will require for installation -- Debian's packaging tools, for example. If these facts were submitted, it would be easy to notice that all tools but one agree on the prereqs for a given distribution, exposing a likely bug somewhere.

So, the use of structured reports is clearly a big win, and indexed metadata largely follows from that. If we know how to provide some indexable data for a fact, it's easier to find facts.

The next likely win is in the way that we associate test reports with distribution files. Every fact is associated with one resource. We use an identifier for our resources that is universal. We call it a URI. For example, if you wanted to submit a report about Config::INI's latest release, you would submit a fact about cpan:///distfile/RJBS/Config-INI-0.014.tar.gz

By using URIs for resources, the metabase is wide open for expansion. You could use it as an address book by submitting vCard facts about mailto URIs. You could submit ratings by submitting a fact about a distribution. There is nothing CPAN-specific in the metabase, even though it was designed first and foremost to be a repository for CPAN testers data. It just worked out that by keeping it simple, we'd made it very generic.

Resources also provide some metadata, so you can easily get a list of all facts about distfile resources where the uploader was DAGOLDEN.

Finally, the Metabase code is free and easy to install. If you want, you can run your own metabase for anything you like. One obvious use is to run your own test report collector. You could have your company's internal deployment system use all the standard CPAN tools -- from CPAN.pm to the same web interface -- for testing its internal code. You'd just submit test reports to your internal server. You could even easily relay reports for non-private code to the public repository. I'm really hoping to eventually do some work to make Smolder cooperate with Metabase, so test reports will contain TAP archives and you can tell Smolder to get its data from a Metabase server.

There's lots more I could say, but I've probably rambled a lot, already. I'm really tired.

There aren't too many major hurdles between the current state of code and getting real data into the system. I'm hoping that we'll have a "stock" testing client submitting data to a test installation within the week -- David suggested he might have it working tonight. After that, it will all be about testing, polish, and scaling.

Well, and a bunch of other problems... but that's okay. It's really coming along nicely.