kba.pipeline is a document processing pipeline that assembles streamcorpus objects from raw data sets for us in TREC KBA.

1006305073 all-stream-ids.suc.txt 884018982 all-stream-ids.doc_ids.suc.txt

The kba.pipeline python module contains tools for processing streamcorpus.StreamItems stored in Chunks. It includes transform functions for getting clean_html, clean_visible text, creating labels from hyperlinks to particular sites (e.g. Wikipedia), and taggers like LingPipe and Stanford CoreNLP, that make Tokens and Sentences.

To create a python2.7 virtualenv, do this:

wget http://www.python.org/ftp/python/2.7.3/Python-2.7.3.tgz tar xzf Python-2.7.3.tgz cd Python-2.7 ./configure --prefix /data/trec-kba/installs/py27 make install cd .. wget http://pypi.python.org/packages/source/v/virtualenv/virtualenv-1.8.4.tar.gz#md5=1c7e56a7f895b2e71558f96e365ee7a7 tar xzf virtualenv-1.8.4.tar.gz cd virtualenv-1.8.4/ /data/trec-kba/installs/py27/bin/python setup.py install cd .. /data/trec-kba/installs/py27/bin/virtualenv --distribute -p /data/trec-kba/installs/py27/bin/python py27env

Easiest to put this entire repo at a path like

/data/trec-kba/installs/trec-kba-data

which is hardcoded into these three files:

scripts/spinn3r-transform.sh scripts/spinn3r-transform.submit configs/spinn3r-transform.yaml

Then, you need these two other directories:

/data/trec-kba/keys ---- from the trec-kba-secret-keys.tar.gz that is in the Dropbox /data/trec-kba/third/lingpipe-4.1.0 --- also in the dropbox

As a test run this:

## first go inside the virtualenv source /data/trec-kba/installs/py27env/bin/activate

## install all the python libraries make install

## run a simple test make john-smith-simple

and if that works, then try

make john-smith

To try doing the real pull/push from AWS, you can put the input paths here:

/data/trec-kba/installs/trec-kba-data/spinn3r-transform zcat spinn3r-transform-input-paths.txt.gz | split -l 150 -a 4 b=0; for a in ls ?????; do mv $a input.$b; let b=$b+1; done;

and then locally as a test:

cat /data/trec-kba/installs/trec-kba-data/spinn3r-transform/input.0 | python -m kba.pipeline.run configs/spinn3r-transform.yaml

and then, after seeing that work edit the submit script to have as many jobs as their are input files:

condor_submit scripts/spinn3r-transform.submit

There is one key problem with this, which we discussed on the phone: when the job dies, it starts over on the input list. Let's discuss using the zookeeper "task_queue" stage.

To use the zookeeper task queue, you must install zookeeper on a computer that your cluster can access. Here is an example zookeeper config:

# The number of milliseconds of each tick tickTime=10000 # The number of ticks that the initial # synchronization phase can take initLimit=10 # The number of ticks that can pass between # sending a request and getting an acknowledgement syncLimit=5 # the directory where the snapshot is stored. dataDir=/var/zookeeper # the port at which the clients will connect clientPort=2181 server.0=localhost:2888:3888 maxClientCnxns=2000

Note the large maxClientCnxns for running with many nodes in condor, and also not the 10sec tickTime, which is needed to avoid frequent session timeouts from condor slots that are working hard.

To make a job run off the zookeeper task queue, make these changes:

configs/spinn3r-transform.yaml:

• task_queue: stdin
• #task_queue: stdin
• task_queue: zookeeper
• zookeeper:
• namespace: spinn3r-transform
• zookeeper_address: mitas-2.csail.mit.edu:2181

scripts/spinn3r-transform.submit:

-Input = /data/trec-kba/spinn3r-transform/input.$(PROCESS) + +## disable stdin because we are using task_queue: zookeeper +#Input = /data/trec-kba/spinn3r-transform/input.$(PROCESS)

Important: Also update the number of jobs at the end of the .submit file.

and then do these steps on the command line:

## see the help text python -m kba.pipeline.load configs/spinn3r-transform.yaml -h

## load the data python -m kba.pipeline.load configs/spinn3r-transform.yaml --load spinn3r-transform-input-paths.txt

## check the counts -- might take a bit to run, so background and come back to it python -m kba.pipeline.load configs/spinn3r-transform.yaml --counts >& counts &

## launch the jobs condor_submit scripts/spinn3r-transform.submit

## watch the logs for the jobs tail -f ../spinn3r-transform/{err,out}*

Periodically check the --counts on the queue and see how fast it is going. Do we need to turn off the lingpipe stage?