We support rolling upgrades from Zen1 by keeping the master as a Zen1 node
until there are no more Zen1 nodes in the cluster, using the following
principles:
- Zen1 nodes will never vote for Zen2 nodes
- Zen2 nodes will, while not bootstrapped, vote for Zen1 nodes
- Zen2 nodes that were previously part of a mixed cluster will automatically
(and unsafely) bootstrap themselves when the last Zen1 node leaves.
This commit makes FormatDateTimeFormatter and DateFormatter apis close
to each other, so that the former can be removed in favor of the latter.
This PR does not change the uses of FormatDateTimeFormatter yet, so that
that future change can be purely mechanical.
This is related to #35975. It implements a basic restore functionality
for the CcrRepository. When the restore process is kicked off, it
configures the new index as expected for a follower index. This means
that the index has a different uuid, the version is not incremented, and
the Ccr metadata is installed.
When the restore shard method is called, an empty shard is initialized.
This commit removes the parseDefaulting method from DateFormatter,
bringing it more inline with the joda equivalent
FormatDateTimeFormatter. This method was only needed for the java
time implementation of DateMathParser. Instead, a DateFormatter now
returns an implementation of DateMathParser for the given format,
allowing the java time implementation to construct the appropriate date
math parser internally.
* Add warning if cluster fails to form fast enough
Today if a leader is not discovered or elected then nodes are essentially
silent at INFO and above, and log copiously at DEBUG and below. A short delay
when electing a leader is not unusual, for instance if other nodes have not yet
started, but a persistent failure to elect a leader is a problem worthy of log
messages in the default configuration.
With this change, while there is no leader each node outputs a WARN-level log
message every 10 seconds (by default) indicating as such, describing the
current discovery state and the current quorum(s).
* Add note about whether the discovered nodes form a quorum or not
* Introduce separate ClusterFormationFailureHelper
... and back out the unnecessary changes elsewhere
* It can be volatile
In #34474, we added a new assertion to ensure that the
LocalCheckpointTracker is always consistent with Lucene index. However,
we reset LocalCheckpoinTracker in testDedupByPrimaryTerm cause this
assertion to be violated.
This commit removes resetCheckpoint from LocalCheckpointTracker and
rewrites testDedupByPrimaryTerm without resetting the local checkpoint.
Relates #34474
Add support for inlined user dictionary in Nori
This change adds a new option called `user_dictionary_rules` to the
Nori a tokenizer`. It can be used to set additional tokenization rules
to the Korean tokenizer directly in the settings (instead of using a file).
Closes#35842
In real deployments it is important that clusters are properly configured to
avoid accidentally forming multiple independent clusters at cluster
bootstrapping time. However we also expect to be able to unpack Elasticsearch
and start up one or more nodes without any up-front configuration, and have
them do their best to find each other and form a cluster after a few seconds.
This change adds a delayed automatic bootstrapping process to nodes that start
up with no relevant settings set to support the desired out-of-the-box
experience without compromising safety in properly-configured deployments.
The conversion of timezones in JodaCompatibleZonedDateTime from joda to
java time requires the use of the DateUtils class to cater for corner
cases.
Closes#36306
Includes:
LUCENE-8594: DV update are broken for updates on new field
LUCENE-8590: Optimize DocValues update datastructures
LUCENE-8593: Specialize single value numeric DV updates
Relates #36286
The test testLookupSeqNoByIdInLucene fails because it expects if any
change should be visible after a flush. However, that flush might be
ignored if the waitIfOngoing parameter is false (the default value), and
there is an ongoing flush triggered after merge is running.
Closes#35823
This commit moves back to use explicit dependsOn for test tasks on
check. Not all tasks extending RandomizedTestingTask should be run by
check directly.
* Add deprecation warnings to `Rest*TermVectorsAction`, plus tests in `Rest*TermVectorsActionTests`.
* Deprecate relevant methods on the Java HLRC requests/ responses.
* Update documentation (for both the REST API and Java HLRC).
