I’ve released Whistlepig 0.10. This release changes the index format in a backwards-incompatible way. Indexes created by pre-0.10 Whistlepigs will not be readable by Whistlepig 0.10.

The good news going is that I have also added proper versioning on the index formats now, and so providing upgrade paths will be possible in the future, should I find it necessary to introduce further backwards-incompatible changes.

This change was necessary primarily for multi-process support. As of Whistlepig 0.10, you can have multiple processes reading from, and writing to, the same index. This means that Whistlepig may finally be useful for e.g., Rails apps, where you have multiple server processes all reading from the same index.

Whistlepig provides this multi-process concurrency by using pthread read-write locks. From a performance perspective, locking is not a very good form of concurrency. The original plan, way back when I was designing Whistlepig, was to use lock-free algorithms to allow concurrent access (at least, single-writer and multi-reader) and remain performant. Unfortunately, the C memory model isn’t as robust as the JVM memory model and, as far as I can tell, doesn’t support the atomicity guarantees necessary to accomplish this (see e.g. this Quora question and my follow-up comment). FWIW, it look like the C11 standard may address these issues (Jens Gustedt has more on this) but that’s not going to help me today.

So, locking it it is, for now. The read-write locking mechanism at least allows multiple reader processes to access the same segment at the same time, but a single writer will block any other writers and any other readers. The implication is that if the proportion of your write traffic is high relative to your read traffic, you will waste a lot of time waiting to acquire locks. So if you need high-throughput writes from Whistlepig, you probably don’t want to glom it all together into a single index, and instead should shard documents across multiple indexes. High-throughput reads with occasional writes should be fine.

The benefit that locking does provide is that it frees me up to implement more sophisticated data structures for postings lists. Right now Whistlepig essentially builds giant linked lists of all postings, which is very naive and wasteful, but does lend itself to a lock-free approach. If I am guaranteed exclusive write access, I can be much smarter about how I represent the postings. The learning in traditional IR applications is that smaller postings lists can improve search time as well. I have some concrete ideas on how to do this that should find themselves into future Whistlepig releases.

I’ve released version 0.9 of Whistlepig, my minimalist real-time full-text search engine. This is a beta release. There are no known bugs, but I wouldn’t use it to power the space shuttle just yet.

Changes since the previous release include:

• A new Query#term_map method, which allows you to programmatically alter queries after they’ve been parsed. The most obvious use case for this is to do case-folding, which can be tricky to do on the unparsed string due to the “OR” operator.
• Various bugfixes and some additional safety-check code.

Do a gem install whistlepig to get it, or download the tarball here.

Whistlepig 0.2 is out already. This time it should actually compile under OS X. I’m having a grand old time figuring out all the differences in flags that Ruby when compiling gems under different architectures. But not so grand that I want to learn automake/autoconf.

As part of making sure things were compiling on different platforms, I ran make test-integration and noticed that on my dual-boot laptop runs it at 8000k/s on Linux but only 6200k/s on OS X. I was expecting a difference in Linux’s favor, but not quite that large! As another point of reference, my mid-range Linux desktop gives me 9500 k/s.

Today I released the very first version of Whistlepig, a minimalist realtime full-text search index.

Side projects apparently take a lot longer when you have a job and a baby, because it’s taken me over 6 months to get to the point where I have something releasable. And there are so many obvious improvements to make. But all known bugs are squashed, and it’s good enough to use, so, it’s out.

The README has a good description of what Whistlepig is, so here I thought I’d talk about the why. Why write yet another inverted index?

The unfortunate fact is that you have too many choices already: Lucene, obviously, and its derivatives like SOLR, and if you’re shy of the JVM, Xapian and Sphinx. Ferret used to be a good choice in the Ruby world until Dave Balmain absconded and no one had the cojones to maintain his code. I’ve used each of these things.

But they are all very heavy-weight solutions, and they all suffer from what I call the “TREC mentality”. In early TREC competitions, you were given a big, static corpus, which you indexed at your leisure, and then you were given a bunch of queries, which were all long descriptions of what documents someone was interested in. It would be something like “I am interested in documents about Mayan architecture, but only during the pre-conquistador period, and specifically I am not interested in such and such” and so on. These competitions were great in that they spurred advances in search engineering, but the result is that almost every inverted index implementation today is optimized for precisely the case of static corpora and large queries.

In the intervening 30 years, the use case for full-text search has far exceeded the library-science-style applications of the early TREC competitions. There are many applications where you don’t need tf-idf scores and the Okapi formula or even necessarily stemming. You just want recent things that match your query, and you value control and transparency over some kind of fuzzy natural language matching. Search in GMail (or Sup of, course!) comes to mind, or searching within the posts on this blog.

That’s one part of the reason why existing solutions are not ideal. The other part is that inverted indexes are so optimized for speed and for size that even little things like wanting documents from last to first can be drastically slower than using the standard ordering. For example, Sup wants documents in reverse chronological order; Xapian is fastest in increasing docid order; so we play crazy games to map dates to docids:

DOCID_SCALE = 2.0**32
TIME_SCALE = 2.0**27
MIDDLE_DATE = Time.gm(2011)
def assign_docid m, truncated_date
  t = (truncated_date.to_i - MIDDLE_DATE.to_i).to_f
  docid = (DOCID_SCALE - DOCID_SCALE /
    (Math::E**(-(t/TIME_SCALE)) + 1)).to_i
  while docid > 0 and docid_exists? docid
    docid -= 1
  end
  docid > 0 ? docid : nil
end

This snippet is courtesy of Rich Lane, who should be credited in history books as the first person to find a use for a logistic curve in an email client.

If you try and use something like Xapian or Sphinx for these applications, you have to play games like that for performance. And when new documents arrive, you have to play further games to get them into the index sooner rather than later. And all the while you’re turning off 90% of the features anyways.

So that leads us to the world of realtime search, which explicitly values recent documents over older ones. It’s “realtime” where new documents arrive on the fly and must be made available to queries as soon as they arrive. If you’re in that situation, you typically also care more about more recent documents that older ones anyways. Those are the two tenets of realtime search: documents are available immediately, and recent documents are more important than older ones.

Whistlepig is my attempt to capture those two tenets in as few lines of code as possible, while still being reasonable performant. I do this by stripping away all the vestigial TREC functionality of relevance, ranking, sorting, tf-idf, etc. You get documents in LIFO order, and that’s it. Whistlepig doesn’t return anything besides the docid either: if you need something more than the id, you have to fit that into a separate store somewhere. It turns out if you throw that stuff away, you can accomplish the rest of the search problem without a tremendous amount of code. Like any C program, it’s 5% algorithm and 95% bookkeeping.

There is one wrinkle that I actually add to the model: I allow adding and removing labels from documents. Every other aspect of a document is fixed in Whistlepig—you can’t even delete it from the index once it’s been added—but labels are mutable. And of course you can intermingle labels with the other components of your query. Almost every realtime search application I can dream up would benefit from this functionality, so there you go.

My hope for Whistlepig is that it becomes the default choice for realtime search applications, especially in the Ruby world, which hasn’t had a good in-process search solution since Ferret bit the dust. And if I mysteriously disappear like Dave did, I also hope that the codebase is small enough and simple enough that taking it over doesn’t seem like a herculean effort.