The equalTo logic of ShardRouting doesn't take version and unassignedInfo into the account when compares shard routings. Since cluster state diff relies on equal to detect the changes that needs to be sent to other cluster, this omission might lead to changes not being properly propagated to other nodes in the cluster.
Closes#12387
In order to test the way plugins with configuration are installed and removed
we need a plugin with configuration in the repository. The simplest way to
get one is to make an "example" plugin.
In the process of making this example it became aparent that cat actions were
difficult to create outside of the org.elasticsearch.rest.action.cat package
because key methods in AbstractCatAction were package private. This makes them
protected and uses them to create the example configured plugin.
Relates to #12717 but is only one step of many to close it.
the default classloader. It had all kinds of leniency in how the
classname was found, and simply cannot work with plugins having isolated
classloaders.
This change removes that method. Some of the uses of it were for custom
extension points, like custom repository or discovery types. A lot were
just there to plugin mock implementations for tests. For the settings
that were legitimate, all now support plugins adding the given setting
via onModule. For those that were specific to tests for mocks, they now
use Classes.loadClass (a helper around Class.forName). This is a
temporary measure until (in a future PR) tests can change the
implementation via package private statics.
I also removed a number of unnecessary intermediate modules, added a
"jvm-example" plugin that can be filled in in the future as a smoke test
for breaking plugins, and gave some documentation to "spawn" modules
interface.
closes#12643closes#12656
This was a straight up bug found in #12753. If only one type existed,
the compatibility check for a new type was not strict, so changes to
an updateable setting like search_analyzer got through (but only
partially). This change fixes the check and adds tests (which were
previously a TODO).
This also fixes a bug in dynamic field creation which woudln't copy
fielddata settings when duplicating a pre-existing field with the
same name.
closes#12753
This PR prevents setting timezone on ValueFormatter.DateTime. Instead
the timezone information needed when printing buckets key-as-string
information is provided at constrution time of the ValueFormatter, making
sure we don't overwrite any constants. This, however, made it necessary to
be able to access the timezone information when resolving the format
in ValueSourseParser, so the `time_zone` parameter is now parsed alongside
the `format` parameter in ValueSourceParser rather than in DateHistogramParser.
Closes#12531
Today we accept store types like `nio_fs`, `nioFs`, `niofs` etc.
this commit removes the leniency and only accepts plain values without underscore.
Yet, this commit also has a BWC layer that upgrades existing indices to the new version.
Affected enums are:
* `nio_fs`
* `mmap_fs`
* `simple_fs`
If an URL specified with --url on the command line cannot be reached,
the plugin manager tries URLs at download.elastic.co automatically.
This can lead to download wrong plugins and also exposes potentially
the name of an internal plugin to the download service.
This fix ensures, that the plugin manager simply aborts, if the specified
URL cannot be downloaded.
Moved the license checker config into the parent pom, and overrede
the license dir/target-to-check in distributions/pom.
Disabled the license checker explicitly for projects which run integration
tests but have no licenses dir:
* core
* distribution
* qa
* plugins/delete-by-query
* plugins/mapper-size
* plugins/site-example
Closes#12752Closes#12754
This commit adds basic support to track the number of times scripts are
compiled and compiled scripts are evicted from the script cache. These
statistics are tracked at the node level.
Closes#12673
ClusterState has 3 different methods to access RoutingNodes:
* #routingNodes() - mutable version
* #getRoutingNodes() - delegates to #getReadOnlyRoutingNodes()
* #getReadOnlyRoutingNodes() - it's docs say `NOTE, the routing nodes are mutable, use them just for read operations`
The latter also reuses the instance that it creates. This has several problems beside the obvious:
* creating RoutingNodes is costly and should be done only if really needed ie. use cached version as much as possible
* the common case is ReadOnly but all kinds of things are called
* mutable version are only needed in one place and should only be used in the AllocationService
* RoutingNodes can freeze it's ShardRoutings but doesn't
* RoutingNodes should check if it's read-only or not
This commit fixed all the problems and special cases the mutable case such that all accesses via ClusterState#getRoutingNodes()
is read-only and RoutingNodes enforces this.
This is one of our esoteric metadata mappers so I think we should distribute
it in a plugin rather than in elasticsearch core.
This introduces one limitation: the value of the `_size` parameter is not
retrievable for documents that are only in the transaction log.
The `multi_match` query groups terms that have the same analyzer together and
then applies the boost of the first query in each group. This is not necessary
given that boosts for each term are already applied another way.
Versions can be tricky with linux vendors and such. To help debug any possible issues, we should output a better version.
Today:
```
[elasticsearch] java.lang.RuntimeException: Java version: 1.7.0_55 suffers from critical bug https://bugs.openjdk.java.net/browse/JDK-8024830 which can cause data corruption.
[elasticsearch] Please upgrade the JVM, see http://www.elastic.co/guide/en/elasticsearch/reference/current/_installation.html for current recommendations.
[elasticsearch] If you absolutely cannot upgrade, please add -XX:-UseSuperWord to the JAVA_OPTS environment variable.
[elasticsearch] Upgrading is preferred, this workaround will result in degraded performance.
[elasticsearch] at org.elasticsearch.bootstrap.JVMCheck.check(JVMCheck.java:121)
[elasticsearch] at org.elasticsearch.bootstrap.Bootstrap.main(Bootstrap.java:270)
[elasticsearch] at org.elasticsearch.bootstrap.Elasticsearch.main(Elasticsearch.java:28)
```
With patch:
```
java.lang.RuntimeException: Java version: Oracle Corporation 1.7.0_40 [Java HotSpot(TM) 64-Bit Server VM 24.0-b56] suffers from critical bug https://bugs.openjdk.java.net/browse/JDK-8024830 which can cause data corruption.
Please upgrade the JVM, see http://www.elastic.co/guide/en/elasticsearch/reference/current/_installation.html for current recommendations.