* For each REST yml test, create one version without types, and another legacy version that retains types (called *_with_types.yml).
* Merging the zen2 branch has made it impossible to reproduce the test failure here, likely as a result of the state now being written atomically
* Closes#36189
Some IDEs (specifically Eclipse 4.8.0) needs some explicit casting for type
inference. This was added in a recent change but we should also have comments so
we remember why this needs to be there (at least for now).
Currently, we only log that WriteStateException has occurred, in
GatewayMetaState.
This PR goal is to re-throw WriteStateException to upper layers.
If dirty flag is set to false, we wrap WriteStateException in
UncheckedIOException, we prefer unchecked exception not to add
explicit throws in the multitude of layers.
If dirty flag is set to true - the world is broken. And we need to
halt the JVM. Instead of explicit halting in GatewayMetaState, we
prefer to throw IOError, which will be subsequently handled by
ElasticsearchUncaughtExceptionHandler and JVM will be halted.
This PR also adds tests for WriteStateException.
The Eclipse IDE java compiler seems to need some special hints about what types
some functions used in the tests return. Correcting this for some test that were
newly merged to master.
This change removes the custom serialization of the total hits
and reuses the shard's serialization of Lucene#read/writeTopDocs
in the client code. This also removes the incorrect assertion that
trips randomly in bwc tests.
Closes#36284
Currently we use Math.random() in a few places in the tests which makes these
tests not reproducable with the random seed mechanism that comes with
ESTestCase. The change removes those instances.
This commit hides ClusterStates that have a STATE_NOT_RECOVERED_BLOCK from
ClusterStateAppliers. This is needed, because some appliers, such as IngestService, rely on
the fact, that cluster states with STATE_NOT_RECOVERED_BLOCK won't contain anything useful.
Once the state is recovered it's fully available for the appliers. This commit also switches many of
the remaining tests that require state persistence/recovery from Zen1 to Zen2.
This commit changes the format of the `hits.total` in the search response to be an object with
a `value` and a `relation`. The `value` indicates the number of hits that match the query and the
`relation` indicates whether the number is accurate (in which case the relation is equals to `eq`)
or a lower bound of the total (in which case it is equals to `gte`).
This change also adds a parameter called `rest_total_hits_as_int` that can be used in the
search APIs to opt out from this change (retrieve the total hits as a number in the rest response).
Note that currently all search responses are accurate (`track_total_hits: true`) or they don't contain
`hits.total` (`track_total_hits: true`). We'll add a way to get a lower bound of the total hits in a
follow up (to allow numbers to be passed to `track_total_hits`).
Relates #33028
Closes#19528
* Port a test Colin wrote for the TDigest library to validate TDigests storing over 2*MAXINT values. This appears to have been fixed in version 3.2 of TDigest, which Elasticsearch has been using for some time, so no changes were necessary to resolve this issue.
The shard deletion logic (triggered by IndicesStore), which also leads to index metadata deletion on
non-master-eligible data nodes, currently races against the new cluster state persistence logic
triggered by accepting cluster states. One thread is writing the index metadata while another one is
deleting the index metadata, leading to exceptions and assertions tripping (see below). The solution
proposed by this PR is to move the cluster state persistence of non-master-eligible nodes back to
the cluster applier service, just as it used to be for Zen1. This ensures that the index metadata
deletion logic, which is triggered by the shard deletion logic, runs on the same thread on which we
persist the cluster state.
Closes#35435
- make it easier to add additional testing tasks with the proper configuration and add some where they were missing.
- mute or fix failing tests
- add a check as part of testing conventions to find classes not included in any testing task.
This commit replaces usages of Streamable with Writeable for the
BaseTasksResponse / TransportTasksAction classes and subclasses of
these classes.
Note that where possible response fields were made final.
Relates to #34389
This commit adds an empty CcrRepository snapshot/restore repository.
When a new cluster is registered in the remote cluster settings, a new
CcrRepository is registered for that cluster.
