We recently run into two issues where mapping weren't serialized in a consistent manner (#10302 and #10318). We rely on this consistency to do a byte level checl that mappings we get from the master are indentical to the one we have locally. Mistakes here can cause endless refresh mapping loops.
This commit adds an assert that verifies this upon every update from the master.
Setting the number of replicas on a closed index can leave the index
in an unopenable state since we might not be able to recover a quorum.
This commit simply prevents updating this setting on a closed index.
Closes#9566
The ShardFilterCache relies on the fact that it's
closed once the last reader on the shard is closed.
This is only guaranteed once the Store and all its
references are closed. This commit moves the closing
into the internal callback mechanism we use for deleting
shard data etc. to close the cache once we have all
searchers released.
fails
In the event that reading from the latest commit fails, we should fall
back to reading from the `Store` using the traditional
`Directory.listAll()`
Related to #11361
In #9474, we exported the hostname in the bin/elasticsearch scripts so that it
could be used as a variable in the elasticsearch.yml file but did not do the same
for plugin manager. When using the hostname variable in elasticsearch.yml and
trying to use the plugin manager, initialization will fail because the property could
not be resolved. This change will allow the hostname to be resolved in the same
manner as the service scripts.
Closes#10902
In #11072 we are adding a check that will prevent opening of old indices. However, this check doesn't take into consideration the fact that indices can be made compatible with the current version through upgrade API. In order to make compatibility check aware of the upgrade, the upgrade API should write a new setting `index.version.minimum_compatible` that will indicate the minimum compatible version of lucene this index is compatible with and `index.version.upgraded` that will indicate the version of elasticsearch that performed the upgrade.
Closes#11095
This commit adds a shell script which:
* move current elasticsearch core source in `/core`
* fetch `elasticsearch-parent` project in `/`
* fetch plugins in `/plugins`
* change `groupId` for all plugins to `org.elasticsearch.plugin` so versions won't conflict in maven central
* remove plugins/name/dev-tools dir which is not needed anymore
* remove plugins/name/.git dir
* remove plugins/name/LICENSE and plugins/name/CONTRIBUTING files
* clean `core/pom.xml` of useless settings that are inherited from parent project.
* `core/pom.xml` is adapted to change location of rest tests definition (`../`)
* change core name to `Elasticsearch Core`
* remove `plugins` dir from `.gitignore`
Plugins added:
* Analysis
* analysis-kuromoji
* analysis-smartcn
* analysis-stempel
* analysis-phonetic
* analysis-icu
* Mapper
* mapper-attachments
* Language
* lang-python
* lang-mvel
* lang-javascript
* Cloud
* cloud-gce
* cloud-azure
* cloud-aws
River plugins are ignored but might be added if we want to.
Todo:
* check and adapt our release tool. It now has to upload all submodules as well.
* adapt Jenkins jobs
The wildcard cat API REST tests relied on bulk.max and bulk.min in
the thread_pool response. However due to the thread pool types being
randomized in InternalTestCluster, the min/max values were not guaranteed
to exist (the cached thread pool type is unbounded and thus does not have a
max value).
In order to prevent this, the test has been removed and now the cat
nodes test is used for wildcard testing, which always returns stats
about the heap.
The Directory#makeLock API is trappy and can easily lead to unexpected
lock release if native locks are used. see LUCENE-6507 for details.
This commit consolidates the lock lock into one place and only returns
the lock instance if we actually acquired it.
Aggregations like Sampler and TopHits that require access to scores did not work if the query has size param set to zero. The assumption was that the Lucene query scoring logic was not required in these cases.
Added a Junit test to demonstrate the issue and a fix which relies on earlier creation of Collector wrappers so that Collector.needsScores() calls work for all search operations.
Closes#11119
On Windows, colons ':' are illegal in file names and since we use a Path to
check if the test is blacklisted, tests with a colon in the test section name
will fail. This change simply removes the colon from the name when matching
against the blacklist.
This catches UnsatifiedLinkError when attempting to load the JNA Native
class, in cases where there are error loading the native libraries that JNA
needs to function.
Today, JNA is a optional dependency in the build but when running tests or running
with mlockall set to true, JNA must be on the classpath for Windows systems since
we always try to load JNA classes when using mlockall.
The old Natives class was renamed to JNANatives, and a new Natives class is
introduced without any direct imports on JNA classes. The Natives class checks to
see if JNA classes are available at startup. If the classes are available the Natives
class will delegate to the JNANatives class. If the classes are not available the
Natives class will not use the JNANatives class, which results in no additional attempts
to load JNA classes.
Additionally, all of the JNA classes were moved to the bootstrap package and made
package private as this is the only place they should be called from.
Closes#11360
The tests for the recently added added wildcard feature were
relying on order of the hashmap being used, which could be
different.
The implementation now ensures, that the header fields are
parsed in the order they have been added.
When we highlight on fields using wildcards then fields might match that cannot
be highlighted by the specified highlighter. The whole search request then
failed. Instead, check that the field can be highlighted and ignore the field
if it can't.
In addition ignore the exception thrown by plain highlighter if a field conatins
terms larger than 32766.
closes#9881
The Moving average predict code generated incorrect keys if the key for the first bucket of the histogram was < 0. This fix makes the moving average use the rounding class from the histogram to generate the keys for the new buckets.
Closes#11369
Britta Weber
Boaz Leskes
Simon Willnauer
Michael McCandless
Adrien Grand
Alexander Reelsen
Igor Motov
Lee Hinman
jaymode
David Pilato
markharwood
javanna
Robert Muir
Zachary Tong
Clinton Gormley
Colin Goodheart-Smithe
Tanguy Leroux