If you absolutely cannot upgrade, please add -XX:-UseSuperWord to the JAVA_OPTS environment variable.
Upgrading is preferred, this workaround will result in degraded performance.
at org.elasticsearch.bootstrap.JVMCheck.check(JVMCheck.java:121)
at org.elasticsearch.bootstrap.Bootstrap.main(Bootstrap.java:270)
at org.elasticsearch.bootstrap.Elasticsearch.main(Elasticsearch.java:28)
```
This commit adds a new API to allow scripts to say whether they need scores.
In practice, only the `expression` script engine makes use of it correctly,
other engines just return `true` since they can't predict whether they'll
need scores. This should make scripted aggregations and `function_query`
faster as we'll now be able to pass needsScores=false to Query.createWeight.
Adding a named query that is null can lead to a NullPointerException
when copying the named queries. This is due to an implementation detail
in QueryParseContent.copyNamedQueries. In particular, this method uses
com.google.common.collect.ImmutableMap.copyOf. A documented requirement
of ImmutableMap is that none of the entries have a null key nor null
value. Therefore, we should not add such queries to the namedQueries
map. This will not change any behavior since Map.get returns null if no
entry with the given key exists anyway.
Closes#12683
This change improves the `function_score` query to not compute scores at all
when they are not needed, and to not compute scores on the underlying query
when the combine function is to replace the score with the scores of the
functions.
This commit makes it possible to serialize arbitrary objects by having them extend Writeable. When reading them though, we need to be able to identify which object we have to create, based on its name. This is useful for queries once we move to parsing on the coordinating node, as well as with aggregations and so on.
Introduced a new abstraction called NamedWriteable, which is supported by StreamOutput and StreamInput through writeNamedWriteable and readNamedWriteable methods. A new NamedWriteableRegistry is introduced also where named writeable prototypes need to be registered so that we are able to retrieve the proper instance of the writeable given its name and then de-serialize it calling readFrom against it.
Closes#12393
In order to avoid extra reroutes, `RoutingService` should avoid
scheduling a reroute of any shards where the delay is negative. To make
sure that we don't encounter a race condition between the
GatewayAllocator thinking a shard is delayed and RoutingService thinking
it is not, the GatewayAllocator will update the RoutingService with the
last time it checked in order to use a consistent "view" of the delay.
Resolves#12456
Relates to #12515 and #12456
When postFilter generates a token that is identical to the input term
DirectCandidateGenerator should not preFilter this token. If postFilter
and preFilter are the same analyzer instance it would fail with :
"TokenStream contract violation: close() call missing"
This is a forward port of @nomoa's #12670
Today we try to detect if there is an index existing in the directory
and if not we create one. This can be tricky and errof prone since we
rely on the filesystem without taking the context into account when the
engine gets created. We know in all situations if the index should be created
so we can just use this infromation and rely on the lucene index writer to barf
if we hit a situations where we can't append to an index while we should.
Today we miss to throw / rethrow an recovery exception if it happens during
the finalization of phase 1 if the source files are not affected. Even worse
this can cause some dataloss if the reason for this exception is a failure of
deleting a corruption marker or similar pre-existing corruptions since we continue
with the recovery and mark the target shared as started which will in-turn open
an engine with an empty index.
This makes the output of EsThreadPoolExecutor#toString less pretty but
we no longer have funky hacky that rely on the specific format of the
toString produced by ThreadPoolExecutor which isn't part of its API and
could change with any JVM version and break the output.
Improving the toString allows for nicer error reporting. Also cleaned up
the way that EsRejectedExecutionException notices that it was rejected
from a shutdown thread pool. I left javadocs about how its not 100% correct
but good enough for most uses.
The improved toString on EsThreadPoolExecutor mean every one of them needs
a name. In most cases the name to use is obvious. In tests I use the name
of the test method and in real thread pools I use the name of the thread
pool. In non-ThreadPool executors I use the thread's name.
Closes#9732
In order to support the URL notation including a user/pass combination
(like http://user:pass@host/plugin.zip) the auth info needs to be added
manually.
Today this is "unofficial" as conf/scripts, but some people
want to share scripts across different nodes and so on. Because
they cannot configure it, they are forced to use dirty hacks
like symbolic links, which isnt going to work: we aren't going
to recursively scan conf/ and add permissions to all link targets
underneath it, thats crazy.
I really hate adding yet another configuration knob here, but
users resorting to using symlinks are going to be frustrated,
and do things in a more insecure way.
The URL to download the main elasticsearch plugins did not match
what the S3 wageon is supposed to write to.
In addition this PR adds support for snapshots to access the
snapshot S3 bucket, so we can possibly download snapshot versions
of plugins.
The format of the URLs stems from #12270Closes#12632
Conflicting mappings that were allowed before v2.0 can cause runaway shard failures on upgrade. This commit adds a check that prevents a cluster from starting if it contains such indices as well as restoring such indices from a snapshot into already running cluster.
Closes#11857
Currently when we delete files belonging to deleted snapshots we issue one delete command to underlying snapshot store at a time. Some repositories can benefit from bulk deletes of multiple files.
Closes#12533
This method syncs the translog unless it's already synced. If the engine
is alreayd closed we are guaranteed to be synced already such that we can just
ignore this exception.
Closes#12603
this change was added recently which uses default timezone for the creation
date on CAT endpoints. We should be consistent and use UTC across the board.
This commit adds #getDefaultTimzone() to forbidden API and fixes the REST tests.
Relates to #11688
The Streams.copyTo(String|Bytes)FromClasspath() methods resolve resources using org.elasticsearch.io.Streams classloader. This is fine in elasticsearch core and when running tests but if used in a plugin this can lead to FileNotFoundExceptions at runtime because plugin are loaded in a dedicated classloader.
Most of the abstract base test classes we have were previously @Ignored.
