Chapter 2: Architecture

Logback's architecture

Logback's basic architecture is sufficiently generic so as to apply under different circumstances. At present time, logback is divided into three modules, Core, Classic and Access.

The core module lays the groundwork for the other two modules. The classic module extends core. The classic module corresponds to a significantly improved version of log4j. Logback-classic natively implements the SLF4J API so that you can readily switch back and forth between logback and other logging systems such as log4j or JDK14 Logging. The third module called access integrates with Servlet containers to provide HTTP-access log functionality. The access module will be covered in a separate document.

In the reminder of this document, we will write "logback" to refer to the logback classic module.

Logger, Appenders and Layouts

Logback has three main types: Logger, Appender and Layout. These three types of components work together to enable developers to log messages according to message type and level, and to control at runtime how these messages are formatted and where they are reported.

The Logger class is part of the classic module. On the other hand, the Appender and Layout interfaces are part of the core module. For the sake of genericity, logback-core has no notion of loggers.

Logger context

The first and foremost advantage of any logging API over plain System.out.println resides in its ability to disable certain log statements while allowing others to print unhindered. This capability assumes that the logging space, that is, the space of all possible logging statements, is categorized according to some developer-chosen criteria. In logback, this categorization is an inherent part of loggers.

Loggers are named entities. Their names are case-sensitive and they follow the hierarchical naming rule:

For example, the logger named "com.foo" is a parent of the logger named "com.foo.Bar". Similarly, "java" is a parent of "java.util" and an ancestor of "java.util.Vector". This naming scheme should be familiar to most developers.

The root logger resides at the top of the logger hierarchy. It is exceptional in that it is part of every hierarchy at its inception. Like every logger, it can be retrieved by its name, as follows:

Logger rootLogger = LoggerFactory.getLogger(LoggerContext.ROOT_NAME);

All other loggers are also retrieved with the class static getLogger method found in the org.slf4j.LoggerFactory class. This method takes the name of the desired logger as a parameter. Some of the basic methods in the Logger interface are listed below.

package org.slf4j; 
public interface Logger {

  // Printing methods: 
  public void debug(String message);
  public void info(String message); 
  public void warn(String message); 
  public void error(String message); 
  public void fatal(String message); 
}

Loggers may be assigned levels. The set of possible levels, that is DEBUG, INFO, WARN and ERROR are defined in the ch.qos.logback.classic.Level class. Note that in logback, the level class is final and cannot be derived, as a much more flexible approach exist in the form of Marker objects.

If a given logger is not assigned a level, then it inherits one from its closest ancestor with an assigned level. More formally:

To ensure that all loggers can eventually inherit a level, the root logger always has an assigned level. By default, this level is DEBUG.

Below are four examples with various assigned level values and the resulting effective (inherited) levels according to the level inheritance rule.

In example 1 above, only the root logger is assigned a level. This level value, DEBUG, is inherited by the other loggers X, X.Y and X.Y.Z

In example 2 above, all loggers have an assigned level value. Level inheritence does not come into play.

In example 3 above, the loggers root, X and X.Y.Z are assigned the levels DEBUG, INFO and ERROR respectively. Logger X.Y inherits its level value from its parent X.

In example 4 above, the loggers root and X and are assigned the levels DEBUG and INFO respectively. The loggers X.Y and X.Y.Z inherit their level value from their nearest parent X, which has an assigned level.

Printing methods

By definition, the printing method determines the level of a logging request. For example, if L is a logger instance, then the statement L.info("..") is a logging statement of level INFO.

A logging request is said to be enabled if its level is higher than or equal to the level of its logger. Otherwise, the request is said to be disabled. A logger without an assigned level will inherit one from the context. This rule is summarized below.

This rule is at the heart of logback. It assumes that levels are ordered as follows: DEBUG < INFO < WARN < ERROR< OFF.

In a more graphic way, here is how the selection rule works. In the following table, the vertical header shows the the level of the logging request, designated by p, while the horizontal header shows effective level of the logger, designated by q.

Here is an example of the basic selection rule.

// get a logger instance named "com.foo", with an INFO level. 
Logger logger = LoggerFactory.getLogger("com.foo");
//set its Level to INFO
logger.setLevel(Level. INFO);
Logger barlogger = LoggerFactory.getLogger("com.foo.Bar");

// This request is enabled, because WARN >= INFO
logger.warn("Low fuel level.");

// This request is disabled, because DEBUG < INFO. 
logger.debug("Starting search for nearest gas station.");

// The logger instance barlogger, named "com.foo.Bar", 
// will inherit its level from the logger named 
// "com.foo" Thus, the following request is enabled 
// because INFO >= INFO. 
barlogger.info("Located nearest gas station.");

// This request is disabled, because DEBUG < INFO. 
barlogger.debug("Exiting gas station search");

Retrieving Loggers

Calling the LoggerFactory.getLogger method with the same name will always return a reference to the exact same logger object.

