Robot Librarian

[This is in response to a thread on the blacklight mailing list about getting MARC data into Solr.]

What’s the question?

The question came up, “How much time do we spend processing the MARC vs trying to push it into Solr?”. Bob Haschart found that even with a pretty damn complicated processing stage, pushing the data to solr was still, at best, taking at least as long as the processing stage.

I’m interested because I’ve been struggling to write a solrmarc-like system that runs under JRuby. Architecturally, the big difference between my stuff and solrmac is that I use the StreamingUpdateSolrServer (on Erik Hatcher’s suggestion). So I thought I’d check how things break down for me.

Here are my numbers running under JRuby (using MARC4J as the marc implementation) with the Solr StreamingUpdateSolrServer. Obviously, there are a lot of differences between this and solrmarc, but I’m hoping that while it’s not comparing apples to apples, it’s at least comparing apples to some sort of processed cheese-like product.

What work is being done on what?

The data set is a file of 18,881 MARC records in marc-binary format. It’s probably not big enough to get a great idea of how things will run over the long (many millions of records) haul, but it’ll do for this rough-cut stuff.

I break my processing down into five categories:

• Read the records into marc4j objects and do nothing. This is a baseline of sorts.
• The “normal” fields are anything that you could do with SolrMarc without a custom routine; the actual processing is done in JRuby.
• Custom fields are generated with JRuby code, but these are things that in solmarc would require a custom routine.
• The big “allfields” field is text from tags 100 through 900.
• The “to_xml” routine is just calling the underlying marc4j XML output and stuffing it into a string.

The schema used is our normal UMICH schema except for High Level Browse (which appear in the our catalog as “Academic Discipline”). The code for that is written in Java, and I just call it from JRuby when I’m using it. I excluded it because it’s incredibly expensive, both at startup time (when it loads a giant database of call-number ranges and associated categories) and for processing — there’s a lot of call-number normalization, long-string comparisons, some modified binary searches, etc. etc. etc. It’s expensive. Trust me.

The Solr server itself is on a different, incredibly-beefy machine, and is emptied out before each invocation that involves actually pushing data to it (with a delete-by-query :).

How fast were things on my desktop?

• 18,881 records in marc-binary format
• Times are in seconds, run on my desktop
• Remember, you can’t compare these numbers to Bob’s because we’re doing different things to different data.

Total Seconds	Description
19	Just read the records with marc4j and do nothing.
85	Read and do 35 “normal” fields (no custom)
104	Read, 35 normal, 15 custom fields
110	Read, normal, custom, allfields
129	Read, normal, custom, allfields, to_xml
136	Read, normal, custom, allfields, to_xml, 2-threaded SUSS, commit every 5K docs
142	Read, normal, custom, allfields, to_xml, 1-threaded SUSS, commit every 5k docs
124	Read, normal, custom, allfields, to_xmx, 1-threaded SUSS, commit every 5k docs, 2 threads doing processing

We can also break the same numbers down as:

Seconds	Description
19	read the records and do nothing
66	process the 35 normal fields
19	process the 15 custom fields
6	generate the “allfields” field
19	generate the XML (yowza!)
7	send to solr with two threads
13	send to solr with one thread

Or like this:

Seconds	Description
129	do all the reading and processing
13	send to solr with one thread

Why does solr processing seem so much faster for me?

There are a lot of reasons why my submit-to-solr might seem like less of a burden. The ones I can think of off the top of my head are:

• SUSS is just faster than whatever solrmarc does.
• My processing stage is so much slower than solrmac’s (due to algorithms or jruby-vs-java, I don’t know) that the “push to solr” portion of it gets swallowed up by the slowness of the of overall code.
• The Solr server is so much faster than my desktop that my poor little desktop can’t send it data fast enough to work it.

For my setup, obviously adding a processing thread is a lot more beneficial than adding a SUSS thread. My desktop doesn’t have that many threads lying around (adding a third processing thread actually slowed things down), so I moved the code to a beefier machine to see what happened.

Trying the same thing on a beefy machine

This is the exact same code and data, but on a beefy machine (16 cores, gobs of memory).

time	SUSS Threads	Processing Threads
70	1	1 (was 142 seconds on the desktop)
47	1	2
39	1	3
35	1	4
68	2	1
48	2	2
38	2	3
34	2	4

So, on my hardware anyway, there’s a sweet spot with one suss thread and three processing threads. YMMV, of course.

What have we learned?

I’m not sure, to be honest. It’s logistically difficult for me to do the same process in solrmarc because I’d have to rebuild everything without the HLB stuff. I guess for me, what I’ve learned that if I’m going to continue working on my code, the places to focus my attention are threading (obviously) and MARC-XML generation.

For reasons I’m still not entirely clear on (I wasn’t there), the Code4Lib 2010 conference this week inspired renewed interest in a JSON-based format for MARC data.