However, there were also some other tests ignored. Having two ways to
quiet tests is confusing, and clearly it has caused some tests
to get lost in the fold.
This change moves all base test classes to use the "TestCase" suffix,
which is not picked up by the test class name pattern. It also removes
@Ignore from (almost) all tests, and adds it to forbidden apis.
And since we were renaming, I shorted base test class names to use
"ES" instead of "Elasticsearch". I type this a lot of types a day,
and I have heard others express a similar desire for a shorter name.
closes#10659
Now that integ tests are moved into `mvn verify`, we don't really have
a need for @Slow, and especially not @Integration. This removes
uses of the first, and completely removes uses of the latter.
Tasks can be registered with a timeout, which runs as a task in a separate
threadpool. The idea is that the timeout runner cancels the main task when
the time is out, and the timeout runner is cancelled when the main task
starts executing. However, the following statement:
```java
timeoutFuture = timer.schedule(new Runnable() {
@Override
public void run() {
if (remove(TieBreakingPrioritizedRunnable.this)) {
runAndClean(timeoutCallback);
}
}
}, timeValue.nanos(), TimeUnit.NANOSECONDS);
```
is not atomic: the removal task is first started, and then the (volatile)
variable is assigned. As a consequence, there is a short window that allows
a timeout task to wait until the time is out even if the task is already
completed.
See http://build-us-00.elastic.co/job/es_core_17_centos/496/ for an example of
such a failure.
Previously only the first aggregation in a buckets_path was check to make sure the aggregation existed. Now the whole path is checked to ensure an aggregation exists at each element in the buckets_path
Closes#12360
Its not going to work: its blocked by security policy
and will just add a confusing SecurityException to the mix, and
bogusly give an exit status of 0 when in fact something bad happened.
Finally, if ES can't startup, it is a serious problem, there is
no sense in hiding the reason why: deliver the full stack trace.
The repository verification process should create a subdirectory to make sure we check permission of newly created directories in case elasticsearch processes on different nodes are running using different uids and creating blobs with incompatible permissions.
Closes#11611
Previously we issued a reroute when a node went over the high watermark
in order to move shards away from the node. This change tracks nodes
that have previously been over the high or low watermarks and issues a
reroute when the node goes back underneath the watermark.
This allows shards that may be unassigned to be assigned back to a node
that was previously over the low watermark but no longer is.
Resolves#12422
As windows has different line endings and this has
already been fixed by another test, the method has been
moved into CliToolTestCase.
In addition one test has been removed, as it was redundant.
In order to ensure, we have the same experience across operating systems
and shells, this commit uses the java CLI parser instead of the shell
getopt parsing to parse arguments.
This also allows for support for paths, which contain spaces.
Also commons-cli depdency was upgraded to 1.3.1 and tests have been added.
Changes
* new exit code, OK_AND_EXIT, allowing to tell the caller to exit, as everything
went as expected (e.g. when running a version output)
BWC breaking:
* execute() returns an ExitStatus instead of an integer, otherwise there is no
possibility to signal by a command, if the JVM should be exited after a run.
This affects plugins, that have command line tools
* -v used to be version, but is a verbose flag by default in the current CLI infra,
must be -V or --version now
* -X has been removed - the current implementation was useless anyway, as
it prefixed those properties with "es.". You should use
ES_JAVA_OPTS/JAVA_OPTS for JVM configuration
Date math index name resolution enables you to search a range of time-series indices, rather than searching all of your time-series indices and filtering the the results or maintaining aliases. Limiting the number of indices that are searched reduces the load on the cluster and improves execution performance. For example, if you are searching for errors in your daily logs, you can use a date math name template to restrict the search to the past two days.
The added `ExpressionResolver` implementation that is responsible for resolving date math expressions in index names. This resolver is evaluated before wildcard expressions are evaluated.
The supported format: `<static_name{date_math_expr{date_format|timezone_id}}>` and the date math expressions must be enclosed within angle brackets. The `date_format` is optional and defaults to `YYYY.MM.dd`. The `timezone_id` id is optional too and defaults to `utc`.
The `{` character can be escaped by places `\\` before it.
Closes#12059
When we rewrite to a MatchNoTermsQuery we were throwing out the boost which
could could lead to funky changes when the query against _all was in a
bool query.
Previously we would write the entire ByteBuffer to the stream to serialise the HDRHistogram even if it was not all needed. Now we only write the bytes that are actually written to in the ByteBuffer.
I suggest updating the BulkProcessor so that it prevents users from building a BulkProcessor with a null client. If the client is null, then the class should throw an exception with a message that describes the cause of the exception. Earlier today, when I passed a null client to the builder() method, later received an exception, and attempted to get the exception's message in my afterBulk() listener (e.g. e.getMessage()), the message was null. It took me a while to pinpoint the cause of the problem.
This can happen in two ways:
1. The _all field is disabled.
2. There are documents in the index, the _all field is enabled, but there are
no fields in any of the documents.
In both of these cases we now rewrite the query to a MatchNoDocsQuery which
should be safe because there isn't anything to match.
Closes#12439
When a node discovers shard content on disk which isn't used, we reach out to all other nodes that supposed to have the shard active. Only once all of those have confirmed the shard active, the shard has no unassigned copies *and* no cluster state change have happened in the mean while, do we go and delete the shard folder.
Currently, after removing a shard, the IndicesStores checks the indices services if that has no more shard active for this index and if so, it tries to delete the entire index folder (unless on master node, where we keep the index metadata around). This is wrong as both the check and the protections in IndicesServices.deleteIndexStore make sure that there isn't any shard *in use* from that index. However, it may be the we erroneously delete other unused shard copies on disk, without the proper safety guards described above.
Normally, this is not a problem as the missing copy will be recovered from another shard copy on another node (although a shame). However, in extremely rare cases involving multiple node failures/restarts where all shard copies are not available (i.e., shard is red) there are race conditions which can cause all shard copies to be deleted.
