Before publishing a cluster state the master connects to the nodes that are added in the cluster state. When publishing fails, however, it does not disconnect from these nodes, leaving NodeConnectionsService out of sync with the currently applied cluster state.
The `_cat/nodes` API might not be available in all clusters for instance
if they have authorization enabled. This change falls back to the previously
used method of using the '/' endpoint to fetch the nodes version, this is best
effort and will emit a warning.
Lucene 6.3 is expected to be released in the next weeks so it'd be good to give
it some integration testing. I had to upgrade randomized-testing too so that
both Lucene and Elasticsearch are on the same version.
Today we only use a single node to send requests to when we run REST tests.
In some cases we have more than one node (ie. in the BWC case) where we should
send requests to all nodes in a round-robin fashion. This change passes all
available node endpoints to the rest test.
Additionally, this change adds the setting of `discovery.zen.minimum_master_nodes`
to the cluster formation forcing the nodes to wait for all other nodes until the cluster
is formed. This allows for a more realistic master election and allows all master eligable
nodes to become master while before always the first node in the cluster became the master.
This also adds logging to each test run to log the master nodes version and the minimum node
version in the cluster to help debugging BWC test failures.
This fixes our cluster formation task to run REST tests against a mixed version cluster.
Yet, due to some limitations in our test framework `indices.rollover` tests are currently
disabled for the BWC case since they select the current master as the merge node which
happens to be a BWC node and we can't relocate all shards to it since the primaries are on
a higher version node. This will be fixed in a followup.
Closes#21142
Note: This has been cherry-picked from 5.0 and fixes several rest tests
as well as a BWC break in `OsStats.java`
The network disruption type "network delay" continues delaying existing requests even after the disruption has been cleared. This commit ensures that the requests get to execute right after the delay rule is cleared.
This commit fixes responses to HEAD requests so that the value of the
Content-Length is correct per the HTTP spec. Namely, the value of this
header should be equal to the Content-Length if the request were not a
HEAD request.
This commit also fixes a memory leak on HEAD requests to the main action
that arose from the bytes on a builder not being released due to them
being dropped on the floor to ensure that the response to the main
action did not have a body.
Relates #21123
Currently test that check that equals() and hashCode() are working as expected
for classes implementing them are quiet similar. This change moves common
assertions in this method to a common utility class. In addition, another common
utility function in most of these test classes that creates copies of input
object by running them through a StreamOutput and reading them back in, is moved
to ESTestCase so it can be shared across all these classes.
Closes#20629
Today the request interceptor can't support async calls since the response
of the async call would execute on a different thread ie. a client or listener
thread. This means in-turn that the intercepted handler is not executed with the
thread it was supposed to run and therefor can, if it's executing blocking
operations, potentially deadlock an entire server.
* Move all zen discovery classes into o.e.discovery.zen
This collapses sub packages of zen into zen. These all had just a couple
classes each, and there is really no reason to have the subpackages.
* fix checkstyle
`LocalDiscovery` is a discovery implementation that uses static in memory maps to keep track of current live nodes. This is used extensively in our tests in order to speed up cluster formation (i.e., shortcut the 3 second ping period used by `ZenDiscovery` by default). This is sad as that mean that most of the test run using a different discovery semantics than what is used in production. Instead of replacing the entire discovery logic, we can use a similar approach to only shortcut the pinging components.
This commit changes the current REST API parser to make it fail and throw an exception when a REST specification file contains a duplicated parameters, or path, or method, or path part.
This is important to allow any test to use RandomQueryBuilder#createQuery()
since some of the query builders that are used in this test test the length
of the types array and otherwise will thow NPE if the test is not a subclass
of AbstractQueryTestCase.
The cache relies on the equals() method so we just need to make sure script
queries can never be equals, even to themselves in the case that a weight
is used to produce a Scorer on the same segment multiple times.
Closes#20763
Today we throw an assertion error if we release an AbstractArray more than once.
Yet, it's recommended to implement close methods such that they can be invoked
more than once. Guaranteed single release calls are hard to implement and some
situations might not be tested causing for instance `CircuitBreaker` to operate on
corrupted memory stats.