This is implemented using a new concept of "internal repositories".
RepositoryPlugin now allows implementations to return factories for
"internal repositories". The "internal repositories" are different from
normal repositories in that they cannot be registered through the
external repository api. Additionally, "internal repositories" are local
to a node and are not stored in the cluster state.
The repository will be unregistered if the remote cluster is removed.
* Move `createRepository` call out of cluster state tasks
* Now only `RepositoriesService#applyClusterState` manipulates `this.repositories`
* Closes#9488
This commit makes `document`, `update`, `explain`, `termvectors` and `mapping`
typeless APIs work on indices that have a type whose name is not `_doc`.
Unfortunately, this needs to be a bit of a hack since I didn't want calls with
random type names to see documents with the type name that the user had chosen
upon type creation.
The `explain` and `termvectors` do not support being called without a type for
now so the test is just using `_doc` as a type for now, we will need to fix
tests later but this shouldn't require further changes server-side since passing
`_doc` as a type name is what typeless APIs do internally anyway.
Relates #35190
It is important that all shards of a given index have the same
`indexCreatedVersionMajor` to Lucene, or eg. merging those shards is going to
be considered illegal. At the moment, we use the latest Lucene version when
creating a shard, which could cause shards to have different created versions
eg. in case of forced allocation. This commit makes sure to reuse the
appropriate Lucene version in order to avoid such issues.
Closes#33826
Today we configure the soft-deletes field iff soft-deletes enabled.
Although this choice was correct, it prevents an engine with
soft-deletes disabled from opening a Lucene index with soft-deletes.
Moreover, this change should not have any side-effect if a Lucene index
does not have any soft-deletes.
Relates #36141
This commit implements proper metadata recovery for Zen2.
GatewayService is responsible for the recovery. In Zen1 GatewayService
creates an instance of Gateway, that is used to reach out to other cluster
nodes, get their state and calculate the most up-to-date state based on
versions. After that Gateway performs upgrade and archival of
ClusterSettings and closes bad indices. Then recovered state is passed to GatewayService.GatewayRecoveryListener that mixes up current state
and restored state, removes state not recovered block, creates the
routing table and performs re-routing.
In Zen2 we should perform this kind of logic on cluster startup, except
mixing state (because there is nothing to mix) and opening routing table.
This commit refactors out all `ClusterUpdate` functions in a separate class
`ClusterStateUpdaters`, which is used by `Gateway` and `GatewayService`
in case of Zen1, and by `GatewayMetaState` and `GatewayService` in case of
Zen2.
This commit also switches all integration tests that are already using Zen2 from
InMemoryPersistedState to GatewayMetaState.
This commit changes how an operation which requires all index shard
operations permits is executed when a primary term update is required:
the operation and the update are combined so that the operation is
executed after the primary term update under the same blocking
operation.
Closes#35850
Co-authored-by: Yannick Welsch <yannick@welsch.lu>
This test suite can stop all the shared master-eligible nodes, which breaks the
cluster since any non-shared master-eligible nodes are stopped first in the
reset process between tests.
Since this test suite can leave the cluster in this somewhat broken state, it
seems best that it uses a new cluster for each test.
This change adds a soft limit to open scroll contexts that can be controlled with the dynamic cluster setting `search.max_open_scroll_context` (defaults to 500).
Today if a node `A` sends a peers request to another node `B` then `B` will
react by sending a peers request back to `A`. However if `A` is not
master-eligible then this reaction is pointless and fails with an exception
saying `non-master-eligible node found`, adding noise to the logs. This change
suppresses this response to non-master-eligible nodes.
Given that we check the max buckets limit on each shard when collecting the buckets, and that non final reduction cannot add buckets (see #35921), there is no point in counting and checking the number of buckets as part of non final reduction phases.
Such check is still needed though in the final reduction phases to make sure that the number of returned buckets is not above the allowed threshold.