Instead, we should change the decision to clean up an index folder to based on checking the index directory for being empty and containing no shards.
This change creates a proper `distribution` modules in which we have today packaging for
all of our four current packages:
* zip
* tar.gz
* rpm
* deb
Licenes have moved into the distribution project as well. So have the config/ and the bin/ directory
from the core/ project.
The RPM package is now built, if rpmbuild exists.
The bats tests have been moved as well.
Also the zip distribution now executes the REST integration tests.
As Robert pointed out on #12465, it has the undesirable property of relying on
the operating system. So it would be better to use a simple rule such as
checking whether the file name starts with a dot.
When an index is opened it will not be assigned to a node but also not have closed state
anymore. Before we only checked if an index either is closed or assigned to the data node
and therefore the change from close->open was not written.
Some of the test for meta data are redundant. Also, since they
somewhat test service disruptions (start master with empty
data folder) we might move them to DiscoveryWithServiceDisruptionsTests.
Also, this commit adds a test for
https://github.com/elastic/elasticsearch/issues/11665
If we cache the type filter and we e.g. set its boost which is now settable on all queries, the boost will change for all subsequent queries. We should rather create a new query every time.
Major changes:
* Changed MetaData to holds alias and index lookup information into a single TreeMap instead of two separate maps.
* Moved the building of the alias / index lookup to the metadata builder.
Settings are currently parsed by looping over the tokens until an END_OBJECT token is reached. However, this does not mean that the end of
the settings stream was reached. This can occur, for example, when parsing a YAML settings file with inconsistent indentation. Currently
in this case, some settings will be silently ignored. This commit forces a check that we have in fact reached the end of the settings
stream.
Closes#12382
HDRHistogram has been added as an option in the percentiles and percentile_ranks aggregation. It has one option `number_significant_digits` which controls the accuracy and memory size for the algorithm
Closes#8324
The ThrottlingAllocationDecider is responsible to limit the number of incoming/local recoveries on a node. It therefore shouldn't count shards marked as relocating which represent the source of the recovery.
Closes#12409
If there are multiple levels of filtered leaf readers then with the unwrap() method it immediately returns the most inner leaf reader and thus skipping of over any other filtered leaf reader that may be instance of ElasticsearchLeafReader. By using #getDelegate() method we can check each filter reader layer if it is instance of ElasticsearchLeafReader, so that we never skip over any wrapped filtered leaf reader and lose the shard id.
Also, as set of utility methods was introduced on ShardRouting to do various types of matching with other shard routings, giving control about what exactly should be matched (same shard id, same allocation id, all but version and shard info etc.). This is useful here, but also prepares the grounds for the change needed in #12387 (making ShardRouting.equals be strict and perform exact equality).
Closes#12397
When a replica is initializing from the primary, and we find a better node that has full sync id match, it is better to cancel the existing replica allocation and allocate it to the new node with sync id match (eventually)
Today, when a user provides settings and specifies a classloader to be used, the classloader gets
dropped when we copy the settings to check for prompt entries. This change copies the classloader
when replacing the prompt placeholders and adds a test to ensure the InternalSettingsPreparer
always retains the classloader.
Closes#12340
JarHell has a low level check, but its more of a best effort one,
only checking if X-Compile-Target-JDK is set in the manifest. This
is the case for all lucene- and elasticsearch- generated jars, but
lets just be explicit for plugins.
The TransportSingleCustomOperationAction `prefer_local` option has been removed as it isn't worth the effort.
The TransportSingleShardAction will execute the operation on the receiving node if a concrete list doesn't provide a list of candite shards routings to perform the operation on.
Change #12371 broke fielddata on the `_parent` child for indices created before
2.0. This pull request adds back caching of the `_parent` fielddata for indices
created before 2.0 and cleans some related stuff. For instance
DocumentTypeListener doesn't need to take care of removed mappings anymore since
mappings can't be removed in 2.0.
The rangeQuery() method in MappedFieldType and some overwriting subtypes takes
a nullable QueryParseContext argument which is unused and can be deleted.
This is also useful for the current query parsing refactoring, since
we want to avoid passing the context object around unnecessarily.
IndicesStoreIntegrationTests.indexCleanup() tests if the shard files on disk are actually removed
after relocation. In particular it tests the following:
Whenever a node deletes a shard because it was relocated somewhere else, it first
checks if enough other copies are started somewhere else. The node sends a
ShardActiveRequest to the nodes that should have a copy.
The nodes that receive this request check if the shard is in state STARTED in which case
they respond with true. If they have the shard in POST_RECOVERY they register a cluster state
observer that checks at each update if the shard has moved to STARTED and respond with true when
this happens.
To test that the cluster state observer mechanism actually works, the latter can be triggered by
blocking the cluster state processing when a recover starts and only unblocking it shortly
after the node receives the ShardActiveRequest.
This is more reliable than using random cluster state processing delays
because the random delays make it hard to reason about different timeouts
that can be reached.