UpdateHelper, MetaDataIndexUpgradeService, and some recovery
stuff.
Move ClusterSettings to nullable ctor parameter of TransportService
so it isn't forgotten.
This change proposes the removal of all non-tcp transport implementations. The
mock transport can be used by default to run tests instead of local transport that has
roughly the same performance compared to TCP or at least not noticeably slower.
This is a master only change, deprecation notice in 5.x will be committed as a
separate change.
Today SearchContext expose the current context as a thread local which makes any kind of sane interface design very very hard. This PR removes the thread local entirely and instead passes the relevant context anywhere needed. This simplifies state management dramatically and will allow for a much leaner SearchContext interface down the road.
LongGCDisruption suspends and resumes node threads but respects several
`unsafe` class name patterns where it's unsafe to suspend. For instance
log4j uses a global lock so we can't suspend a thread that is currently
calling into log4j. The same is true for the security manager, it's similar
to log4j a shared resource between the test and the node that is _suspended_.
This change adds `java.lang.SecrityManager` to the unsafe patterns.
This prevents test framework deadlocking if a nodes thread is supended
while it's calling into the security manager that uses synchronized maps etc.
today it's not possible to use date-math efficiently with the `_rollover`
API. This change adds support for date-math in the target index as well as
support for preserving the math logic when an existing index that was created with
a date math expression all subsequent indices are created with the same expression.
The logging listener tests started failing after
953a8a959b when the tests are run with
tests.es.logger.level set to any level other than debug. This is because
these tests were based around the assumption that the default logging
level was info, which was the case before that commit fixed setting the
default logging level via that system property. This commit fixes these
failing tests by adjusting this assumption to account for the fact that
the default logging level could be different.
Pipe in the `tests.es.logger.level` system property to the log4j config file used in tests. We still default to info. Also adapts the logger name to use the first letter of packages.
Many of our unit tests instantiate an `AllocationService`, which requires having a `GatewayAllocator`. Today almost all of our test use a class called `NoopGatewayAllocator` which does nothing, effectively leaving all shard assignments to the balanced allocator. This is sad as it means we test a system that behaves differently than our production logic in very basic things. For example, a started primary that is lost will be assigned to a node that didn't use to have it.
This PR removes `NoopGatewayAllocator` in favor of a new `TestGatewayAllocator` that inherits the standard `GatewayAllocator` and overrides shard information fetching to return information based on historical assignments the allocator has done. The only exception is `BalanceConfigurationTests` which does test only the balancer and I opted to not have it work around the `GatewayAllocator` being in it's way.
Changes the API of GatewayAllocator#applyStartedShards and
GatewayAllocator#applyFailedShards to take both a RoutingAllocation
and a list of shards to apply. This allows better mock allocators
to be created as being done in #20637.
Closes#20642
Removes the FailedRerouteAllocation class and StartedRerouteAllocation
class, as they were just wrappers for RerouteAllocation that stored
started and failed shards, but these started and failed shards can
be passed in directly to the methods that needed them, removing the
need for this wrapper class and extra level of indirection.
Closes#20626
Today we hold on to all possible tokenizers, tokenfilters etc. when we create
an index service on a node. This was mainly done to allow the `_analyze` API to
directly access all these primitive. We fixed this in #19827 and can now get rid of
the AnalysisService entirely and replace it with a simple map like class. This
ensures we don't create a gazillion long living objects that are entirely useless since
they are never used in most of the indices. Also those objects might consume a considerable
amount of memory since they might load stopwords or synonyms etc.
Closes#19828
With the switch to Log4j 2 throughout our code base, the logger usage checker was temporarily disabled. This commit
adapts the checks to work with Log4j 2 and re-enables the Gradle checks.
Closes#20243
This commit changes the default behavior of `_flush` to block if other flushes are ongoing.
This also removes the use of `FlushNotAllowedException` and instead simply return immediately
by skipping the flush. Users should be aware if they set this option that the flush might or might
not flush everything to disk ie. no transactional behavior of some sort.
