In our tests you have to explicitly shut down watcher rather than shut
down the node it is running on because of thread leak detection. Just
shutting down the node that it is running on will cause it to start up
on another node if there is another one running and then not properly
shut down. This is probably something that should be fixed in watcher
somehow but for now lets just be more careful with the tests.
Closeselastic/elasticsearch#2365Closeselastic/elasticsearch#2588
Original commit: elastic/x-pack-elasticsearch@fb8a172972
Previously core Elasticsearch had methods in a test class for removing
and adding appenders. However, these methods were moved to production
code. This commit adjusts x-plugins for this change.
Original commit: elastic/x-pack-elasticsearch@83e37ef65a
This particular change focuses on upgrading the source of a watch when it comes to scripts that have no language specified explicitly.
The default language in version 5 changed to painless from whatever is specified in `script.default_lang` setting (this defaulted to groovy). In order to make sure that scripts in watcher remain to work we should rewrite the search source upon startup and set the legacy default language explicitly. The legacy script language is now controlled by `script.legacy.default_lang` setting and that defaults to groovy.
Changing the source upon startup should do the trick and only change the source of watches with scripts that don't have an explicit language set. For new watches the default language used in scripts is painless and because we now always serialize the language explicitly in scripts these watches won't be changed on startup.
The upgrade logic added here tries to upgrade scripts in the following places in a watch:
* script condition
* script transform
* any script defined inside of a search input
Original commit: elastic/x-pack-elasticsearch@4d578819eb
This commit cleans most of the methods of XContentBuilder so that:
- Jackson's convenience methods are used instead of our custom ones (ie field(String,long) now uses Jackson's writeNumberField(String, long) instead of calling writeField(String) then writeNumber(long))
- null checks are added for all field names and values
- methods are grouped by type in the class source
- methods have the same parameters names
- duplicated methods like field(String, String...) and array(String, String...) are removed
- varargs methods now have the "array" name to reflect that it builds arrays
- unused methods like field(String,BigDecimal) are removed
- all methods now follow the execution path: field(String,?) -> field(String) then value(?), and value(?) -> writeSomething() method. Methods to build arrays also follow the same execution path.
Original commit: elastic/x-pack-elasticsearch@d83f3aa6e2
This fixes a bug I found with a customer when he updated from 1.x to 2.x.
Due to an BWC incompatible change in the watch history mapping and a thread
pool rejection during execution a watch was not removed from the triggered
watches and tried to be executed again.
While trying to fix it it turned out that the execution of the failure
test case was still done in the transport thread and thus required some
offloading to another thread pool.
Original commit: elastic/x-pack-elasticsearch@df04ce31f2
If someone deletes the watch index (i.e. by deleting all indices), the watcher
in memory store still contains all the watches and tries to execute watches -
which results in exceptions as the watch itself cannot be updated anymore.
In order to minimize this problem (it cant be get rid of completely), we should
act accordingly if the watch index goes missing (either deleted or closed) and
clear out the memory representation of watches in the watchstore as well as trying
to finish all the current executions.
Closeselastic/elasticsearch#2794
Original commit: elastic/x-pack-elasticsearch@12d98cd566