closes#11989
Closes#12367
Squashed commit of the following:
commit 9453c411798121aa5439c52e95301f60a022ba5f
Merge: 3511a9c 828d8c7
Author: Robert Muir <rmuir@apache.org>
Date: Wed Jul 22 08:22:41 2015 -0400
Merge branch 'master' into refactor_pluginservice
commit 3511a9c616503c447de9f0df9b4e9db3e22abd58
Author: Ryan Ernst <ryan@iernst.net>
Date: Tue Jul 21 21:50:15 2015 -0700
Remove duplicated constant
commit 4a9b5b4621b0ef2e74c1e017d9c8cf624dd27713
Author: Ryan Ernst <ryan@iernst.net>
Date: Tue Jul 21 21:01:57 2015 -0700
Add check that plugin must specify at least site or jvm
commit 19aef2f0596153a549ef4b7f4483694de41e101b
Author: Ryan Ernst <ryan@iernst.net>
Date: Tue Jul 21 20:52:58 2015 -0700
Change plugin "plugin" property to "classname"
commit 07ae396f30ed592b7499a086adca72d3f327fe4c
Author: Robert Muir <rmuir@apache.org>
Date: Tue Jul 21 23:36:05 2015 -0400
remove test with no methods
commit 550e73bf3d0f94562f4dde95239409dc5a24ce25
Author: Robert Muir <rmuir@apache.org>
Date: Tue Jul 21 23:31:58 2015 -0400
fix loading to use classname
commit 04463aed12046da0da5cac2a24c3ace51a79f799
Author: Robert Muir <rmuir@apache.org>
Date: Tue Jul 21 23:24:19 2015 -0400
rename to classname
commit 9f3afadd1caf89448c2eb913757036da48758b2d
Author: Ryan Ernst <ryan@iernst.net>
Date: Tue Jul 21 20:18:46 2015 -0700
moved PluginInfo and refactored parsing from properties file
commit df63ccc1b8b7cc64d3e59d23f6c8e827825eba87
Author: Robert Muir <rmuir@apache.org>
Date: Tue Jul 21 23:08:26 2015 -0400
fix test
commit c7febd844be358707823186a8c7a2d21e37540c9
Author: Robert Muir <rmuir@apache.org>
Date: Tue Jul 21 23:03:44 2015 -0400
remove test
commit 017b3410cf9d2b7fca1b8653e6f1ebe2f2519257
Author: Robert Muir <rmuir@apache.org>
Date: Tue Jul 21 22:58:31 2015 -0400
fix test
commit c9922938df48041ad43bbb3ed6746f71bc846629
Merge: ad59af4 01ea89a
Author: Robert Muir <rmuir@apache.org>
Date: Tue Jul 21 22:37:28 2015 -0400
Merge branch 'master' into refactor_pluginservice
commit ad59af465e1f1ac58897e63e0c25fcce641148a7
Author: Areek Zillur <areek.zillur@elasticsearch.com>
Date: Tue Jul 21 19:30:26 2015 -0400
[TEST] Verify expected number of nodes in cluster before issuing shardStores request
commit f0f5a1e087255215b93656550fbc6bd89b8b3205
Author: Lee Hinman <lee@writequit.org>
Date: Tue Jul 21 11:27:28 2015 -0600
Ignore EngineClosedException during translog fysnc
When performing an operation on a primary, the state is captured and the
operation is performed on the primary shard. The original request is
then modified to increment the version of the operation as preparation
for it to be sent to the replicas.
If the request first fails on the primary during the translog sync
(because the Engine is already closed due to shadow primaries closing
the engine on relocation), then the operation is retried on the new primary
after being modified for the replica shards. It will then fail due to the
version being incorrect (the document does not yet exist but the request
expects a version of "1").
Order of operations:
- Request is executed against primary
- Request is modified (version incremented) so it can be sent to replicas
- Engine's translog is fsync'd if necessary (failing, and throwing an exception)
- Modified request is retried against new primary
This change ignores the exception where the engine is already closed
when syncing the translog (similar to how we ignore exceptions when
refreshing the shard if the ?refresh=true flag is used).
commit 4ac68bb1658688550ced0c4f479dee6d8b617777
Author: Shay Banon <kimchy@gmail.com>
Date: Tue Jul 21 22:37:29 2015 +0200
Replica allocator unit tests
First batch of unit tests to verify the behavior of replica allocator
commit 94609fc5943c8d85adc751b553847ab4cebe58a3
Author: Jason Tedor <jason@tedor.me>
Date: Tue Jul 21 14:04:46 2015 -0400
Correctly list blobs in Azure storage to prevent snapshot corruption and do not unnecessarily duplicate Lucene segments in Azure Storage
This commit addresses an issue that was leading to snapshot corruption for snapshots stored as blobs in Azure Storage.
The underlying issue is that in cases when multiple snapshots of an index were taken and persisted into Azure Storage, snapshots subsequent
to the first would repeatedly overwrite the snapshot files. This issue does render useless all snapshots except the final snapshot.
The root cause of this is due to String concatenation involving null. In particular, to list all of the blobs in a snapshot directory in
Azure the code would use the method listBlobsByPrefix where the prefix is null. In the listBlobsByPrefix method, the path keyPath + prefix
is constructed. However, per 5.1.11, 5.4 and 15.18.1 of the Java Language Specification, the reference null is first converted to the string
"null" before performing the concatenation. This leads to no blobs being returned and therefore the snapshot mechanism would operate as if
it were writing the first snapshot of the index. The fix is simply to check if prefix is null and handle the concatenation accordingly.
Upon fixing this issue so that subsequent snapshots would no longer overwrite earlier snapshots, it was discovered that the snapshot metadata
returned by the listBlobsByPrefix method was not sufficient for the snapshot layer to detect whether or not the Lucene segments had already
been copied to the Azure storage layer in an earlier snapshot. This led the snapshot layer to unnecessarily duplicate these Lucene segments
in Azure Storage.
The root cause of this is due to known behavior in the CloudBlobContainer.getBlockBlobReference method in the Azure API. Namely, this method
does not fetch blob attributes from Azure. As such, the lengths of all the blobs appeared to the snapshot layer to be of length zero and
therefore they would compare as not equal to any new blobs that the snapshot layer is going to persist. To remediate this, the method
CloudBlockBlob.downloadAttributes must be invoked. This will fetch the attributes from Azure Storage so that a proper comparison of the
blobs can be performed.
