GNIP STREAM - COLLECTOR METRICS

Command line tool for:

 * Saving JSON stream of activities to disk in date-ordered directory tree
 * See real-time rule production
 * See real-time token frequency (uses redis store)
 * Measure social media latency (time from creation time stamp to delivery)

Requires Python urllib2 and, for token frequency, Redis and the python
redis module.

The tool is driven from a config file named gnip.cfg. To get started, 
enter your Gnip login credentials and stream URL and choose the tool 
function you desire.

Usage:

All examples run with Gnip Powertrack rule set:

tebow
broncos
football
win
gnip
beer -root

Example 1: Rule production rates
--------------------------------
In gnip.cfg, set

...
# print out rule production rates
processtype=rules
...

Scott-Hendricksons-MacBook-Pro ~/IdeaProjects/RuleProduction/src> ./GnipStreamCollectorMetrics.py 
(2012-01-06 13:07:53.989700) sample 154 tweets (20 seconds)
win: .......... (  91 tweets matched)   4.4592 tweets/second
football: ..... (  35 tweets matched)   1.7151 tweets/second
beer -root: ... (  24 tweets matched)   1.1761 tweets/second
tebow: ........ (   5 tweets matched)   0.2450 tweets/second
(2012-01-06 13:08:14.484888) sample 158 tweets (20 seconds)
win: .......... (  95 tweets matched)   4.6359 tweets/second
football: ..... (  28 tweets matched)   1.3664 tweets/second
beer -root: ... (  24 tweets matched)   1.1712 tweets/second
tebow: ........ (   9 tweets matched)   0.4392 tweets/second
broncos: ...... (   5 tweets matched)   0.2440 tweets/second
(2012-01-06 13:08:34.731178) sample 162 tweets (20 seconds)
win: .......... ( 110 tweets matched)   5.4359 tweets/second
football: ..... (  28 tweets matched)   1.3837 tweets/second
beer -root: ... (  21 tweets matched)   1.0378 tweets/second
broncos: ...... (   2 tweets matched)   0.0988 tweets/second
tebow: ........ (   2 tweets matched)   0.0988 tweets/second
(2012-01-06 13:08:55.285500) sample 189 tweets (20 seconds)
win: .......... ( 110 tweets matched)   5.3634 tweets/second
beer -root: ... (  39 tweets matched)   1.9016 tweets/second
football: ..... (  35 tweets matched)   1.7065 tweets/second
tebow: ........ (   4 tweets matched)   0.1950 tweets/second
broncos: ...... (   1 tweets matched)   0.0488 tweets/second
(2012-01-06 13:09:15.276119) sample 176 tweets (20 seconds)
win: .......... ( 111 tweets matched)   5.5474 tweets/second
football: ..... (  35 tweets matched)   1.7492 tweets/second
beer -root: ... (  18 tweets matched)   0.8996 tweets/second
tebow: ........ (   8 tweets matched)   0.3998 tweets/second
broncos: ...... (   5 tweets matched)   0.2499 tweets/second
...


The parameter "rollduration" sets the time between reports.  You may need to adjust this based
on the productivity of your rule set. 


Example 2: English token frequency (Requires redis)
---------------------------------------------------

Start your local redis server and perform a "flushall" to clear the database.

In gnip.cfg, set

...
# store stream contents in redis
processtype=redis
...

Start the client:

Scott-Hendricksons-MacBook-Pro ~/IdeaProjects/RuleProduction/src> ./GnipStreamCollectorMetrics.py 
(no output expected)
...


To see token frequencies, let the client run for a few minutes to collect data.  Then run the
RedisFreq.py tool to see what is stored in redis.  The token counts accumulated in the
redis store decay over 90 seconds so the store will reach steady state size after a few minutes.

You can monitor redis with a variety of tools. For example, with 
the redis-cli tool is include with redis.  Execute the command "dbsize"
to see the number of keys in the store.  Like this: 

redis 127.0.0.1:6379> dbsize
(integer) 4097
redis 127.0.0.1:6379> 

This shows that there are 4097 redis keys in the store.

Running the RedisFreq.py script after a few minutes gives the 100 more frequent
tokens from the redis store.  A sample of the counts and fractions
first few tokens is shown below.  This data shows 15627 tokens counted with
"win" being the most common and comprising about 6% of the tokens counted.

(Edit the stopwords list in the script "Redis.py" if you want to
include/exclude different tokens.)


Scott-Hendricksons-MacBook-Pro ~/IdeaProjects/RuleProduction/src> ./RedisFreq.py 
TotalTokensCount ......... 15627 (1.00000)
TotalRuleMatchCount ...... 1613 (0.10322)
win ...................... 1006 (0.06438)
football ................. 286 (0.01830)
beer ..................... 237 (0.01517)
follow ................... 124 (0.00793)
chance ................... 109 (0.00698)
giveaway ................. 99 (0.00634)
free ..................... 99 (0.00634)
please ................... 73 (0.00467)
enter .................... 69 (0.00442)
new ...................... 66 (0.00422)
game ..................... 64 (0.00410)
out ...................... 55 (0.00352)
play ..................... 54 (0.00346)
good ..................... 50 (0.00320)
tebow .................... 47 (0.00301)
team ..................... 46 (0.00294)
today .................... 45 (0.00288)
tonight .................. 44 (0.00282)
one ...................... 44 (0.00282)
2012 ..................... 43 (0.00275)
last ..................... 41 (0.00262)
via ...................... 40 (0.00256)
want ..................... 40 (0.00256)
love ..................... 39 (0.00250)
lol ...................... 39 (0.00250)
time ..................... 39 (0.00250)
day ...................... 38 (0.00243)
thanks ................... 38 (0.00243)
over ..................... 38 (0.00243)
see ...................... 37 (0.00237)
...


