This drop the "top level" pipeline aggregators from the aggregation
result tree which should save a little memory and a few serialization
bytes. Perhaps more imporantly, this provides a mechanism by which we
can remove *all* pipelines from the aggregation result tree. This will
save quite a bit of space when pipelines are deep in the tree.
Sadly, doing this isn't simple because of backwards compatibility. Nodes
before 7.7.0 *need* those pipelines. We provide them by setting passing
a `Supplier<PipelineTree>` into the root of the aggregation tree that we
only call if we need to serialize to a version before 7.7.0.
This solution works for cross cluster search because we always reduce
the aggregations in each remote cluster and then forward them back to
the coordinating node. Its quite possible that the coordinating node
needs the pipeline (say it is version 7.1.0) and the gateway node in the
remote cluster doesn't (version 7.7.0). In that case the data nodes
won't send the pipeline aggregations back to the gateway node.
Critically, the gateway node *will* send the pipeline aggregations back
to the coordinating node. This is all managed with that
`Supplier<PipelineTree>`, but *how* it is managed is a bit tricky.
Fixes an issue where the elasticsearch-node command-line tools would not work correctly
because PersistentTasksCustomMetaData contains named XContent from plugins. This PR
makes it so that the parsing for all custom metadata is skipped, even if the core system would
know how to handle it.
Closes#53549
DocsClientYamlTestSuiteIT sometimes fails for CCR
related tests because tests are started before the license
is fully applied and active within the cluster. The first
tests to be executed then fails with the error noticed
in #53430. This can be easily reproduced locally by
only running CCR docs tests.
This commit adds some @Before logic in
DocsClientYamlTestSuiteIT so that it waits for the
license to be active before running CCR tests.
Closes#53430
This moves the pipeline aggregation validation from the data node to the
coordinating node so that we, eventually, can stop sending pipeline
aggregations to the data nodes entirely. In fact, it moves it into the
"request validation" stage so multiple errors can be accumulated and
sent back to the requester for the entire request. We can't always take
advantage of that, but it'll be nice for folks not to have to play
whack-a-mole with validation.
This is implemented by replacing `PipelineAggretionBuilder#validate`
with:
```
protected abstract void validate(ValidationContext context);
```
The `ValidationContext` handles the accumulation of validation failures,
provides access to the aggregation's siblings, and implements a few
validation utility methods.
Benchmarking showed that the effect of the ExitableDirectoryReader
is reduced considerably when checking every 8191 docs. Moreover,
set the cancellable task before calling QueryPhase#preProcess()
and make sure we don't wrap with an ExitableDirectoryReader at all
when lowLevelCancellation is set to false to avoid completely any
performance impact.
Follows: #52822
Follows: #53166
Follows: #53496
(cherry picked from commit cdc377e8e74d3ca6c231c36dc5e80621aab47c69)
Use sequence numbers and force merge UUID to determine whether a shard has changed or not instead before falling back to comparing files to get incremental snapshots on primary fail-over.
This change adds a "grant API key action"
POST /_security/api_key/grant
that creates a new API key using the privileges of one user ("the
system user") to execute the action, but creates the API key with
the roles of the second user ("the end user").
This allows a system (such as Kibana) to create API keys representing
the identity and access of an authenticated user without requiring
that user to have permission to create API keys on their own.
This also creates a new QA project for security on trial licenses and runs
the API key tests there
Backport of: #52886
Today in the `CoordinatorTests` each node uses multiple threadpools. This is
mostly fine as they are almost completely stateless, except for the
`ThreadContext`: by using multiple threadpools we cannot make assertions that
the thread context is/isn't preserved as we expect. This commit consolidates
the threadpool instances in use so that each node uses just one.
This fixes two issues:
1. Currently, the future here is never resolved on assertion error so a failing test would take a full minute
to complete until the future times out.
2. S3 tests overide this method to busy assert on this method. This only works if an assertion error makes it
to the calling thread.
