INTRODUCING: ZOHMG.
-------------------

Zohmg is a data store for aggregation of multi-dimensional time series data. It
is built on top of Hadoop, Dumbo and HBase. The core idea is to pre-compute
aggregates and store them in a read-efficient manner – Zohmg is
wasteful with storage in order to answer queries faster.

This README assumes a working installation of zohmg. Please see INSTALL for
installation instructions.

Zohmg is alpha software. Be gentle.


CONTACT
-------

IRC:           #zohmg at freenode
Code:          http://github.com/zohmg/zohmg/tree/master
User list:     http://groups.google.com/group/zohmg-user
Dev list:      http://groups.google.com/group/zohmg-dev


RUN THE ZOHMG
-------------

Zohmg is installed as a command line script. Try it out:

  $> zohmg help


AN EXAMPLE: TELEVISION!
-----------------------

Imagine that you are the director of operations of a web-based
television channel. Your company streams videos, called "clips".

Every time a clip is streamed across the intertubes, your software
makes a note of it in a log file. The log contains information about
the clip (id, length, producer) and about the user watching it
(country of origin, player).

Your objective is to analyze the logs to find out how many people
watch a certain clip over time, broken down by country and player,
etc.

The rest of this text will show you by example how to make sense of
logs using Zohmg.


THE ANATOMY OF YOUR DATA SET
----------------------------

Each line of the logfile has the following space-delimited format:
 
timestamp clipid producerid length usercountry player love

Examples:
1245775664 1440 201 722 GB VLC 0
1245775680 1394 710 2512 GB VLC 1
1245776010 1440 201 722 DE QUICKTIME 0

The timestamp is a UNIX timestamp, the length is counted in seconds
and the id's are all integers. Usercountry is a two-letter ISO
standard code while the player is an arbitrary string.

The last field -- "love" -- is a boolean that indicates whether the user
clicked the heart shaped icon, meaning that she was truly moved by the
clip.


DO YOU SPEAK ZOHMG?
-------------------

In the parlance of Zohmg, clip and producer as well as country and
player are dimensions. 1440, 'GB', 'VLC', etc, are called attributes of
those dimensions.

The length of a clip, whether it was loved, etc, are called
measurements. In the configuration, which we will take a look at
shortly, we define the units of measurements. Measurements must be
integers so that they can be summed.

Simple enough!


CREATE YOUR FIRST 'PROJECT'
---------------------------

Every Zohmg project lives in its own directory. A project is created
like so:

  $> zohmg create television

This creates a project directory named 'television'. Its contents are:

 config     - environment and dataset configuration.
 lib        - eggs or jars that will be automatically included in job jar.
 mappers    - mapreduce mappers (you will write these!)


CONFIGURE YOUR 'PROJECT'
------------------------

The next step is to configure environment.py and dataset.yaml.


config/environment.py:

Define HADOOP_HOME and set the paths for all three jars (hadoop, hadoop-streaming,
hbase). You might need to run 'ant package' in $HADOOP_HOME to have
the streaming jar built for you.


config/dataset.yaml:

Defining your data set means defining dimensions, projections and
units.

The dimensions lets Zohmg know what your data looks like while the
projections hints at what queries you will want to ask. Once you get
more comfortable with Zohmg you will want to optimize your projections
for efficiency.

For example, if you wish to find out how many times a certain clip has
been watched broken down by country, you would want to set up a
projection where clip is the first dimension and country is the second
one.

For the television station, something like the following will do just
fine.


## config/dataset.yaml

dataset: television

dimensions:
 - clip
 - producer
 - country
 - player

projections:
 - clip
 - player
 - country
 - clip-country
 - producer-country
 - producer-clip-player

units:
 - plays
 - loves
 - seconds


After you've edited environment.py and dataset.yaml:

  $> zohmg setup

This command creates an HBase table with the same name as your dataset.

Verify that the table was created:

  $> hbase shell
  hbase(main):001:0> list
  television
  1 row(s) in 0.1337 seconds

Brilliant!


DATA IMPORT
-----------

After the project is created and setup correctly it is time to import some
data.

Data import is a process in two steps: write a map function that analyzes
the data line for line, and run that function over the data. The data
is normally stored on HDFS, but it is also possible to run on local data.


WRITE A  MAPPER
----------------

First we'll write the mapper. It will have this signature:

  def map(key, value)

The 'key' argument defines the line number of the input data and is
usually (but not always!) not very interesting. The 'value' argument is a
string - it represents a single line of input.

Analyzing a single line of input is straightforward: split the line
on spaces and collect the debris.

