BalaBit blog

Update: The syntax of this feature has slightly changed due to discussions on the mailing list (e.g. change the log keyword to channel), and I’ve updated it to use the current syntax.

syslog-ng is often referred as a very flexible application when it comes to processing logs. Over the years however, I began to feel that some things are a bit more difficult to achieve in the configuration language than it should be. For instance it is sometimes too rigid when you need a combination of parsers (patterndb with db-parser) and rewrite rules to achieve the goal you wanted. Parsers and rewrite rules are distinct part of the configuration, it is not possible to combine them into a single functionality. Also, declaring objects first and then referencing them later, makes the configuration easy to read, however sometimes it is quite cumbersome, when you only need to invert the result of an already existing filter.

To solve this situation, I’ve set out to implement an idea I had on mind for some time now. It is quite difficult to describe the feature in clear and concise words, as it is a combination of various changes that together makes syslog-ng configuration more flexible and easier to use, without sacrificing readability. Curious? Please read on.

In-line objects

Perhaps the simplest of all features is that you can now define the contents of a given object right on the spot, without having to use a separate statement. For example, earlier you had to write:

log {
  source(s_local);
  filter(f_postfix);
  destination(d_postfix);
};

Sometimes, f_postfix filter is only used once and is trivial. This can now be written as:

log {
 source(s_local);
 filter { program("^postfix/"); };
 destination(d_postfix);
};

Furthermore both the source() and destination() options can be written in-line, you simply use braces instead of parentheses. The same functionality applies to everything: sources, destinations, filters, parsers and rewrite rules.

Junctions

A limited form of junctions has been supported since syslog-ng 3.0 in the form of “embedded log statements”, which has been generalized now. Within syslog-ng, when a message is received it is dispatched to a log processing path or pipeline, which carries out the task at hand. A junction is a point in the log processing path where the processing is performed on multiple independent branches, each doing its own specific thing with the message.

The limited functionality in 3.0 only allowed the processing tree to split (or fork) into independent branches, each of the branches was a “sink”, where processing also ended. Configuration example:

log {
  source(s_all); filter(f);
  log { filter(f1); destination(d1); };
  log { filter(f2); destination(d2); };
};

This sample forks the processing path into two branches starting with the “log” keyword within the top-level log statement. The first branch evaluates the filter f1 and the writes matching messages to the d1 destination, effectively sending all messages that match (f AND f1) to d1. Likewise, d2 receives all messages that match (f AND f2).

The limitation of the embedded log statement concept was simple: it could only be listed at the very end of a log statement, and the end-result of the branches couldn’t be processed further. Effectively the message at the end of each branch “fell off”. Junctions on the other hand makes it possible to do things to messages once the branches converge to the same point again. Repeating the sample above, it is now possible to write:

log {
  source(s_all); filter(f);
  junction {
    channel { filter(f1); destination(d1); };
    channel { filter(f2); destination(d2); };
  };
  destination(d_all);
};

The new thing is that you can now add processing after the branches finish their processing. A bit more useful example would be:

log {
  source(s_apache_files);
  source(s_syslog);
  junction {
    channel { filter(f_apache_files); rewrite(r_apache_remove_file_header); parser(p_apache); flags(final); };
    channel { filter(f_apache_syslog); parser(p_apache); flags(final); };
  };
  destination(d_files);
};

This example does an alternative processing of incoming logs based on where the message came from.

Everything is a log expression

This feature is probably the most complicated, however provides very nice properties and expressiveness to the configuration. From now on, not just the well known log statement allows the specification of log processing rules, but all the objects in the syslog-ng configuration file can use the same expressive power.

It is now possible to use embedded log statements, junctions and in-line object definitions within source, destination, filter, rewrite and parser definitions. Huh, you could ask: what does it bring to me as a benefit? Well, until now, objects of different types were separate entities, connected using log statements, with this change a source can also specify a rewrite rule and that combination used as a log source in a log statement.

For instance, a usual source definition looked like this:

source s_apache {
  file("/var/log/apache/error.log");
};

If you wanted to process this log file in a specific way, you needed to define the accompanying processing rules (parsers and rewrite expressions) and combine them in a log statement. But how about this:

source s_apache {
  channel {
    source { file("/var/log/apache/error.log"); };
    parser(p_apache_parser); };
  };
};

log { source(s_apache); ... };

Can you see? The s_apache source used a file source and the reference of a specific parser and all messages read from the apache error log file would be processed by that parser. The log statement is just as simple as if s_apache would be a “normal” source definition. This feature allows pairing the essential log preprocessing functionality very close to the source itself, making it very easy to write and read the log statements. As an added bonus, it becomes very easy to distribute application specific source & parser definitions as an SCL configuration snippet.

Where?

This stuff is available in the syslog-ng 3.4 git tree, on master. It passes the included regression test, so it is at least dogfoodable. The nice thing about the implementation is that it only slightly increased the code size, but brought a lot of new features. If you have trouble getting the code from git, let me know, I’m willing to create an alpha release, so that it becomes easier to play with it.

Feedback

I see a lot of potential in this functionality, however my examples may have not been the best ones. I would really appreciate any kind of feedback, please be sure to send those to the syslog-ng mailing list or post me as a private email.