[Hadoop-common-user] Web Analytics Use case?

Benjamin,

That is kind of the exact case for Hadoop.

Hadoop is a system that is built for handling very large datasets, and
delivering processed results. HBase is built for AdHoc data, so
instead of having complicated table joins etc, you have very large
rows (multiple columns) with aggregate data, then use HBase to return
results from that.

We currently use hadoop/hbase to collect and process lots of data,
then take the results from the processing to populate a SOLR Index,
and a MySQL database which is then used to feed the front ends. It
seems to work pretty good in that it greatly reduces the number of
rows and the size of the queries in the DB/index.

We are exploring using HBase to feed the front-ends in place of the
MySQL DBs, so far the jury is out on the performance but it does look
promising.

-John

Hi John,

Thanks a lot for the fast answer. I was unsure because we would like to avoid aggregating the data so that our users can come up with all kinds of filters and conditions for your queries and always drill down to single users of their website. I am not sure how this works when SQL is not directly available? We are currently using complex sql queries for this, these would then have to be rewritten in form of Map/reduce tasks which provide the final result?

Or how would one go about to actually replace an RDBMS system?

Thanks a lot,
Benjamin



_______________________________________
Benjamin Dageroth, Business Development Manager
Webtrekk GmbH
Boxhagener Str. 76-78, 10245 Berlin
fon 030 - 755 415 - 360
fax 030 - 755 415 - 100
[email protected]
http://www.webtrekk.com
Amtsgericht Berlin, HRB 93435 B
Geschäftsführer Christian Sauer


_______________________________________


-----Ursprüngliche Nachricht-----
Von: John Martyniak
Gesendet: Dienstag, 3. November 2009 15:09
An: [email protected]
Betreff: Re: Web Analytics Use case?

Benjamin,

That is kind of the exact case for Hadoop.

Hadoop is a system that is built for handling very large datasets, and
delivering processed results. HBase is built for AdHoc data, so
instead of having complicated table joins etc, you have very large
rows (multiple columns) with aggregate data, then use HBase to return
results from that.

We currently use hadoop/hbase to collect and process lots of data,
then take the results from the processing to populate a SOLR Index,
and a MySQL database which is then used to feed the front ends. It
seems to work pretty good in that it greatly reduces the number of
rows and the size of the queries in the DB/index.

We are exploring using HBase to feed the front-ends in place of the
MySQL DBs, so far the jury is out on the performance but it does look
promising.

-John

I think you should break down your question into 3 ...

1) How do I collect the raw data ?
2) How do I process the raw data that I collect above to produce useful information ?
3) How do I store the information for large scale consumption needs ?


In (1), looks like you can collecting raw data using tracking pixel, but where do you store these raw data ? The best way to store them map depends on your answer in (2).

In (2), what algorithms are you choosing to process the raw data ? Are they parallelizable ? Hadoop is certainly a candidate that you should consider. But be reminded that Hadoop is a batch-oriented processing framework which means your processing will be chopped into chunks. Double check your application is OK with this processing model.

In (3), what is the size of the output information and how it will be consumed ? Most likely your analytic application has a high tolerance on data integrity issue and that's why you can use a more scalable Non-SQL DB which sacrifice some data integrity but provide a lot of scalability. Note that the decision in (3) is completely independent of (2).

Rgds,
Ricky
http://horicky.blogspot.com

-----Original Message-----
From: Benjamin Dageroth
Sent: Tuesday, November 03, 2009 7:16 AM
To: [email protected]
Subject: AW: Web Analytics Use case?

Hi John,

Thanks a lot for the fast answer. I was unsure because we would like to avoid aggregating the data so that our users can come up with all kinds of filters and conditions for your queries and always drill down to single users of their website. I am not sure how this works when SQL is not directly available? We are currently using complex sql queries for this, these would then have to be rewritten in form of Map/reduce tasks which provide the final result?

Or how would one go about to actually replace an RDBMS system?

Thanks a lot,
Benjamin



_______________________________________
Benjamin Dageroth, Business Development Manager
Webtrekk GmbH
Boxhagener Str. 76-78, 10245 Berlin
fon 030 - 755 415 - 360
fax 030 - 755 415 - 100
[email protected]
http://www.webtrekk.com
Amtsgericht Berlin, HRB 93435 B
Geschäftsführer Christian Sauer


_______________________________________


-----Ursprüngliche Nachricht-----
Von: John Martyniak
Gesendet: Dienstag, 3. November 2009 15:09
An: [email protected]
Betreff: Re: Web Analytics Use case?

Benjamin,

That is kind of the exact case for Hadoop.

Hadoop is a system that is built for handling very large datasets, and
delivering processed results. HBase is built for AdHoc data, so
instead of having complicated table joins etc, you have very large
rows (multiple columns) with aggregate data, then use HBase to return
results from that.

We currently use hadoop/hbase to collect and process lots of data,
then take the results from the processing to populate a SOLR Index,
and a MySQL database which is then used to feed the front ends. It
seems to work pretty good in that it greatly reduces the number of
rows and the size of the queries in the DB/index.

We are exploring using HBase to feed the front-ends in place of the
MySQL DBs, so far the jury is out on the performance but it does look
promising.

-John

Benjamin,

Well instead of SQL you have code that you can use to manipulate the
data. If it was possible I would see if there was some way that you
can pre-process as much of the data as possible to put into HBase, and
then use any additional Map/Reduce jobs to provide any additional
customizations.