The redis store also contains rule match counts.  Rules are stored in redis as strings
surrounded by "[...]". To see rule counts with redis-cli:

redis 127.0.0.1:6379> get "[tebow]"
"127"
redis 127.0.0.1:6379> get "[win]"
"2240"
redis 127.0.0.1:6379> get "[football]"
"675"
redis 127.0.0.1:6379> 
...


Example 3: Latency measurements
--------------------------------
Latency estimates in this case include the python processes (and inherent variability) but
will give some idea of an application can accomplish.  It should be fairly straight forward
to build an application that beats these measurements.

The latency data is output to standard out the measurement for each activity on a separate
line.  

Be sure to set the rollduration to 1 sec. The rollduration is
included in the measured latency.

In gnip.cfg, set

...
rollduration=1
...


Scott-Hendricksons-MacBook-Pro ~/IdeaProjects/RuleProduction/src> ./GnipStreamCollectorMetrics.py 
2012-01-06 20:25:52.841085, 10.841085
2012-01-06 20:25:52.849420, 10.849420
2012-01-06 20:25:52.851427, 10.851427
2012-01-06 20:25:53.954821, 11.954821
2012-01-06 20:25:53.957109, 11.957109
2012-01-06 20:25:53.959620, 12.959620
2012-01-06 20:25:53.962573, 14.962573
2012-01-06 20:25:53.964144, 11.964144
2012-01-06 20:25:53.965350, 11.965350
2012-01-06 20:25:53.966468, 11.966468
2012-01-06 20:25:53.967509, 11.967509
2012-01-06 20:25:53.968822, 12.968822
2012-01-06 20:25:53.970103, 11.970103
2012-01-06 20:25:53.971735, 10.971735
2012-01-06 20:25:55.282562, 13.282562
2012-01-06 20:25:55.285407, 13.285407
2012-01-06 20:25:55.287489, 12.287489
2012-01-06 20:25:55.289636, 13.289636
...

In the sample shown, the average latency is around 12 s.  Gnip documentations estimates
that the stream enhancement adds about 10s of latency on average.  The 
GnipStreamCollectorMetrics.py adds approximately 1.5s on average.

An R utility is provided to create boxplots of latencies. (See ./R/plotLatency.r)

Example 4: Files
-----------------

To store data in files, edit gnip.cfg to set the appropriate rollduration and set
the filepath.  Don't forget to set the processtype to "files"

...
filepath=/Users/scotthendrickson/IdeaProjects/RuleProduction/data
...

...
# store stream contents in files by year, month, day, hour, min
processtype=files
...

Run script.

Scott-Hendricksons-MacBook-Pro ~/IdeaProjects/RuleProduction/src> ./GnipStreamCollectorMetrics.py 
(no output expected)

When you look in the data path:

Scott-Hendricksons-MacBook-Pro ~/IdeaProjects/RuleProduction/data> tree
.
└── 2012
    └── 01
        └── 06
            └── 14
                ├── Powertrack_201201061448497544.gz
                ├── Powertrack_201201061448783761.gz
                └── Powertrack_201201061449705606.gz
...

New directories will be created for every hour.  Here the rollduration is set to 
60s so a new file is created every minute.


If your input stream is JSON-formated, you can additionally parse it with Gnacs before writing to disk:

...
# store stream contents in files by year, month, day, hour, min
processtype=files-gnacs
...

Your cfg file must have a [gnacs] block, with a mandatory parameter 'delim' and an optional parameter
'options', which sets the Gnacs options.

...
[gnacs]
options=gulist
delim=|
,,,

When using 'files-gnacs', the output directory structure and files names are the same as for 'files'.

Example 5: Rules
-----------------

To store counts of rule matches, edit gnip.cfg to set the appropriate rollduration and set
the filepath.  Set the processtype to "rules". The output will have the same structure as
"file" and "file-gnacs".

...
filepath=/Users/scotthendrickson/IdeaProjects/RuleProduction/data
...

...
processtype=rules
...

Run script.

Scott-Hendricksons-MacBook-Pro ~/IdeaProjects/RuleProduction/src> ./GnipStreamCollectorMetrics.py 
(no output expected)

When you look in the data path:

Scott-Hendricksons-MacBook-Pro ~/IdeaProjects/RuleProduction/data> tree
.
└── 2012
    └── 01
        └── 06
            └── 14
                ├── Powertrack_201201061448497544.counts
                ├── Powertrack_201201061448783761.counts
                └── Powertrack_201201061449705606.counts
...

New directories will be created for every hour.  Here the rollduration is set to 
60s so a new file is created every minute.

Enjoy!