Closeselastic/elasticsearch-cloud-azure#51, closeselastic/elasticsearch-cloud-azure#99
commit cf1d481ce5dda0a45805e42f3b2e0e1e5d028b9e
Author: Lee Hinman <lee@writequit.org>
Date: Mon Jul 20 08:41:55 2015 -0600
Unit tests for `nodesAndVersions` on shared filesystems
With the `recover_on_any_node` setting, these unit tests check that the
correct node list and versions are returned.
commit 3c27cc32395c3624f7c794904d9ea4faf2eccbfb
Author: Robert Muir <rmuir@apache.org>
Date: Tue Jul 21 14:15:59 2015 -0400
don't fail junit4 integration tests if there are no tests.
instead fail the failsafe plugin, which means the external cluster will still get shut down
commit 95d2756c5a8c21a157fa844273fc83dfa3c00aea
Author: Alexander Reelsen <alexander@reelsen.net>
Date: Tue Jul 21 17:16:53 2015 +0200
Testing: Fix help displaying tests under windows
The help files are using a unix based file separator, where as
the test relies on the help being based on the file system separator.
This commit fixes the test to remove all `\r` characters before
comparing strings.
The test has also been moved into its own CliToolTestCase, as it does
not need to be an integration test.
commit 944f06ea36bd836f007f8eaade8f571d6140aad9
Author: Clinton Gormley <clint@traveljury.com>
Date: Tue Jul 21 18:04:52 2015 +0200
Refactored check_license_and_sha.pl to accept a license dir and package path
In preparation for the move to building the core zip, tar.gz, rpm, and deb as separate modules, refactored check_license_and_sha.pl to:
* accept a license dir and path to the package to check on the command line
* to be able to extract zip, tar.gz, deb, and rpm
* all packages except rpm will work on Windows
commit 2585431e8dfa5c82a2cc5b304cd03eee9bed7a4c
Author: Chris Earle <pickypg@users.noreply.github.com>
Date: Tue Jul 21 08:35:28 2015 -0700
Updating breaking changes
- field names cannot be mapped with `.` in them
- fixed asciidoc issue where the list was not recognized as a list
commit de299b9d3f4615b12e2226a1e2eff5a38ecaf15f
Author: Shay Banon <kimchy@gmail.com>
Date: Tue Jul 21 13:27:52 2015 +0200
Replace primaryPostAllocated flag and use UnassignedInfo
There is no need to maintain additional state as to if a primary was allocated post api creation on the index routing table, we hold all this information already in the UnassignedInfo class.
closes#12374
commit 43080bff40f60bedce5bdbc92df302f73aeb9cae
Author: Alexander Reelsen <alexander@reelsen.net>
Date: Tue Jul 21 15:45:05 2015 +0200
PluginManager: Fix bin/plugin calls in scripts/bats test
The release and smoke test python scripts used to install
plugins in the old fashion.
Also the BATS testing suite installed/removed plugins in that
way. Here the marvel tests have been removed, as marvel currently
does not work with the master branch.
In addition documentation has been updated as well, where it was
still missing.
commit b81ccba48993bc13c7678e6d979fd96998499233
Author: Boaz Leskes <b.leskes@gmail.com>
Date: Tue Jul 21 11:37:50 2015 +0200
Discovery: make sure NodeJoinController.ElectionCallback is always called from the update cluster state thread
This is important for correct handling of the joining thread. This causes assertions to trip in our test runs. See http://build-us-00.elastic.co/job/es_g1gc_master_metal/11653/ as an example
Closes#12372
commit 331853790bf29e34fb248ebc4c1ba585b44f5cab
Author: Boaz Leskes <b.leskes@gmail.com>
Date: Tue Jul 21 15:54:36 2015 +0200
Remove left over no commit from TransportReplicationAction
It asks to double check thread pool rejection. I did and don't see problems with it.
commit e5724931bbc1603e37faa977af4235507f4811f5
Author: Alexander Reelsen <alexander@reelsen.net>
Date: Tue Jul 21 15:31:57 2015 +0200
CliTool: Various PluginManager fixes
The new plugin manager parser was not called correctly in the scripts.
In addition the plugin manager now creates a plugins/ directory in case
it does not exist.
Also the integration tests called the plugin manager in the deprecated way.
commit 7a815a370f83ff12ffb12717ac2fe62571311279
Author: Alexander Reelsen <alexander@reelsen.net>
Date: Tue Jul 21 13:54:18 2015 +0200
CLITool: Port PluginManager to use CLITool
In order to unify the handling and reuse the CLITool infrastructure
the plugin manager should make use of this as well.
This obsolets the -i and --install options but requires the user
to use `install` as the first argument of the CLI.
This is basically just a port of the existing functionality, which
is also the reason why this is not a refactoring of the plugin manager,
which will come in a separate commit.
commit 7f171eba7b71ac5682a355684b6da703ffbfccc7
Author: Martijn van Groningen <martijn.v.groningen@gmail.com>
Date: Tue Jul 21 10:44:21 2015 +0200
Remove custom execute local logic in TransportSingleShardAction and TransportInstanceSingleOperationAction and rely on transport service to execute locally. (forking thread etc.)
Change TransportInstanceSingleOperationAction to have shardActionHandler to, so we can execute locally without endless spinning.
commit 0f38e3eca6b570f74b552e70b4673f47934442e1
Author: Ryan Ernst <ryan@iernst.net>
Date: Tue Jul 21 17:36:12 2015 -0700
More readMetadata tests and pickiness
commit 880b47281bd69bd37807e8252934321b089c9f8e
Author: Ryan Ernst <ryan@iernst.net>
Date: Tue Jul 21 14:42:09 2015 -0700
Started unit tests for plugin service
commit cd7c8ddd7b8c4f3457824b493bffb19c156c7899
Author: Robert Muir <rmuir@apache.org>
Date: Tue Jul 21 07:21:07 2015 -0400
fix tests
commit 673454f0b14f072f66ed70e32110fae4f7aad642
Author: Robert Muir <rmuir@apache.org>
Date: Tue Jul 21 06:58:25 2015 -0400
refactor pluginservice
When performing an operation on a primary, the state is captured and the
operation is performed on the primary shard. The original request is
then modified to increment the version of the operation as preparation
for it to be sent to the replicas.
