[
https://issues.apache.org/jira/browse/LOG4J2-3621?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17621321#comment-17621321
]
Piotr Karwasz commented on LOG4J2-3621:
---------------------------------------
Hi [~adwsingh],
In my opinion a breaking change occurs only if the documented order changes.
That is what happened between {{2.17.2}} and {{2.18.0}}: the documented order
changed from "environment variables < log4j2.component.properties < system
properties" to "system properties < environment variables <
log4j2.component.properties". Changes that fix the code to behave as documented
are just bug fixes.
So if you change from "system properties < environment variables < legacy
system properties" to "system properties < legacy system properties <
environment variables", that is a bug fix. If someone relies on the bug (as you
did in 2.17.2 when the doc said the {{LOG4J_CONFIGURATION_FILE}} overrides
{{log4j.configurationFile}}), he will need to change his configuration.
Regarding the token-based cache: in 2.17.2 it was necessary, because the
normalized cache was broken. The Log4j2 code always performs lookups using
*legacy* property names, while the normalized cache used normalized keys. A
lookup for {{log4j.configurationFile}} in the normalized cache was always a
miss and the code would fall back on the token-based cache.
This has been fixed in 2.18.0: the normalized cache uses legacy names as keys
and values corresponding to normalized keys as values. The token-based cache
should not be necessary.
That said: if you want to keep the preemptive caching of all properties, I
would keep also the token-based cache. If you want to remove the preemptive
cache and cache property values when they are looked up for the first time I
think we can forget about tokens.
Remark: most of the property lookups return {{null}}, so caching both positive
and negative results would be a performance improvement. E.g. we can use a
{{Map<String, Optional<String>>}}.
> Log4J 2.19 breaks contract of order of loading of System Properties
> -------------------------------------------------------------------
>
> Key: LOG4J2-3621
> URL: https://issues.apache.org/jira/browse/LOG4J2-3621
> Project: Log4j 2
> Issue Type: Bug
> Components: Configuration
> Affects Versions: 2.19.0, 2.19.1
> Reporter: Adwait Kumar Singh
> Priority: Major
>
> [This change|https://github.com/apache/logging-log4j2/pull/742] broke one of
> our systems on upgrading to 2.19.
> In our applications we had both LOG4J_CONFIGURATION_FILE environment variable
> as well as log4j.configurationFile system property set.
> In version 2.17.2 "log4j.configurationFile" gets priority vs in 2.19
> "LOG4J_CONFIGURATION_FILE" gets priority. This also breaks the contract
> mentioned in the
> [documentation|https://logging.apache.org/log4j/2.x/manual/configuration.html#SystemProperties].
> This is happening because of the normalization code here,
> [https://github.com/apache/logging-log4j2/blob/release-2.x/log4j-api/src/main/java/org/apache/logging/log4j/util/PropertiesUtil.java#L503-L526]
>
> When we are trying to normalize, we are checking if the source contains the
> normalKey. In case both log4j.configurationFile and LOG4J_CONFIGURATION_FILE
> are present, the following sequence happens,
> # log4j.configurationFile does not get inserted into the normalized map
> because the normal key is log4j2.configurationFile which is not present in
> the SystemPropertiesSource.
> # Then when we hit the EnvironmentPropertiesSource, log4j.configurationFile
> is normalized to LOG4J_CONFIGURATION_FILE and then an entry is made in the
> normalized map with key = log4j.configurationFile, but value of
> LOG4J_CONFIGURATION_FILE.
> # During look up with first look into normalized map, so now we got the
> wrong value (EnvironmentVariable instead of SystemProperty).
>
> I am aware changing "log4j.configurationFile" to "log4j2.configurationFile"
> can fix the issue, however this is clearly a backward incompatible change
> which will require this change across a lot of consumers who want to upgrade
> to log4j 2.19
> I can think of two ways to fix this,
> # We make an entry into the normalized map with the actual key if the
> normalized key is not present in the source, OR
> # While fetching we prefer literal map over normalized map.
> Would like to know which of the approaches would be better, can raise a PR
> accordingly.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