When I initially looked at MARC-HASH almost a year ago, I was mostly looking for something that wasn’t such a pain in the butt to work with, something that would marshall into multiple formats easily and would simply and easily round-trip.

Now, though, a lot of us are looking for a MARC format that (a) doesn’t suffer from the length limitations of binary MARC, but (b) is less painful (both in code and processing time) than MARC-XML, and it’s worth re-visiting.

For at least a few folks, un-marshaling time is a factor, since no matter what you’re doing, processing XML is slower than most other options. Much of that expense is due to functionality and data-safety features that just aren’t a big win with a brain-dead format like MARC, so it’s worth looking at alternatives.

What is MARC-HASH?

At some point, we’ll want a real spec, but right now it’s just this:

So, a short example:

How's the speed?

I think it's important to separate the format marc-hash from the eventual marshaling format -- partly because someone might actually want to use something other than JSON as a final format, but mostly because it forces implementations to separate hash-creation from json-creation and makes it easy to swap in different json libraries when a faster one comes along.

Having said that, in real life people are mostly concerned about JSON. So, let's look at JSON performance.

The MARC-Binary and MARC-XML files are normal files, as you'd expect. The JSON file is "Newline-Delimited JSON" -- a single JSON record on each line.

The benchmark code looks like this:

  # Unmarshal
  x.report("MARC Binary") do
    reader = MARC::Reader.new('test.mrc')
    reader.each do |r|
      title = r['245']['a']
    end
  end


# Marshal
  x.report("MARC Binary") do 
    reader = MARC::Reader.new('test.mrc')
    writer = MARC::Writer.new('benchout.mrc')
    reader.each do |r|
      writer.write(r)
    end
    writer.close
  end

Under MRI, I used the nokogiri XML parser and the yajl JSON gem. Under JRUby, it was the jstax XML parser and the json-jruby JSON gem.

The test file is a set of 18,831 records I've been using for all my benchmarking of late. It's nothing special; just a nice size.

Marshalling Speed (read from binary marc, dump to given format)

Times are in seconds on my Macbook laptop, using ruby-marc.

Format	Ruby 1.87	Ruby 1.9	JRuby 1.4	Jruby 1.4 --1.9
XML	393	443	188	356
MARC Binary	36	23	23	25
JSON/ NDJ	31	19	25	ERROR

Unmarshalling speed (from pre-created file)

Again, times are in seconds

Format	Ruby 1.87	Ruby 1.9	JRuby 1.4	Jruby 1.4 --1.9
XML	113	89	75	89
MARC Binary	29	16	16	19
JSON/ NDJ	17	9	13	16

And so...

I'm not sure what else to say. The format is totally brain-dead. It round-trips. It's fast enough. It has no length limitations. We define it to have to be UTF-8. The NDJ format is easy to understand and allows streaming of large document collections.

If folks are interested in implementing this across other libraries, that'd be great. Any thoughts?

[Note: in this post I'm just going to focus on the "get stuff into Solr" part. My normal focus -- MARC data -- will make an appearance in the next post when I talk about using this in addition to / instead of solrmarc.]

Working with Solr

I love me the Solr. I love everything about it except that the best way to interact with it is via Java. I don’t so much love me the java.

So…taking Erik Hatcher’s lead and advice, as I will do whenever he offers either, I wrote some code to work within JRuby to deal with Solr.

Getting the code

I’ve added the gems to gemcutter, if you want to play along at home:

• jruby_producer_consumer (github, rdoc.info) Ruby syntax for threaded operations under jruby
• jruby_streaming_update_solr_server (github, rdoc.info) Ruby syntax on top of the Java class of the same name
• marc4j4r (github, rdoc.info) Ruby syntax on top of the marc4j java library.

WARNING: None of these gems have a 1.0 version tag on them, and that means that the API may change a titch in the future. Also, the fact that they’re released as gems means that it’s easy to release gems, not that I’m not an idiot.

The basics: Using SolrInputDocument and StreamingUpdateSolrServer

OK, with the disclaimer out of the way, let’s look at some code.

Nothing really fancy in there — just a few things worth noting:

• An suss object will take a hash (again, anything that responds to #each_pair) or a SolrInputDoc
• You can use either strings or symbols to represent Solr field names
• Values can be either a single value, or an array of multiple values

And there are three ways to get data into a doc:

• Via << [field, value(s)] (additive)
• Via doc.merge! hash (additive)
• Via doc[field] = value (replaces)

Adding Threads

I also went down the garden path of threading things. There are an awful lot of operations that are not threadsafe (e.g., reading a line from a file) but once you’ve got a bunch of records to worth with, turning them into Solr documents is usually thread-safe.