If the request first fails on the primary during the translog sync
(because the Engine is already closed due to shadow primaries closing
the engine on relocation), then the operation is retried on the new primary
after being modified for the replica shards. It will then fail due to the
version being incorrect (the document does not yet exist but the request
expects a version of "1").
Order of operations:
- Request is executed against primary
- Request is modified (version incremented) so it can be sent to replicas
- Engine's translog is fsync'd if necessary (failing, and throwing an exception)
- Modified request is retried against new primary
This change ignores the exception where the engine is already closed
when syncing the translog (similar to how we ignore exceptions when
refreshing the shard if the ?refresh=true flag is used).
The help files are using a unix based file separator, where as
the test relies on the help being based on the file system separator.
This commit fixes the test to remove all `\r` characters before
comparing strings.
The test has also been moved into its own CliToolTestCase, as it does
not need to be an integration test.
There is no need to maintain additional state as to if a primary was allocated post api creation on the index routing table, we hold all this information already in the UnassignedInfo class.
closes#12374
The release and smoke test python scripts used to install
plugins in the old fashion.
Also the BATS testing suite installed/removed plugins in that
way. Here the marvel tests have been removed, as marvel currently
does not work with the master branch.
In addition documentation has been updated as well, where it was
still missing.
This dependency was used in order for mapping updates that change the fielddata
format to take effect immediately. And the way it worked was by clearing the
cache of fielddata instances that were already loaded. However, we do not need
to cache the already loaded (logical) fielddata instances, they are cheap to
regenerate. Note that the fielddata _caches_ are still kept around so that we
don't keep on rebuilding costly (physical) fielddata values.
The new plugin manager parser was not called correctly in the scripts.
In addition the plugin manager now creates a plugins/ directory in case
it does not exist.
Also the integration tests called the plugin manager in the deprecated way.
In order to unify the handling and reuse the CLITool infrastructure
the plugin manager should make use of this as well.
This obsolets the -i and --install options but requires the user
to use `install` as the first argument of the CLI.
This is basically just a port of the existing functionality, which
is also the reason why this is not a refactoring of the plugin manager,
which will come in a separate commit.
The `_index` field is now a completely virtual field thanks
to #12027. It is no longer necessary to index the actual value
of the index name.
closes#12329
The index name was passed along through many levels of mapping parsing,
just so that it could be used for _index. However, the index name
is really metadata that should exist alongside things like type and
id in SourceToParse.
This change moves index name to SourceToParse, and eliminates it from the
DocumentMapperParser.
Today we grant read+write+delete access to any files underneath the home.
But we have to remove this, if we want to have improved security of files
underneath elasticsearch.
Fold ignored unassigned to a UnassignedShards and have simpler handling of them. Also remove the trapy way of adding an ignored unassigned shards today directly to the list, and have dedicated methods for it.
This change also removes the useless moving of unassigned shards to the end, since anyhow we first, sort those unassigned shards, and second, we now have persistent "store exceptions" that should not cause "dead letter" shard allocation.
Break it into more manageable code by separating allocation primaries and allocating replicas. Start adding basic unit tests for primary shard allocator.
Our thread pools have support for timeout on a task. To support this, a special background task is schedule to run at timeout. That background task fires and check if the main task is still in the executor queue and then cancels it if needed. Currently we schedule this background task before adding the main task to the queue. If the timeout is very small (in tests we often use numbers like 2 ms) the background task can fire before the main one is added to the queue causing the timeout to be missed.
See http://build-us-00.elastic.co/job/es_g1gc_master_metal/11780/testReport/junit/org.elasticsearch.cluster/ClusterServiceTests/testTimeoutUpdateTask/Closes#12319
On top of that:
1) A relocation target shards' allocation id is changed to include the allocation id of the source shard under relocatingId (similar to shard routing semantics)
2) The logic around state change for finalize shard relocation is simplified - one simple start the target shard (we previously had unused logic around relocating state)
Closes#12299
While the GeoJSON spec does say a polygon is represented as an array of LinearRings (where a LinearRing is defined as a 'closed' array of points), the coerce parameter provides users with flexibility to have ES automatically close polygons. This addresses situations like those integrated with twitter (where GeoJSON polygons are not closed) such that our users do not have to write extra code to close the polygon. This code change adds the optional coerce parameter to the GeoShapeFieldMapper.
closes#11131
Currently this target is "yet another way" to run elasticsearch,
which we can't maintain. It also has the problem that it doesnt
ensure its running on the latest source code, doesn't configure
any scratch space properly, won't work with securitymanager, list
goes on.
Even if we made it work, it would break every day, since its untested.
Instead, `mvn package -Drun -DskipTests` will run packaging, and then
startup bin/elasticsearch (like integration tests, but in foreground).
It also enables debugger socket on port 8000, for people that like
IDE debuggers and not system.out.println.
Its a little slower to get started because of all the shading/RPM/DEB
building going on in `package` but that is just what it is right now
until that stuff is moved out.
failsafe uses surefire, which sucks. It also mean integ tests act alien right now.
I would rather have the consistency, e.g. things formatted the same way, running integ tests under security manager, etc.
Store information reports on which nodes shard copies exist, the shard
copy version, indicating how recent they are, and any exceptions
encountered while opening the shard index or from earlier engine failure.
closes#10952
When adding a script to the Groovy classloader, the script name is used
as the class identifier in the classloader. This means that in order not
to break JVM Classloader convention, that script must always be
available by that name. As a result, modifying a script with the same
content over and over causes it to be loaded with a different name (due
to the incrementing integer).
This is particularly bad when something like chef or puppet replaces the
on-disk script file with the same content over and over every time a
machine is converged.
This change makes the script name the SHA1 hash of the script itself,
meaning that replacing a script with the same text will use the same
script name.