For example, in

Logger x = LoggerFactory.getLogger("wombat"); 
Logger y = LoggerFactory.getLogger("wombat");

x and y refer to exactly the same logger object.

Thus, it is possible to configure a logger and then to retrieve the same instance somewhere else in the code without passing around references. In fundamental contradiction to biological parenthood, where parents always preceed their children, logback loggers can be created and configured in any order. In particular, a "parent" logger will find and link to its descendants even if it is instantiated after them.

Configuration of the logback environment is typically done at application initialization. The preferred way is by reading a configuration file. This approach will be discussed shortly.

Logback makes it easy to name loggers by software component. This can be accomplished by instantiating a logger in each class, with the logger name equal to the fully qualified name of the class. This is a useful and straightforward method of defining loggers. As the log output bears the name of the generating logger, this naming strategy makes it easy to identify the origin of a log message. However, this is only one possible, albeit common, strategy for naming loggers. Logback does not restrict the possible set of loggers. As a developer, you are free to name loggers as you wish.

Nevertheless, naming loggers after the class where they are located seems to be the best general strategy known so far.

Appenders and Layouts

The ability to selectively enable or disable logging requests based on their logger is only part of the picture. Logback allows logging requests to print to multiple destinations. In logback speak, an output destination is called an appender. Currently, appenders exist for the console, files, remote socket servers, to MySQL, PostgreSQL, Oracle and other databases, JMS, and remote UNIX Syslog daemons.

More than one appender can be attached to a logger.

The addAppender method adds an appender to a given logger. Each enabled logging request for a given logger will be forwarded to all the appenders in that logger as well as the appenders higher in the hierarchy. In other words, appenders are inherited additively from the logger hierarchy. For example, if a console appender is added to the root logger, then all enabled logging requests will at least print on the console. If in addition a file appender is added to a logger, say L, then enabled logging requests for L and L's children will print on a file and on the console. It is possible to override this default behavior so that appender accumulation is no longer additive by setting the additivity flag of a logger to false.

The rules governing appender additivity are summarized below.

More often than not, users wish to customize not only the output destination but also the output format. This is accomplished by associating a layout with an appender. The layout is responsible for formatting the logging request according to the user's wishes, whereas an appender takes care of sending the formatted output to its destination. The PatternLayout, part of the standard logback distribution, lets the user specify the output format according to conversion patterns similar to the C language printf function.

For example, the PatternLayout with the conversion pattern "%-4relative [%thread] %-5level %logger{32} - %msg%n" will output something akin to:

176  [main] DEBUG chapter2.HelloWorld2 - Hello world.

The first field is the number of milliseconds elapsed since the start of the program. The second field is the thread making the log request. The third field is the level of the log request. The fourth field is the name of the logger associated with the log request. The text after the '-' is the message of the request.

Parameterized logging

Given that loggers in logback-classic implement the SLF4J's Logger interface, certain printing methods admit more than one parameter. These printing method variants are mainly intended to improve performance while minimizing the impact on the readability of the code.

For some Logger logger, writing,

logger.debug("Entry number: " + i + " is " + String.valueOf(entry[i]));

incurs the cost of constructing the message parameter, that is converting both integer i and entry[i] to a String, and concatenating intermediate strings. This, regardless of whether the message will be logged or not.

One possible way to avoid the cost of parameter construction is by surrounding the log statement with a test. Here is an example.

if(logger.isDebugEnabled()) { 
  logger.debug("Entry number: " + i + " is " + String.valueOf(entry[i]));
}

This way you will not incur the cost of parameter construction if debugging is disabled for logger. On the other hand, if the logger is enabled for the DEBUG level, you will incur the cost of evaluating whether the logger is enabled or not, twice: once in debugEnabled and once in debug. This is an insignificant overhead because evaluating a logger takes less than 1% of the time it takes to actually log a request.

Better alternative

There exists a convenient alternative based on message formats. Assuming entry is an object, you can write:

Object entry = new SomeObject(); 
logger.debug("The entry is {}.", entry);

After evaluting whether to log or not, and only if the decision is positive, will the logger implementation format the message and replace the '{}' pair with the string value of entry. In other words, this form does not incur the cost of parameter construction in case the log statement is disabled.

The following two lines will yield the exact same output. However, in case of a disabled logging statement, the second variant will outperform the first variant by a factor of at least 30.

logger.debug("The new entry is "+entry+".");
logger.debug("The new entry is {}.", entry);

A two argument variant is also availalble. For example, you can write:

logger.debug("The new entry is {}. It replaces {}.", entry, oldEntry);

If three or more arguments need to be passed, an Object[] variant is also availalble. For example, you can write:

Object[] paramArray = {newVal, below, above};
logger.debug("Value {} was inserted between {} and {}.", paramArray);

A peak under the hood

After we have introduced the essential logback components, we are now ready to describe the steps that the logback framework takes when the user invokes a logger's printing method. Let us now analyze the steps logback takes when the user invokes the info() method of a logger named com.wombat.

