I would suggest others review the code in question, but here is what I see.

LoggerFactory initializes dynamically. The first caller binds to the implementation of the LoggerFactory by calling StaticLoggerBinder.getSingleton(), but only if it is the first caller. Other callers are given a substitute logger factory until the binding is complete. Log4j’s implementation creates a Log4jLoggerFactory instance but performs no initialization at this time, so the amount of time that SLF4J would be in this state should be very small. Regardless, events could still be routed to the substitute logger. If we want to ensure that these events are not discarded then we have to implement the new method.

Note that due to the way Log4j initializes we don’t have this kind of race going on anywhere. Ceki seems to imply that the race is in the underlying logging framework, but it is not. Or perhaps he is just assuming that all log events are coming through SLF4J, which may or may not actually be the case.  The race is entirely within SLF4J’s initialization logic. Anyone using just the Log4j API will not have this problem.

As I said previously, A Log4j LoggerContext always has a configuration, at least until it is shut down, and logging an event requires a constructed LoggerContext. A LoggerContext has a default configuration until the actual configuration file is processed and the full configuration constructed. The default configuration logs errors to the console. So any log events that actually make it to Log4j will have something done with them, but maybe not what the user wants.  If we wanted to have any events that occur before configuration is completed subject to that configuration the default configuration would have to capture them and then republish the events when the configuration is stopped, which is similar in concept to what SLF4J is doing. But we can’t rely on SLF4J’s substitute loggers for this as SLF4J may not even be in the picture.

So as I see it, the discussion on whether to implement the support for the new SLF4J method is completely separate from enhancing support of Log4j logging during startup.


On Apr 23, 2016, at 11:16 AM, Ceki Gulcu <ceki@qos.ch> wrote:

SLF4J will replay the events it captured during the initialization of the underlying logging framework, log4j2 in this case, assuming log4j2 supports SLF4J replay via the Logger.log(o.s.event.LoggingEvent) method.

What Raplh seems to suggest is to duplicate the SLF4J replay functionality within log4j2. However, given that SLF4J captures events generated during initialization, log4j2 cannot see them unless it supports SLF4J replay. For logging frameworks without replay support, i.e. those which lack the Logger.log(o.s.event.LoggingEvent) method, slf4j will simply drop the events it captured. Note that log4j1, logback prior to 1.1.4, and slf4j-jul and slf4j-simple prior to version 1.7.15 all lacked replay support but continue to work just fine under slf4j-api 1.7.15+.

Also note that if N is the version of slf4j-api and M is the slf4j-api version with which a binding is compiled, then slf4j ensures binary compatibility for all N and M in the 1.6.x and the 1.7.x series. For example, the slf4j-api-1.6.0.jar/slf4j-simple-1.7.21.jar combination will work just fine, same goes for the slf4j-simple-1.7.21.jar/slf4j-api-1.6.0.jar combination.

Compatibility is broken when the binding voluntarily supports event replay, as was the case with logback versions 1.1.4+ which require SLF4J version 1.7.15+. The question is whether log4j2 wants to support event replay at the cost of dropping compatibility with earlier versions of slf4j-api.

I hope this further clarifies the matter,


On 4/23/2016 19:00, Ralph Goers wrote:
I suppose that depends on the definition of “event loss”.  You can’t log
without a LoggerContext and the LoggerContext is initialized with a
default configuration, which means errors will be logged to the console.
We could create a default configuration that buffers the events and logs
then again when stop is called.


On Apr 23, 2016, at 7:18 AM, Ceki Gulcu <ceki@qos.ch
<mailto:ceki@qos.ch>> wrote:

You are welcome.

In the principle, the  event loss issue (fixed by event replay
post-initialization) depends on the time it takes for the underlying
implementation to initialize. Unless you can guarantee that log4j2
initializes instantly**, during SLF4J initialization event loss will
occur with log4j2 as well. Note the event loss issue is limited to
applications using slf4j which are multi-threaded early on.

> Would you mind sharing which applications perform reflection on
> the org.slf4j.Logger implementation (and why)?