I don't think that you can "replace" the RDBMS without re-visualizing
the data, meaning that you will need to re-model it so that it fits
into HBase architecture, which means no relationships.

By the way most of this can be done, it just requires some work, and a
rethinking of the way that you do things, both for Map/Reduce and HBase.

-John

Hey,

Hadoop, HBase and Hive really scale for web analytics, I have been into web
analytics using Hadoop for more than a year.

In my case, I periodically rotate logs and put them on the HDFS. (I should
think of writing directly to HDFS; but it's not a critical issue for me
right now.)
When the log files are on the HDFS somehow, a single map/reduce job runs on
the newly introduced data line by line.

The key point here is, we need to think web analytics as a series of
abstraction on the raw data. Each abstraction analysis might symbolize a
map/reduce job.

The big question arises just right here, What does the initial analysis do
for the log files.

Abstraction #1:
I assume each log line either represents a pageview or an event; we can
generalize an event as a pageview too, and surely I will do so!
An event comes with some valuable information such as,
[Session identifier, unique visitor identifier, browser and locale related
data, page related data, location related data, etc...]

Abstraction #2:
Our map/reduce job should map an Event to a Session Event in order to get a
newer abstraction on the raw data. Session Events should be reduced into
Sessions with respect to the Session Identifiers as keys.
At the end of the first abstraction, we have our session data sorted out as
(key,value) pairs, where the keys are the Session Identifiers, and
presumably the Values should be the Sessions. Which means, now we can store
Sessions in a Key/Value database that in this case conforms to HBase.

One can think of additional abstractions from this point I think, I can come
up with many ideas some of which are fairly mature and some are just dreams
and/or premature thoughts.

Regards,
Utku

Good point. Hadoop can be used as a distributed DB loader.

Just curious. How would you compare this with directly write to HBase (by passing the log, Hadoop step) ?

Rgds,
Ricky

-----Original Message-----
From: Utku Can Topçu
Sent: Tuesday, November 03, 2009 4:39 PM
To: [email protected]
Subject: Re: AW: Web Analytics Use case?

Hey,

Hadoop, HBase and Hive really scale for web analytics, I have been into web
analytics using Hadoop for more than a year.

In my case, I periodically rotate logs and put them on the HDFS. (I should
think of writing directly to HDFS; but it's not a critical issue for me
right now.)
When the log files are on the HDFS somehow, a single map/reduce job runs on
the newly introduced data line by line.

The key point here is, we need to think web analytics as a series of
abstraction on the raw data. Each abstraction analysis might symbolize a
map/reduce job.

The big question arises just right here, What does the initial analysis do
for the log files.

Abstraction #1:
I assume each log line either represents a pageview or an event; we can
generalize an event as a pageview too, and surely I will do so!
An event comes with some valuable information such as,
[Session identifier, unique visitor identifier, browser and locale related
data, page related data, location related data, etc...]

Abstraction #2:
Our map/reduce job should map an Event to a Session Event in order to get a
newer abstraction on the raw data. Session Events should be reduced into
Sessions with respect to the Session Identifiers as keys.
At the end of the first abstraction, we have our session data sorted out as
(key,value) pairs, where the keys are the Session Identifiers, and
presumably the Values should be the Sessions. Which means, now we can store
Sessions in a Key/Value database that in this case conforms to HBase.

One can think of additional abstractions from this point I think, I can come
up with many ideas some of which are fairly mature and some are just dreams
and/or premature thoughts.

Regards,
Utku

The reason why I'm choosing DB loading is the fact that, each Session (WAA
calls this Visit) is composed of multiple Events, where an Event is a line
of Log File. For every session, we can be sure that, each session is
basically composed of one or more lines of log, that results in duplication
of many session constants (i.e IP Addr, User-Agent, Unique Visitor cookie,
etc..). In addition to this, the Reducer goes over the sessions once again
before loading them to the DB so that we can compute some session specific
calculations on the fly.

I hope I was precise and clear enough in expressing the design choice of
mine.

Regards,
Utku

Why can't you do the session specific calculation and aggregation at the spot where session data is gathered ?


One of the main usage of Map/Reduce is when the aggregation is done across a very scattered data set. But it looks like the kind of processing you describe is very localized. I mean the same session pretty much hitting the same server, so you can do the aggregation at the same spot.

Rgds,
Ricky

-----Original Message-----
From: Utku Can Topçu
Sent: Wednesday, November 04, 2009 7:48 AM
To: [email protected]
Subject: Re: AW: Web Analytics Use case?

The reason why I'm choosing DB loading is the fact that, each Session (WAA
calls this Visit) is composed of multiple Events, where an Event is a line
of Log File. For every session, we can be sure that, each session is
basically composed of one or more lines of log, that results in duplication
of many session constants (i.e IP Addr, User-Agent, Unique Visitor cookie,
etc..). In addition to this, the Reducer goes over the sessions once again
before loading them to the DB so that we can compute some session specific
calculations on the fly.

I hope I was precise and clear enough in expressing the design choice of
mine.

Regards,
Utku

Ricky,

You're absolutely right, I've already started developing a new data
collection system that populates sessions on the fly. Till that development
is finished, I felt like I needed to develop a layered approach of
abstractions. Session aggregation might override the initial Hadoop run at
the end of the day.

The case with these log files are that, they are just collections of past 5
years. When I started this, (5 years ago) I had no idea of what I would be
facing :)