Relates somehow to #32125 as we will make use of non final reduction phases in CCS alternate execution mode and that increases the chance that this check trips for nothing when reducing aggs in each remote cluster.
When building a query Lucene distinguishes two cases, queries that require to produce a score and queries that only need to match. We cloned this mechanism in the QueryBuilders in order to be able to produce different queries based on whether they need to produce a score or not. However the only case in es that require this distinction is the BoolQueryBuilder that sets a different minimum_should_match when a `bool` query is built in a filter context..
This behavior doesn't seem right because it makes the matching of `should` clauses different when the score is not required.
Closes#35293
* Moved method `canOpenIndex` is only used in tests -> moved to test CP
* Simplify `org.elasticsearch.index.store.Store#renameTempFilesSafe`
* Delete some dead methods
* Replace Streamable w/ Writeable in BaseTasksRequest and subclasses
This commit replaces usages of Streamable with Writeable for the
BaseTasksRequest / TransportTasksAction classes and subclasses of
these classes.
Relates to #34389
In #36033 we removed a catch block because we thought we were preventing
exceptions by avoiding concurrent elections, missing the obvious fact that some
joins are supposed to be failing.
As a quick fix the catch was reinstated in 3a5dab6d8e
but this change adds finesse by only catching exceptions from the joins that we
expect to fail. It also inlines an always-false parameter to `initialState()`.
Adds about a minute worth of backoffs and retries to saving task
results so it is *much* more likely that a busy cluster won't lose task
results. This isn't an ideal solution to losing task results, but it is
an incremental improvement. If all of the retries fail when still log
the task result, but that is far from ideal.
Closes#33764
Empty buckets don't need to be added when performing an incremental reduction step, they can be added later in the final reduction step. This will allow us to later remove the max buckets limit when performing non final reduction.
This is a follow-up to #35144. That commit made the underlying
connection opening process in TcpTransport asynchronous. However the
method still blocked on the process being complete before returning.
This commit moves the blocking to the ConnectionManager level. This is
another step towards the top-level TransportService api being async.
Prior to #35441 `ConnectionManager` had a `Lifecycle` object to support
the ping runnable. After that commit, the connection amanger only needs
the existing `AtomicBoolean` to indicate if it is running.
New that we test with min_doc_count set to 0 as well, we may end up generating a lot more buckets. This commit adjusts the min bound and max bound, as well as the offset for each randomly generated agg instance so that we don't end up hitting the 10.000 max buckets limit.
Relates to #36064
In this test we were randomizing different values but minDocCount was hardcoded to 1. It's important to test other values, especially `0` as it's the default. The test needed some adapting in the way buckets are randomly generated: all aggs need to share the same interval, minDocCount and emptyBucketInfo. Also assertions need to take into account that more (or less) buckets are expected depending on minDocCount.
CompositeBytesReference#slice has two bugs:
- One that makes it fail if the reference is empty and an empty slice is
created, this is #35950 and is fixed by special-casing empty-slices.
- One performance bug that makes it always create a composite slice when
creating a slice that ends on a boundary, this is fixed by computing `limit`
as the index of the sub reference that holds the last element rather than
the next element after the slice.
Closes#35950
org.elasticsearch.rest.RestController#hasContentType checks to see if the
RestHandler supports the `application/x-ndjson` Content-Type. DeprecationRestHandler
is a wrapper around the real RestHandler, and prior to this change
would always return `false` due to the interface's default supportsContentStream().
This prevents API's that use multi-line JSON from properly being deprecated
resulting in an HTTP 406 error.
This change ensures that the DeprecationRestHandler honors the
supportsContentStream() of the wrapped RestHandler.
Relates to #35958
The support for rest_total_hits_as_int has already been merged to 6x
in #35848 so this change adds this new option to master. The plan was
to add this new option as part of #35848 but we've decided to wait a few
days before merging this breaking change so this commit just handles
the new option as a noop exactly like 6x for now. This will allow
users to migrate to this parameter before #35848 is merged.