1. Get the filter chain decision

If it exists, the TurboFilter chain is invoked. Turbo filters can set a context-wide threshold, or to filter out certain events based on information such as Marker, Level, Logger, message, or the Throwable that are associated with each logging request. If the reply of the filter chain is FilterReply.DENY, then the logging request is dropped. If it is FilterReply.NEUTRAL, then we continue with the next step, i.e. step 2. In case the reply is FilterReply.ACCEPT, we skip the next and directly jump to step 3.

2. Apply the Logger level filter

At this step, logback compares the effective level of the logger with the level of the request. If the logging request is disabled according to this test, then logback will drop the request without further processing. Otherwise, it proceeds to the next step.

3. Create a LoggingEvent object

If the request passed the previous filters, logback will create a ch.qos.logback.classic.LoggingEvent object containing all the relevant parameters of the request, such as the logger of the request, the request level, the message itself, the exception that might have been passed along with the request, the current time, the current thread, various data about the class that issued the logging request and the MDC. Note that some of these fields are initialized lazily, that is only when they are actually needed. The MDC is used to decorate the logging request with additional contextual information. MDC is discussed in a subsequent chapter.

4. Invoking appenders

After the creation of a LoggingEvent object, logback invokes the doAppend() methods of all the applicable appenders, that is, the appenders inherited from the logger context.

All appenders shipped with the logback distribution extend the AppenderBase abstract class that implements the doAppend method in a synchronized block ensuring thread-safety. The doAppend() method of AppenderBase also invokes custom filters attached to the appender, if any such filters exist. Custom filters, which can be dynamically attached to any appender, are presented in a separate chapter.

5. Formatting the output

It is responsibility of the invoked appender to format the logging event. However, some (but not all) appenders delegate the task of formatting the logging event to a layout. A layout formats the LoggingEvent instance and returns the result as a String. Note that some appenders, such as the SocketAppender, do not transform the logging event into a string but serialize it instead. Consequently, they do not require nor have a layout.

6. Sending out the LoggingEvent

After the logging event is fully formatted it is sent to its destination by each appender.

Here is a sequence UML diagram to show how everything works. You might want to click on the image to display its bigger version.

Performance

One of the often-cited arguments against logging is its computational cost. This is a legitimate concern as even moderately sized applications can generate thousands of log requests. Much effort is spent measuring and tweaking logging performance. Independently of these efforts, the user should still be aware of the following performance issues.

1. Logging performance when logging is turned off entirely

You can turn off logging entirely by setting the level of the root logger to Level.OFF, the highest possible level. When logging is turned off entirely, the cost of a log request consists of a method invocation plus an integer comparison. On a 3.2Ghz Pentium D machine this cost is typically around 20 nanoseconds.

However, any method invocation involves the "hidden" cost of parameter construction. For example, for some logger x writing,

x.debug("Entry number: " + i + "is " + entry[i]);

incurs the cost of constructing the message parameter, i.e. converting both integer i and entry[i] to a string, and concatenating intermediate strings, regardless of whether the message will be logged or not.

The cost of parameter construction can be quite high and depends on the size of the parameters involved. To avoid the cost of parameter construction you can take advantage of SLF4J's parametrized logging:

x.debug("Entry number: {} is {}", i, entry[i]);

This variant will not incur the cost of parameter construction. Compared to the previous call to the debug() method, it will be faster by a very wide margin. The message will be formatted only if the request is processed to the appenders. If it is processed, the component that formats the message offers high performance and does not impact negatively the overall process. It respectively takes 2 and 4 microseconds to format a message with 1 and 3 parameters.

Please notice that, despite the performance points that we just discussed, inserting logging statements in tight-loops or very frequently invoked code is a lose-lose proposal and is likely to result in degraded performance. Logging in tight-loops slow down your application even if logging is turned off, and if logging is turned on, they will generate massive (and hence useless) output.

2. The performance of deciding whether to log or not to log when logging is turned on.

In logback, there is no need to walk the logger hierarchy. A logger knows its effective level (that is, its level, once level inheritance has been taken into consideration) when it is created. Should the level of a parent logger be changed, then all child loggers are contacted to take notice of the change. Thus, before accepting or denying a request based on the effective level, the logger can make an quasi-instantaneous decision, without needing to consult its ancestors.

3. Actual logging (formatting and writing to the output device)

This is the cost of formatting the log output and sending it to its target destination. Here again, a serious effort was made to make layouts (formatters) perform as quickly as possible. The same is true for appenders. The typical cost of actually logging is about 9 to 12 microseconds when logging to a file on the local machine. It goes up to several milliseconds when logging to a database on a remote server.

Although feature-rich, one of the foremost design goals of logback was speed of execution, a requirement which is second only to reliability. Some logback components have been rewritten several times to improve performance.