## mappers/mapper.py

import time

def map(key, value):
    # split on space; make sure there are 7 parts.
    parts = value.split(' ')
    if len(parts) < 7: return

    # extract values.
    epoch = parts[0]
    clipid, producerid, length = parts[1:4]
    country, player, love      = parts[4:7]

    # format timestamp as yyyymmdd.
    ymd = "%d%02d%02d" % time.localtime(float(epoch))[0:3]

    # dimension attributes are strings.
    dimensions = {}
    dimensions['clip']     = str(clipid)
    dimensions['producer'] = str(producerid)
    dimensions['country']  = country
    dimensions['player']   = player

    # measurements are integers.
    measurements = {}
    measurements['plays']   = 1
    measurements['seconds'] = int(length)
    measurements['loves']   = int(love)

    yield ymd, dimensions, measurements


The output of the mapper is a three-tuple: the first element is a
string of format yyyymmdd (i.e. "20090601") and the other two elements
are dictionaries.

The mapper's output is fed to a reducer that sums the values of the
units and passes the data on to the underlying data store.



RUN THE MAPPER
--------------

Populate a file with a small data sample:

  $> cat > data/short.log
  1245775664 1440 201 722 GB VLC 0
  1245775680 1394 710 2512 GB VLC 1
  1245776010 1440 201 722 DE QUICKTIME 0
  ^D

Perform a local test-run:

  $> zohmg import mappers/mapper.py data/short.log --local

The first argument to import is the path to the python file containing the map
function, the second is a path on the local file system.

The local run will direct its output to a file instead of writing to
the data store. Inspect it like so:

  $> cat /tmp/zohmg-output
  'clip-1440-country-DE-20090623' '{"unit:length": {"value": 722}}'
  'clip-1440-country-DE-20090623' '{"unit:plays": {"value": 1}}'
  'clip-all-country-DE-20090623'  '{"unit:length": {"value": 722}}'
  [..]


If Hadoop and HBase are up, run the mapper on the cluster:

  $> zohmg import mappers/mapper.py /data/television/20090620.log

Assuming the mapper finished successfully there is now some data in
the HBase table. Verify this by firing up the HBase shell:

  $> hbase shell
  hbase(main):001:0> scan 'television'
  [..]

Lots of data scrolling by? Good! (Interrupt with CTRL-C at your leisure.)


SERVE DATA
----------

Start the Zohmg server:

  $> zohmg server

The Zohmg server listens on port 8086 at localhost by default. Browse
http://localhost:8086/ and have a look!



THE API
-------

Zohmg's data server exposes the data store through an HTTP API. Every
request returns a JSON-formatted string.

The JSON looks something like this:

[{"20090601": {"DE": 270, "US": 21, "SE": 5547}}, {"20090602": {"DE":
9020, "US": 109, "SE": 11497}}, {"20090603": {"DE": 10091, "US": 186,
"SE": 8863}}]


The API is extremely simple: it supports a single type of GET request
and there is no authentication.

There are four required parameters: t0, t1, unit and d0.

The parameters t0 and t1 define the time span. They are strings of the
format "yyyymmdd", i.e. "t0=20090601&t1=20090630".

The unit parameter defines the one unit for which you query, for
example 'unit=plays'.

The parameter d0 defines the base dimension, for example 'd0=country'.

Values for d0 can be defined by setting the parameter d0v to a
comma-separated string of values, for example "d0v=US,SE,DE". If d0v is
empty, all values of the base dimension will be returned.


A typical query string looks like this:

 http://localhost:8086/?t0=20090601&t1=20090630&unit=plays&d0=country&d0v=DE,SE,US

This example query would return the number of clips streamed for the
time span between the first and last of June broken down by the
countries Sweden, Germany and the United States.


The API supports JSONP via the jsonp parameter. By setting this
parameter, the returned JSON is wrapped in a function call.



PLOTTING THE DATA
-----------------

There is an example client bundled with Zohmg. It is served by the Zohmg server
at http://localhost:8086/graph/ and is quite useful for exploring your dataset.

You may also want to peek at the javascript code to gain some inspiration and
insight into how the beast works.

The graphing tool uses Javascript to query the data server, and plots
graphs with the help of Google Charts.



KNOWN ISSUES
------------

Zohmg is alpha software and is still learning how to behave properly in mixed
company. You will spot more than a few minor quirks while using it and might
perhaps even run into the odd show-stopper. If you do, please let us know!

Zohmg currently only supports mappers written in Python. Eventually, you will
also be able to write mappers in Java.