Closes#53508
It's simple to deprecate a field used in an ObjectParser just by adding deprecation
markers to the relevant ParseField objects. The warnings themselves don't currently
have any context - they simply say that a deprecated field has been used, but not
where in the input xcontent it appears. This commit adds the parent object parser
name and XContentLocation to these deprecation messages.
Note that the context is automatically stripped from warning messages when they
are asserted on by integration tests and REST tests, because randomization of
xcontent type during these tests means that the XContentLocation is not constant
Today cluster states are sometimes (rarely) applied in the default context
rather than system context, which means that any appliers which capture their
contexts cannot do things like remote transport actions when security is
enabled.
There are at least two ways that we end up applying the cluster state in the
default context:
1. locally applying a cluster state that indicates that the master has failed
2. the elected master times out while waiting for a response from another node
This commit ensures that cluster states are always applied in the system
context.
Mitigates #53751
* Adds per context settings:
`script.context.${CONTEXT}.cache_max_size` ~
`script.cache.max_size`
`script.context.${CONTEXT}.cache_expire` ~
`script.cache.expire`
`script.context.${CONTEXT}.max_compilations_rate` ~
`script.max_compilations_rate`
* Context cache is used if:
`script.max_compilations_rate=use-context`. This
value is dynamically updatable, so users can
switch back to the general cache if desired.
* Settings for context caches take the first value
that applies:
1) Context specific settings if set, eg
`script.context.ingest.cache_max_size`
2) Correlated general setting is set to the non-default
value, eg `script.cache.max_size`
3) Context default
The reason for 2's inclusion is to allow an easy
transition for users who've customized their general
cache settings.
Using the general cache settings for the context caches
results in higher effective settings, since they are
multiplied across the number of contexts. So a general
cache max size of 200 will become 200 * # of contexts.
However, this behavior it will avoid users snapping to a
value that is too low for them.
Backport of: #52855
Refs: #50152
Re-applies the change from #53523 along with test fixes.
closes#53626closes#53624closes#53622closes#53625
Co-authored-by: Nik Everett <nik9000@gmail.com>
Co-authored-by: Lee Hinman <dakrone@users.noreply.github.com>
Co-authored-by: Jake Landis <jake.landis@elastic.co>
Today it can happen that a transport message fails to send (for example,
because a transport interceptor rejects the request). In this case, the
response handler is never invoked, which can lead to necessary cleanups
not being performed. There are two ways to handle this. One is to expect
every callsite that sends a message to try/catch these exceptions and
handle them appropriately. The other is merely to invoke the response
handler to handle the exception, which is already equipped to handle
transport exceptions.
This begins to clean up how `PipelineAggregator`s and executed.
Previously, we would create the `PipelineAggregator`s on the data nodes
and embed them in the aggregation tree. When it came time to execute the
pipeline aggregation we'd use the `PipelineAggregator`s that were on the
first shard's results. This is inefficient because:
1. The data node needs to make the `PipelineAggregator` only to
serialize it and then throw it away.
2. The coordinating node needs to deserialize all of the
`PipelineAggregator`s even though it only needs one of them.
3. You end up with many `PipelineAggregator` instances when you only
really *need* one per pipeline.
4. `PipelineAggregator` needs to implement serialization.
This begins to undo these by building the `PipelineAggregator`s directly
on the coordinating node and using those instead of the
`PipelineAggregator`s in the aggregtion tree. In a follow up change
we'll stop serializing the `PipelineAggregator`s to node versions that
support this behavior. And, one day, we'll be able to remove
`PipelineAggregator` from the aggregation result tree entirely.
Importantly, this doesn't change how pipeline aggregations are declared
or parsed or requested. They are still part of the `AggregationBuilder`
tree because *that* makes sense.
This change introduces a new API in x-pack basic that allows to track the progress of a search.
Users can submit an asynchronous search through a new endpoint called `_async_search` that
works exactly the same as the `_search` endpoint but instead of blocking and returning the final response when available, it returns a response after a provided `wait_for_completion` time.