My model is that there’s a producer (usually the method #each) from an underlying data object. A thread takes whatever that method yields and sticks the values into a java BlockingQueue awaiting consumption. You then use ProdcuerConsumer#threaded_each (or ProducerConsumer#threaded_each_with_index) to pull items out of the queue and do something useful with them.

I extracted stuff into a library (jruby_producer_consumer) for your viewing pleasure.

CONFUSION ALERT: It’s perhaps unfortunate that the object you send to ProducerConsumer.new(obj) must implement #each and that the ProducerConsumer method #threaded_each calls that underlying #each…well there’s a lot of #each’s floating around. Keep them straight.

So…let’s look at some code to work with consumer threads.

Again, not a lot happening here.

• The “producer” is always one thread, because so little is thread-safe at the ‘each’ level. In this case, there’s a single thread pulling data out of the file and turning it into MARC records, which are added to the internal BlockingQueue. I buffer 5 of these at a pop (the default) so the consumer threads don’t starve. I presume that producing items is cheaper than consuming them, or else this library won’t help you much.
• ProducerConsumer#threaded_each calls the #each method of the underlying object. You can substitute anything that yields, though, as in this example where I call #each_line instead of the default #each

• Keep track of your threads. In this last example, there is one thread getting MARC records and putting them into the PC buffer (no way to change that), three threads consuming those records and sticking them into the suss object, and another two pulling stuff out of the suss object and sending things to Sorl. And, of course, there’s other stuff running on the computer, too. Experiment and figure out what works best for your hardware.
• See the docs for how to mess with what goes into a ProducerConsumer object. It’s entirely possible to use, say, #each_slice. There’s also a convenience method #threaded_each_with_index, but it does not call the underlying #each_with_index, it produces its own index as things are read.

Feedback not only welcome but necessary!

I’ve done a lot of messing around with Ruby in the last 10 days or so, but I’m still basically converting from Perl in my head. Any comments, bugs reports, or whatnot are definitely welcome!

I’ve been looking at making a jruby-based solr indexer for MARC documents, and started off wanting to make sure I could determine if anything I did would be faster than our existing (solrmarc-based) setup.

On our ridiculously-awesome hardware, right now we’re doing about 300 records/second for short files and 250 records/second for a full (6.5 million record) index, giving us a 7-8 hour reindex.

I’ll just post the results without a lot of commentary. I warmed stuff up in all cases, and ran on my desktop (so I could compare to MRI ruby, which isn’t installed on the server) and on the server where we usually run these things.

• The machines are my desktop OSX machine and the beefy linux server where we usually do this stuff
• The platforms are jruby 1.4 –server and MRI ruby 1.87
• The libraries are marc4j and ruby-marc 0.3.3
• The parsers are
  • The standard binary parsers all around
  • A home-grown AlephSequential format reader for the ’seq’ type. AlephSequential is a MARC representation that uses one line for each field. We use it because it doesn’t have length limitations and, not surprisingly, Aleph can spit it out pretty quickly compared to MARC-XML.
  • Whatever marc4j uses internally for MARC-XML
  • ruby-marc’s ‘jstax’ xml parser under jruby (which I wrote and apparently needs some love, see below)
  • ruby-marc’s ‘libxml’ xml parser under MRI ruby
• Seconds is the average of two rounds, with measurements taken after a warmup run in each case.

The test files were 18,881 records in marc-xml, marc-binary, and AlephSequential formats.

MACHINE  PLATFORM LIBRARY     PARSER    SECONDS    REC/SECOND

desktop  jruby    marc4j      binary      4.06       4650
desktop  jruby    marc4j      xml         5.55       3401
desktop  jruby    ruby-marc   binary     17.35       1088
desktop  jruby    ruby-marc   jstax      80.11        236

desktop  ruby     ruby-marc   binary     33.54        562
desktop  ruby     ruby-marc   libxml     46.87        402

server   jruby    marc4j      binary      2.29       8245
server   jruby    marc4j      xml         3.36       5619
server   jruby    marc4j      AlephSeq    3.68       5130
server   jruby    ruby-marc   binary      9.93       1901
server   jruby    ruby-marc   jstax      44.56        424

The quick takeaways, with all the obvious caveats:

• jruby with ruby-marc is twice as fast at binary and twice as slow at xml compared with MRI
• marc4j is four times as fast for binary and about an order of magnitutde faster for xml compared with ruby-marc.
• The server is fast.

We know from previous experience that libxml is the fastest of the current MRI-based marc-xml readers and that jstax is the best of the current jruby-based marc-xml readers. And, finally, we know that many of us can’t use marc-binary format because our records are too big.

If I’m gonna use jruby (which I think I am due to wanting to use the StreamingUpdateSolrServer) I’m gonna need to use marc4j and just wrap it up in some nicer syntax.