Resolves#12212
Just like specifying `?preference=_primary`, this adds the ability to
specify `?preference=_replica` or `?preference=_replica_first` on
requests that support it.
Resolves#12222
This action is a liveness test added in #8763 . It should be excluded, just like the fault detection logic or things become overly chatty.
Closes#12291
Most of our tests call assertSearchResponse which checks whether all shards
were successful. However this is usually not necessary given that all shards
that received documents should be available given the way our indexing works.
The only shards that might not be available are those that did not index
any documents. So removing this assertion would allow us to remove most
ensureGreen/ensureYellow calls while still being able to assert on the content
of the search response since all data have been taken into account.
For now I only removed the ensureYellow/Green calls from SearchQueryTests in
order to not de-stabilize the build, but eventually we should remove most of
them.
Today, the unicast zen test configuration will try to find a open port starting at the internal test cluster's
base port and continuing for 1000 ports. The internal test cluster class assigns a port range of 100 ports
to each JVM. This means that the unicast zen test configuration will try ports in the range for another JVM
and can lead to port conflicts. This change uses the same value for both so that the unicast configuration
does not go into another JVM's port range.
Add a unique allocation id for a shard, helping to uniquely identify a specific allocation taking place to a node.
A special case is relocation, where a transient relocationId is kept around to make sure the target initializing shard (when using RoutingNodes) is using it for its id, and when relocation is done, the transient relocationId becomes the actual id of it.
closes#12242
If an index name is reused but a leftover shard still exists on any node
we fail repeatedly to allocate the shard since we now check the index UUID
before reusing data. This commit allows to recover even if there is such a
leftover shard by deleting the leftover shard.
Closes#10677
During master election each node pings in order to discover other nodes and validate the liveness of existing nodes. Based on this information the node either discovers an existing master or, if enough nodes are found (based on `discovery.zen.minimum_master_nodes>>) a new master will be elected.
Currently, the node that is elected as master will currently update it the cluster state to indicate the result of the election. Other nodes will submit a join request to the newly elected master node. Instead of immediately processing the election result, the elected master
node should wait for the incoming joins from other nodes, thus validating the elections result is properly applied. As soon as enough nodes have sent their joins request (based on the `minimum_master_nodes` settings) the cluster state is modified.
Note that if `minimum_master_nodes` is not set, this change has no effect.
Closes#12161
Require urls for URL repository to be listed in repositories.url.allowed_urls setting. This change ensures that only authorized URLs can be accessed by elasticsearch
The previous strategy (target/xxx + .m2/repository) is obviously broken for multi-module
builds.
Includes hack for crazy jython, which "finds its own jar" then looks for a Lib/ beside it
Lucene deprecated this in 4.0 and we only try best effort to support it.
Folks should only use edit distance rather than some length based
similarity. Yet the formular is simple enough such that users can
still do it in the client if they really need to.
Closes#10638
The old script syntax has been removed from the Java API but the metrics aggregations were missed. This change removes the old script API from the ValuesSourceMetricsAggregationBuilder and removes the relevant test methods for the metrics aggregations.
Today we have a intermediate hierarchy for shard and index exceptions
which makes it hard to introduce generic exceptions like ResourceNotFoundException
intoduced in this commit. This commit breaks up the hierarchy by adding index and shard
as a special internal header that gets rendered for every exception that fills that header.
This commit removes dedicated exceptions like `IndexMissingException` or
`IndexShardMissingException` in favour of `ResourceNotFoundException`
Change the default delayed allocation timeout from 0 (no delayed allocation) to 1m. The value came from a test of having a node with 50 shards being indexed into (so beefy translog requiring flush on shutdown), then shutting it down and starting it back up and waiting for it to join the cluster. This took, on a slow machine, about 30s.
The value is conservatively low and does not try to address a virtual machine / OS restart for now, in order to not have the affect of node going away and users being concerned that shards are not being allocated to the rest of the cluster as a result of that. The setting can always be changed in order to increase the delayed allocation if needed.
closes#12166
The MetaData.clusterUUID is guaranteed to be unique across clusters and is handy (which may or may not have the same human readable cluster name).
Closes#11832
As explained in #11831, we currently have uuid fields on the cluster state, meta data and index metadata. The latter two are persistent across changes are being effectively used as a persistent uuid for this cluster and a persistent uuid for an index. The first (ClusterState.uuid) is ephemeral and changes with every change to the cluster state. This is confusing,
We settled on having the following, new names:
-> ClusterState.uuid -> stateUUID (transient)
-> MetaData.uuid -> clusterUUID (persistent)
-> IndexMetaData.uuid -> indexUUID (persistent).
Closes#11914Closes#11831
Today we throw ElasticsearchException if we can't lock the index. This can cause
problems since some places where we have logic to deal with IOException on shard
deletion won't schedule a retry if we can't lock the index dir for removal. This
is the case on shadow replicas for instance if a shared FS is used. The result
of this is that the delete of an index is never acked.
A change in #12116 introduces closing / cleaning of search ctx even if
the index service was closed due to a relocation of it's last shard. This
is not desired since in that case it's fine to serve the pending requests from
the relocated shard. This commit adds an extra check to ensure that the index is
either removed (delete) or closed via API.
Igor Motov
Ryan Ernst
Nik Everett
Christoph Büscher
Simon Willnauer
Alexander Reelsen
keety
Clinton Gormley
Robert Muir
Jason Tedor
javanna
Adrien Grand
Robert Muir
Lee Hinman
oyiadom
Tanguy Leroux
Britta Weber
Colin Goodheart-Smithe
Martijn van Groningen
szroland
Dave Parfitt
Dave Parfitt
Boaz Leskes
Ruflin
Shay Banon
xuzha
Alex Ksikes
Jay Modi
jaymode
khiraiwa
Areek Zillur
Zachary Tong
Alexander Clare
Nicholas Knize
Michael McCandless
Simon Willnauer
uboness