Stayin Alive with Nokia

Me: What if I told you I was going to implement Google's MapReduce white paper?

You: Erik, that's a gargantuan effort! And why would you do that, when Hadoop already exists?

Me: Now what if I said I could do it in less than 20,000 lines of code?

You: Preposterous!

I'd have thought so too, but Nokia Labs has already done it: Meet the Disco Project. This project implements a mapreduce layer on top of a distributed filesystem, with a nice web admin interface. The package is lean, easy to install, and easy to use. I don't see myself ever launching Hadoop again (outside of work). Nokia did it, indeed they could only have done it, using my two favorite languages: Erlang and Python.

I've been playing with it all weekend. The install is like butter, basically a git clone, a make, and then start the service (_one_ service, not Hadoop's five). There are surprisingly few dependencies. The Erlang side has only Lager for logging and Mochiweb.

I executed the staple word count and grep examples, but wanted to do more. For a test project, I decided I could run queries on my ledger data. Note that I'm interested right now in testing ease of use, not scalability, and so my setup is single-node. Also, my ledger data is hardly a large data set, but it is more than two complete years of transactions, so there are bound to be some interesting queries.

Here's the interface for working with the filesystem:

gentoo01 deps # ddfs -h
        ddfs attrs tag
        ddfs blobs [tag ...]
        ddfs cat [url ...]
        ddfs chtok tag token
        ddfs chunk tag [url ...]
        ddfs cp source_tag target_tag
        ddfs delattr tag attr
        ddfs exists tag
        ddfs find [tag ...]
        ddfs get tag
        ddfs getattr tag attr
        ddfs ls [prefix ...]
        ddfs push tag [file ...]
        ddfs put tag [url ...]
        ddfs rm [tag ...]
        ddfs setattr tag attr val
        ddfs stat [tag ...]
        ddfs tag tag [url ...]
        ddfs touch [tag ...]
        ddfs urls [tag ...]
        ddfs xcat [urls ...]

  -h, --help            display command help
  -s SETTINGS, --settings=SETTINGS
                        use settings file settings
  -v, --verbose         print debugging messages
  -t TOKEN, --token=TOKEN
                        authorization token to use for tags

Loading the raw ledger file in directly is just ddfs push ledger ledger. The first 'ledger' is the tag and the second is the source filename - ddfs references files by tags, rather than the traditional hierarchical directory structure. Also, if you're keeping score against Hadoop, note there is no access control; if you can run ddfs, then you can alter or destroy its entire contents. Also, the web interface (dynamic via jQuery) is concerned only with mapreduce, and so filesystem operations are command-line only.

The main cluster status page is fairly spartan. The eight small lavendar boxes are the eight workers I've configured (for an eight-core machine):

Ledger data is stored as multiline entries:

2011/08/04 Dairy Queen
  Assets:Texans:Checking                      -22.81

2011/08/04 Northside Market
  Assets:Texans:Checking                      -28.39

which I'd first like to denormalize into one-record-per-entry-leg format to make future queries easier:

219bb456-5ef2-54e2-bc92-e0b4f25ff39e    ('2011/08/04', 'Dairy Queen', 'Assets:Texans:Checking', -22.81, 'defcur')
219bb456-5ef2-54e2-bc92-e0b4f25ff39e    ('2011/08/04', 'Dairy Queen', 'Expenses:Dining/Entertainment', 22.81, 'defcur')
3361d9ef-f865-5e06-adec-13fadbc6072c    ('2011/08/04', 'Northside Market', 'Assets:Texans:Checking', -28.39, 'defcur')
3361d9ef-f865-5e06-adec-13fadbc6072c    ('2011/08/04', 'Northside Market', 'Expenses:Gas', 28.39, 'defcur')

Already I'm dreading some upcoming pain. The multiline input format requires special handling of record boundaries, not once but twice: once during input to the filesystem, and then again when mapping. For the first problem, Disco provides a ddfs chunk which can import files to the filesystem in a content-aware way (I skipped this, knowing my input file is smaller than the chunk size). Hadoop solves this differently, by ignoring the issue until the map job starts, then making an extra round-trip to glue together a record that has been split across nodes (informed by custom InputFormat and InputSplit classes). This feels reactive and unnecessarily elaborate.

For the problem of running a 'map' on multiline records, I was pleased at how easy Disco makes this. Here's an abridged snippet from Disco's grep example (jobs are defined with Python):

    def map_reader(fd, size, url, params):
        for line in fd:
            if pattern.match(line):
                yield url, line

Note that the map_reader takes in an I/O stream, and I can read as many lines as I like to yield a map result. Super easy!

In the end, is 110 lines of string processing, a map-only job, to transform raw ledger input to a denormalized flat file as shown above. Most of the code is handling multiple currencies or populating the blank implicit amount value - actual domain problems, not boilerplate. The code is here

Running the job:

gentoo01 jobcode # disco run ledger.Ledger tag://ledger

...which results in this status page:

To view the output, which has been stored in ddfs, I xcat the result file into a pager:

gentoo01 jobcode # ddfs xcat dir://localhost/disco/localhost/5f/Ledger@568:b59e3:b6fb2/.disco/map_shuffle-1-1384743909052908.results | less

xcat is the compression-aware version cat command; ddfs compresses files by default, a feature missing in early Hadoop versions.

By the way, here is the full command list for administering the mapreduce layer:

gentoo01 jobcode # disco -h
        disco client_version 
        disco config 
        disco debug [host]
        disco deref [url ...]
        disco events jobname
        disco job worker [input ...]
        disco jobdict jobname
        disco jobs 
        disco kill jobname ...
        disco mapresults jobname
        disco master_version 
        disco nodaemon 
        disco nodeinfo 
        disco oob jobname
        disco oob get key jobname
        disco pstats jobname
        disco purge jobname ...
        disco restart 
        disco results jobname
        disco run jobclass [input ...]
        disco start 
        disco status 
        disco stop 
        disco submit [file]
        disco test [testname ...]
        disco wait jobname

  -h, --help            display command help
  -s SETTINGS, --settings=SETTINGS
                        use settings file settings
  -v, --verbose         print debugging messages
  -t TOKEN, --token=TOKEN
                        authorization token to use for tags

It's getting late, so I'm going to save further job development for another day. I'm jazzed about it, though, because of how easy I know it will be. Some trivial queries: How much do I spend on dining in a month? How has mortgage principle accelerated downwards? Has my tax withholding corresponded to income?

Consider deploying Disco for your mapreduce needs, assuming your needs include easy to use and elegant.

11/17/2013 by stasibear Permalink