flume-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Paul Chavez <pcha...@verticalsearchworks.com>
Subject RE: Dynamic Key=Value Parsing with an Interceptor?
Date Mon, 11 Nov 2013 23:09:19 GMT
I think there may be two 'out of box' ways to do this kind of thing. First would be using the
regex extract interceptor with multiple serializers keying on various fields. However that's
not really dynamic and just kind of a half-step better from one interceptor for each field
as you mentioned. Second would be to use the morphline interceptor to parse your event body
and insert headers as needed. I have to admit I have no experience with this interceptor but
in reading the documentation it seems designed for this kind of use case.

Ultimately though, when faced with this we opted to push this into the app layer. Is there
a reason the applications can't write these key/value pairs as headers in the first place?
We use an HTTP source and when we wrote the logging class for it on our app side we put similar
functionality in as category/subcategory headers. Then flume doesn't have to have any special
interceptors beyond a default static one in case the headers are completely missing, and we
write to HDFS with tokenized paths so each permutation of those headers gets a separate directory.

If you continue to explore this issue, please keep us updated. I especially would like to
hear some real world morphline examples.

Hope that helps,
Paul Chavez


From: Matt Wise [mailto:matt@nextdoor.com]
Sent: Monday, November 11, 2013 10:04 AM
To: user@flume.apache.org
Subject: Re: Dynamic Key=Value Parsing with an Interceptor?

Anyone have any ideas on the best way to do this?

Matt Wise
Sr. Systems Architect
Nextdoor.com

On Sat, Nov 9, 2013 at 5:28 PM, Matt Wise <matt@nextdoor.com<mailto:matt@nextdoor.com>>
wrote:
Hey we'd like to set up a default format for all of our logging systems... perhaps looking
like this:

  "key1=value1;key2=value2;key3=value3...."

With this pattern, we'd allow developers to define any key/value pairs they want to log, and
separate them with a common separator.

If we did this, what do we need to do in Flume to get Flume to parse out the key=value pairs
into dynamic headers? We pass our data from Flume into both HDFS and ElasticSearch sinks.
We would really like to have these fields dynamically sent to the sinks for much easier parsing
and analysis later.

Any thoughts on this? I know that we can define a unique interceptor for each service that
looks for explicit field names ... but thats a nightmare to manage. I really want something
truly dynamic.

Matt Wise
Sr. Systems Architect
Nextdoor.com


Mime
View raw message