Closes#20569
This commit removes `ByteSizeValue`'s methods that are duplicated (ex: `mbFrac()` and `getMbFrac()`) in order to only keep the `getN` form.
It also renames `mb()` -> `getMb()`, `kb()` -> `getKB()` in order to be more coherent with the `ByteSizeUnit` method names.
This PR introduces backward compatibility index tests to test the rolling upgrade process amongst Elasticsearch instances within the same major version. The test executes in three phases. In the first phase, we form a cluster of 2 ES instances on an old version. In the second phase, we keep one of the nodes from the old cluster, kill the other node, but preserve its data directory and start an instance of the current version of ES using the same data directory as the killed instance. In the third phase, we kill the other old version ES instance from the first phase and launch a new instance, using the same data directory as the killed instance. Therefore, during phase 3, we have fully migrated and have all current versions of ES running. In each phase, we run REST tests that index documents and search them, ensuring at each stage that the documents from the previous phase are still there.
Note that because we haven't released a GA yet of 5.0, the tests currently don't start an old version cluster in the first phase. Once GA is released, this will be changed to make the backward compatibility version 5.0, while the current version in the cluster will be 5.x.
This change removes all guice interaction from Transport, HttpServerTransport,
HttpServer and TransportService. All these classes as well as their subclasses
or extended version configured via plugins are now created by using plain old
bloody java constructors. YAY!
Currently all the reroute-like methods of `AllocationService` return a result object of type `RoutingAllocation.Result`. The result object contains the new `RoutingTable` and `MetaData` plus an indication whether those were changed. The caller is then responsible of updating a cluster state with these. These means that things can easily go wrong and one can take one of these but not the other causing inconsistencies. We already have a utility method on the `ClusterState` builder that does but no one forces you to do so. Also 99% of the callers do the same thing: i.e., check if the result was changed and if so update the very same cluster state that was passed to `AllocationService`. This PR folds this pattern into `AllocationService` and changes almost all it's methods to return a new cluster state (potentially the original one). This saves some 500 lines of code.
The one exception here is the reroute API which executes allocation commands and potentially returns an explanation as well (next to the routing table and metadata). That API now returns a `CommandsResult` object which encapsulate a cluster state and the explanation.
TransportService is such a central part of the core server, replacing
it's implementation is risky and can cause serious issues. This change removes the ability to
plug in TransportService but allows registering a TransportInterceptor that enables
plugins to intercept requests on both the sender and the receiver ends. This is a commonly used
and overwritten functionality but encapsulates the custom code in a contained manner.
During a networking partition, cluster states updates (like mapping changes or shard assignments)
are committed if a majority of the masters node received the update correctly. This means that the current master has access to enough nodes in the cluster to continue to operate correctly. When the network partition heals, the isolated nodes catch up with the current state and get the changes they couldn't receive before. However, if a second partition happens while the cluster
is still recovering from the previous one *and* the old master is put in the minority side, it may be that a new master is elected which did not yet catch up. If that happens, cluster state updates can be lost.
This commit fixed 95% of this rare problem by adding the current cluster state version to `PingResponse` and use them when deciding which master to join (and thus casting the node's vote).
Note: this doesn't fully mitigate the problem as a cluster state update which is issued concurrently with a network partition can be lost if the partition prevents the commit message (part of the two phased commit of cluster state updates) from reaching any single node in the majority side *and* the partition does allow for the master to acknowledge the change. We are working on a more comprehensive fix but that requires considerate work and is targeted at 6.0.
The only repository we can be sure is safe to clean is `fs` so we clean
any snapshots in those repositories after each test. Other repositories
like url and azure tend to throw exceptions rather than let us fetch
their contents during the REST test. So we clean what we can....
Closes#18159
Yannick Welsch
Simon Willnauer
Adrien Grand
Simon Willnauer
Jason Tedor
Igor Motov
Christoph Büscher
Ryan Ernst
javanna
Boaz Leskes
Tanguy Leroux
Nik Everett
Colin Goodheart-Smithe
Ali Beyad