On Thu, 23 Jan 2014, Rainer Gerhards wrote:
On Thu, Jan 23, 2014 at 8:00 AM, David Lang <[email protected]> wrote:
On Thu, 23 Jan 2014, Radu Gheorghe wrote:
2014/1/22 Rainer Gerhards <[email protected]>
On Wed, Jan 22, 2014 at 3:17 PM, Radu Gheorghe <[email protected]
wrote:
Hi Rainer,
I'd like to see an Apache Solr <https://lucene.apache.org/solr/> output
plugin.
You can see in my presentation
here<
http://blog.sematext.com/2013/12/16/video-using-solr-for-
logs-with-rsyslog-flume-fluentd-and-logstash/
at
slides 21-26 that this is what I did as a demo with omprog. The actual
python script that pushes logs to Solr is on slide 26.
Oh, I think that is an excellent example, especially as it is very
brief
and still does useful work. That's one of the core ideas to convey so
that
others become interested in doing a similar thing.
The thing works, but:
- it would only send logs one by one
- it has no error handling
Though I think it should be easy to add.
Would be great if you could help with that over time, as I neither have
the
know how nor environment...
Yep, I could help. If you need some steps for setting up the env just
let
me know.
I think Solr is a really nice destination for logs, offering similar
stuff
to what Elasticsearch does. Plus some, minus some, and they're both
based
on Apache Lucene <https://lucene.apache.org/>. You can see in the same
presentation that other tools already send logs to Solr, and I think
Apache
Flume <https://flume.apache.org/> is the most widely used of them.
Big question now: which changes in rsyslog are required to make you
happier
as an author of such a non-C module? What would you like to see improved?
Batch processing and error handling is what is needed for this script to
work. I had a similar script some time ago before omelasticsearch and it
worked just fine, even at high load (although the final argument for
omelasticsearch was that the script was using way more resources than
rsyslog under high load).
If rsyslog could provide me a code sample that does batch processing and
error handling and I would just import the Solr library and add the
function to post message batches, that would be awesome.
For error handling, things seem pretty straightforward:
- if an error is permanent (eg: invalid JSON/XML) log a warning (where?)
and discard the message
- if the error is temporary (eg: Solr is down), but messages back in the
queue and retry later
What queue? This brings me to the batch processing thing. If you need
batch
processing, you need to read something from the pipe and put it into the
array and do so until you have enough requests in the array to send the
batch to Solr (size/time).
This single-threaded approach has some disadvantages:
- only one thread can send to Solr at the same time. With HTTP requests
it's typically better to have more (I know this from omelasticsearch)
- the process of adding to the array has to wait for the indexing process
to finish
Therefore a more efficient way is to have a thread that takes from the
pipe
and puts into the scripts (tiny) internal queue, an then one or more
threads that take messages from the queue, make arrays and index them to
Solr. And in case of a temporary error, put those messages back to the
queue.
Now, this whole logic is quite complicated (though not extremely
complicated), and it would be a show-stopper IMO if one had to re-do it
for
every such non-C output module.
A better option IMO is to have a guide that said:
- make sure you have omprog
- configure the template like this
- deploy this skeleton script like this (list of languages, link to each
repo and how to add new languages)
- import the library you need to sending to X
- change the "send()" function or whatever the name is to send the array
of
messages to X.
Only one issue left (as far as I see): multiline messages. Let''s say the
template sends JSONs separated by a newline (is there a better way?). If
one field (say $msg) spreads on multiple lines, then the "reading"thread
can't simply read from stdin (non-blocking!) line-by-line. Maybe we can
use
rsyslog's built-in newline escaping and re-construct newlines in the
"send()"threads?
See? Quite a lot of logic. I think providing this for people instead of
having them re-invent it for every script they do would be a huge step
towards adoption.
This is why I am suggesting that instead of ascii based message passing
(and encoding it in JSON, dealing with escaping, linebreaks, etc) we should
use protocol buffers and/or thrift for the message passing.
It would let logs be passed as structures, using optimized parser
libraries (one of the problems with json-c is it's lack of optimization),
allow batches of logs to be passed, allow non-log messages to be passed
without any possibility of confusion, etc.
David, while I basically agree on all the points that you give, I still
stick with the idea of pipes. I think you are overlooking a the work
involved. If we "do it right" in the first place, we'll probably never do
it in any case (at least I am very skeptical if I can find sufficient time
to do the necessary work in rsyslog -- there still is a long todo list).
On the other hand, if we accept to re-invent the wheel at least twice (so
to say), I can stick some of the omprog work into my schedule and so we get
something to actually prove in practice. This would immediately enable
anyone to write/contribute output plugins. With the knowledge gained there,
I could create an input counterpart, most probably also with comparatively
little effort. From there, it is possibly (not 100% sure) also easy to do
an external modification and filter interface. In the end result, we would
have all capabilities to interface to external programs, albeit not in a
perfect way. But folks could actually use it.
Once this is in place for some time, we can review where it hurts and then
decide on a next and better solution. I think this is much the same
approach as the overall "let's do non-c, external plugins" approach --
because by definition they are second class in many ways and really not
something one should do if seriously interested in performance. But we
concluded there is value in that.
For that reasoning I would like to proceed with the pipe interface for now.
Does that make sense to you?
I think it all depends on how much work we are going to have to do.
why do we need to do anything? can't omprog with a JSON template already do
what's needed?
as long as we stick to pipes, we aren't going to be able to get per message
success/failure messages (the buffering in the pipe will prevent that from
working well)
I guess I'm not fully understanding what's being proposed for the short term.
David Lang
_______________________________________________
rsyslog mailing list
http://lists.adiscon.net/mailman/listinfo/rsyslog
http://www.rsyslog.com/professional-services/
What's up with rsyslog? Follow https://twitter.com/rgerhards
NOTE WELL: This is a PUBLIC mailing list, posts are ARCHIVED by a myriad of
sites beyond our control. PLEASE UNSUBSCRIBE and DO NOT POST if you DON'T LIKE
THAT.