The typical example is Spring which may do reflection on logger
instances. If the org.slg4j.Logger implementation offers the
log(org.slf4j.event.LoggingEvent) method, then the
org.slf4j.event.LoggingEvent interface must exist on the class path.
Otherwise, the Spring initialization will fail. The
org.slf4j.event.LoggingEvent interface was introduced in slf4j-api
version 1.7.15. It follows that if you decide to make use of the slf4j
replay feature, then log42 will de facto depend on slf4j version
1.7.15 both at compile and runtime.


**I am presuming here that log4j2 initializes when it creates and
returns its first Logger.

On 4/23/2016 15:54, Remko Popma wrote:
...and thank you for the info, Ceki, that is certainly helpful!


On Saturday, 23 April 2016, Remko Popma <remko.popma@gmail.com
<mailto:remko.popma@gmail.com>> wrote:

  Question: does the replay support solve a problem in SLF4J or in
  Logback initialization?

  If the latter, then perhaps there's no need to implement the new
  method since log4j-slf4j-impl binds to log4j2 anyway, no?


  Sent from my iPhone

   > On 2016/04/23, at 22:12, Ceki Gulcu <ceki@qos.ch
<mailto:ceki@qos.ch> <javascript:;>>
   > Hello,
   > For your information, slf4j-api version 1.7.21 will work with
  current versions of log4j2 just fine albeit without replay support.
  For replay support, log4j2's implementation of org.slf4j.Logger
  interface needs to have a method with the signature
  log(org.slf4j.event.LoggingEvent) in which case events generated
  during SLF4J initialization will be replayed. Note that if log4j2's
  Logger implementation chooses to implement the aforementioned log
  method, log42 will de facto depend on SLF4J version 1.7.15 and later
  both at compile and *runtime*.
   > The runtime dependency might seem surprising but some
  applications perform reflection on the org.slf4j.Logger
  implementation which will fail without slf4j-api 1.7.15 or later
  being present on the classpath.
   > I hope this helps,
   > --
   > Ceki
   >> On 4/23/2016 14:47, Ralph Goers wrote:
   >> We are not fully compatible with this version of SLF4J.  It
had some
   >> initialization race conditions so Substitute Loggers were
   >> Anyway, there is a new interface we need to implement.
   >> Ralph
   >> Begin forwarded message:
   >>> *From:* "Remko Popma (JIRA)" <jira@apache.org
   >>> *Date:* April 23, 2016 at 4:13:12 AM MST
   >>> *To:*log4j-dev@logging.apache.org
   >>> *Subject:* *[jira] [Closed] (LOG4J2-1375) Update SLF4J from
  1.7.13 to
   >>> 1.7.21*
   >>> *Reply-To:* "Log4J Developers List"
   >>> <mailto:log4j-dev@logging.apache.org<javascript:;>>>
   >>>    [
   >>> ]
   >>> Remko Popma closed LOG4J2-1375.
   >>> -------------------------------
   >>>   Resolution: Fixed
   >>>> Update SLF4J from 1.7.13 to 1.7.21
   >>>> ----------------------------------
   >>>>               Key: LOG4J2-1375
   >>>>               URL:
   >>>>           Project: Log4j 2
   >>>>        Issue Type: Improvement
   >>>>        Components: SLF4J Bridge
   >>>>  Affects Versions: 2.5
   >>>>          Reporter: Remko Popma
   >>>>          Assignee: Remko Popma
   >>>>           Fix For: 2.6
   >>>> Update SLF4J from 1.7.13 to 1.7.21
   >>> --
   >>> This message was sent by Atlassian JIRA
   >>> (v6.3.4#6332)
   >>> To unsubscribe, e-mail:
   >>> <mailto:log4j-dev-unsubscribe@logging.apache.org<javascript:;>>
   >>> For additional commands, e-mail:
   >>> <mailto:log4j-dev-help@logging.apache.org<javascript:;>>
   > To unsubscribe, e-mail:log4j-dev-unsubscribe@logging.apache.org
   > For additional commands, e-mail:

To unsubscribe, e-mail:log4j-dev-unsubscribe@logging.apache.org
For additional commands, e-mail:log4j-dev-help@logging.apache.org

To unsubscribe, e-mail: log4j-dev-unsubscribe@logging.apache.org
For additional commands, e-mail: log4j-dev-help@logging.apache.org