Relates #33028
This is related to #34405 and a follow-up to #34753. It makes a number
of changes to our current keepalive pings.
The ping interval configuration is moved to the ConnectionProfile.
The server channel now responds to pings. This makes the keepalive
pings bidirectional.
On the client-side, the pings can now be optimized away. What this
means is that if the channel has received a message or sent a message
since the last pinging round, the ping is not sent for this round.
The nested agg can defer the collection of children if it is nested
under another aggregation. In such case accessing the score in the children
aggregation throws an error because the scorer has already advanced to the next
parent. This change fixes this error by caching the score of the parent in the
nested aggregation. Children aggregations that work on nested documents will be
able to access the _score. Also note that the _score in this case is always the
parent's score, there is no way to retrieve the score of a nested docs in aggregations.
Closes#35985Closes#34555
* [Zen2] Implement Tombstone REST APIs
* Adds REST API for withdrawing votes and clearing vote withdrawls
* Tests added to Netty4 module since we need a real Network impl. for Http endpoints
Today the default for USE_ZEN2 is false and it is overridden in many places. By
defaulting it to true we can be sure that the only places in which Zen2 does
not work are those in which it is explicitly set to false.
Today we sometimes create a setup in which the node is a quorum on its own,
which allows it to win a pre-voting round and schedule an election essentially
at will, causing it to discard all the joins it just received and fail the
test. This change excludes this case, preventing stray elections from ruining
things.
Today any node can win an election. However, the whole point of
master-eligibility is that master-ineligible nodes should not be elected as the
leader; furthermore master-ineligible nodes do not have any outgoing STATE
channels so cannot publish cluster states, so their leadership is ineffective
and disruptive.
This change ensures that the elected leader is master-eligible by preventing
master-ineligible nodes from scheduling an election.
A number of tokenfilters can produce multiple tokens at the same position. This
is a problem when using token chains to parse synonym files, as the SynonymMap
requires that there are no stacked tokens in its input.
This commit ensures that when used to parse synonyms, these tokenfilters either produce
a single version of their input token, or that they throw an error when mappings are
generated. In indexes created in elasticsearch 6.x deprecation warnings are emitted in place
of the error.
* asciifolding and cjk_bigram produce only the folded or bigrammed token
* decompounders, synonyms and keyword_repeat are skipped
* n-grams, word-delimiter-filter, multiplexer, fingerprint and phonetic throw errors
Fixes#34298
The ActiveShardCount is used by cluster state observers to wait for a
given number of shards to be active before returning to the caller. The
current implementation does not work when an index is closed while an
observer is waiting on shards to be active. In this case, a NPE is thrown
and the observer is never notified that the shards won't become active.
This commit fixes the ActiveShardCount.enoughShardsActive() so that it
does not fail when an index is closed, similarly to what is done when an
index is deleted.
In `InternalHistogramTests` we were randomizing different values but `minDocCount` was hardcoded to `1`. It's important to test other values, especially `0` as it's the default. To make this possible, the test needed some adapting in the way buckets are randomly generated: all aggs need to share the same `interval`, `minDocCount` and `emptyBucketInfo`. Also assertions need to take into account that more (or less) buckets are expected depending on `minDocCount`.
This was originated by #35921 and its need to test adding empty buckets as part of the reduce phase.
Also relates to #26856 as one more key comparison needed to use `Double.compare` to properly handle `NaN` values, which was triggered by the increased test coverage.
David Turner
Ryan Ernst
Julie Tibshirani
Tim Brooks
Nhat Nguyen
Jim Ferenczi
Christoph Büscher
Adrien Grand
Alexander Reelsen
Tanguy Leroux
Armin Braun
Andrey Ershov
Yannick Welsch
Jake Landis
jaymode
Jay Modi
Mark Tozzi
Andrey Ershov
Alan Woodward
Alpar Torok
Martijn van Groningen
Andy Bristol
João Barbosa
Luca Cavanna
Boaz Leskes
Nik Everett