I decided to see if I could get Blacklight working under JRuby, starting with running the test suite and working my way up from there.

There was much pain. Much, much pain. Exacerbated by my almost complete lack of knowledge about what I was doing.

This is the procedure I eventually arrived at — if there are places where I made trouble for myself, please let me know!

[And does anyone know how to get jruby's nokogiri to link to a different libxml and stop with the crappy libxml2-version error message every time I run it under OSX???]

Download jruby

Go to jruby.org and download a binary distribution. Extract the tar.gz (or zip or whatever)

I’ll put mine in ~/jruby. Or, at least that’s what I’ll tell you.

tar xzf jruby-1.4.tar.gz

To avoid confusion, let’s make jrake an alias for rake and add the jruby bin directory to the path

cd ~/jruby/bin
ln -s rake jrake
export PATH=`pwd`:$PATH

Download Blacklight

git clone git://github.com/projectblacklight/blacklight.git

Again, well say that I put this in ~/blacklight/

Muck with Blacklight dependencies

Edit the file init.rb to comment out references to libxml and ruby-xslt, as well as nokogiri. My understanding is that the first two are used, at this point, only for the EAD stuff. Both rely on libxml2 which is a C-extension and hence unavailable to JRuby.

Nokogiri gets pulled in during other installs and for some reason jrake will complain later on that it’s got a wrong version or something. So, we’ll just work without that particular net for now.

#### File ~/blacklight/init.rb
# config.gem 'libxml-ruby', :lib=>'libxml', :version=>'1.1.3'
# config.gem 'ruby-xslt', :lib=>'xml/xslt', :version=>'0.9.6'
# config.gem 'nokogiri', :version=>'1.3.3'

Do some initial installs

jgem install -v=2.3.4 rails 
jgem install activerecord-jdbc-adapter jdbc-sqlite3 
             activerecord-jdbcsqlite3-adapter ActiveRecord-JDBC 
jgem install rcov -s http://gemcutter.org --no-rdoc --no-ri
jrake
jrake gems:install

Edit the config/database.yml file

…to change the adapter to jdbcsqlite3 for development and testing.

Edit the databases.rake file

This one was harder to track down. The default rake task has hard-coded database names in the .rake file — jdbcsqlite3 isn’t included. I keep seeing things saying, “Oh, yeah, that’s been fixed…” but, well, it wasn’t for me. I had to do it by hand.

edit ~/jruby/lib/ruby/gems/1.8/gems/rails-2.3.4/lib/tasks/databases.rake

You need to find everywhere there’s a

when "sqlite", "sqlite3" # or when /^sqlite/ in one case

…and change it to

when "sqlite", "sqlite3", "jdbcsqlite3"

Repeat for other databases you want to use (e.g., mysql). For the moment, since I’m only worried about running jrake spec, that’s all I’m gonna do.

Try again

jrake
  Missing these required gems:
   mislav-hanna  = 0.1.11

OK. Not sure why that didn’t come in before. Go head and add it.

jgem install  mislav-hanna

Migrate the databases

jrake

The databases should migrate, and then it’ll poop out because Solr didn’t start.

Fire up solr

Since we’re running jruby, accessing the shell doesn’t work. You’ll have to fire up your test solr instance by hand.

cd ~/blacklight/jetty
java -Djetty.port=8888 -jar start.jar 2>log.jetty

Try it again!

cd ~/blacklight
jrake spec

   ................................................................
   ................................................................
   ....F............................................................
   1)
   'ApplicationHelper Export EndNote should render the correct 
   EndNote text file' FAILED
   expected: "%0 Format\n%E Greer, Lowell. \n%E Lubin, Steven. \n%E Chase, Stephanie, \n%E Brahms, Johannes, \n%E Beethoven, Ludwig van, \n%E Krufft, Nikolaus von, \n%T Music for horn \n%I Harmonia Mundi USA, \n%C [United States] : \n%D p2001. \n",
  got: "%0 Format\n%C [United States] : \n%D p2001. \n%E Greer, Lowell. \n%E Lubin, Steven. \n%E Chase, Stephanie, \n%E Brahms, Johannes, \n%E Beethoven, Ludwig van, \n%E Krufft, Nikolaus von, \n%I Harmonia Mundi USA, \n%T Music for horn \n" (using ##)
./spec/helpers/application_helper_spec.rb:128:

Finished in 15.519 seconds
193 examples, 1 failure

I can live with that for the moment. Anyone know why that spec fails?

Great! How about the features?

jrake features
  (much output)

  59 scenarios (59 passed)
  434 steps (434 passed)
  0m51.186s

And so…

…it appears that, at least on the surface, jruby is a viable platform for Blacklight so long as I don’t actually need any of the libxml stuff. In the next couple days I’ll try and actually get it all up and running and see if I can break it.