````
GET my_index_pattern*/_async_search?wait_for_completion=100ms
{
"aggs": {
"date_histogram": {
"field": "@timestamp",
"fixed_interval": "1h"
}
}
}
````
If after 100ms the final response is not available, a `partial_response` is included in the body:
````
{
"id": "9N3J1m4BgyzUDzqgC15b",
"version": 1,
"is_running": true,
"is_partial": true,
"response": {
"_shards": {
"total": 100,
"successful": 5,
"failed": 0
},
"total_hits": {
"value": 1653433,
"relation": "eq"
},
"aggs": {
...
}
}
}
````
The partial response contains the total number of requested shards, the number of shards that successfully returned and the number of shards that failed.
It also contains the total hits as well as partial aggregations computed from the successful shards.
To continue to monitor the progress of the search users can call the get `_async_search` API like the following:
````
GET _async_search/9N3J1m4BgyzUDzqgC15b/?wait_for_completion=100ms
````
That returns a new response that can contain the same partial response than the previous call if the search didn't progress, in such case the returned `version`
should be the same. If new partial results are available, the version is incremented and the `partial_response` contains the updated progress.
Finally if the response is fully available while or after waiting for completion, the `partial_response` is replaced by a `response` section that contains the usual _search response:
````
{
"id": "9N3J1m4BgyzUDzqgC15b",
"version": 10,
"is_running": false,
"response": {
"is_partial": false,
...
}
}
````
Asynchronous search are stored in a restricted index called `.async-search` if they survive (still running) after the initial submit. Each request has a keep alive that defaults to 5 days but this value can be changed/updated any time:
`````
GET my_index_pattern*/_async_search?wait_for_completion=100ms&keep_alive=10d
`````
The default can be changed when submitting the search, the example above raises the default value for the search to `10d`.
`````
GET _async_search/9N3J1m4BgyzUDzqgC15b/?wait_for_completion=100ms&keep_alive=10d
`````
The time to live for a specific search can be extended when getting the progress/result. In the example above we extend the keep alive to 10 more days.
A background service that runs only on the node that holds the first primary shard of the `async-search` index is responsible for deleting the expired results. It runs every hour but the expiration is also checked by running queries (if they take longer than the keep_alive) and when getting a result.
Like a normal `_search`, if the http channel that is used to submit a request is closed before getting a response, the search is automatically cancelled. Note that this behavior is only for the submit API, subsequent GET requests will not cancel if they are closed.
Asynchronous search are not persistent, if the coordinator node crashes or is restarted during the search, the asynchronous search will stop. To know if the search is still running or not the response contains a field called `is_running` that indicates if the task is up or not. It is the responsibility of the user to resume an asynchronous search that didn't reach a final response by re-submitting the query. However final responses and failures are persisted in a system index that allows
to retrieve a response even if the task finishes.
````
DELETE _async_search/9N3J1m4BgyzUDzqgC15b
````
The response is also not stored if the initial submit action returns a final response. This allows to not add any overhead to queries that completes within the initial `wait_for_completion`.
The `.async-search` index is a restricted index (should be migrated to a system index in +8.0) that is accessible only through the async search APIs. These APIs also ensure that only the user that submitted the initial query can retrieve or delete the running search. Note that admins/superusers would still be able to cancel the search task through the task manager like any other tasks.
Relates #49091
Co-authored-by: Luca Cavanna <javanna@users.noreply.github.com>
Using a Long alone is not strong enough for the id of search contexts
because we reset the id generator whenever a data node is restarted.
This can lead to two issues:
1. Fetch phase can fetch documents from another index
2. A scroll search can return documents from another index
This commit avoids these issues by adding a UUID to SearchContexId.
This commit introduces hidden aliases. These are similar to hidden
indices, in that they are not visible by default, unless explicitly
specified by name or by indicating that hidden indices/aliases are
desired.
The new alias property, `is_hidden` is implemented similarly to
`is_write_index`, except that it must be consistent across all indices
with a given alias - that is, all indices with a given alias must
specify the alias as either hidden, or all specify it as non-hidden,
either explicitly or by omitting the `is_hidden` property.
This commit updates the template used for watch history indices with
the hidden index setting so that new indices will be created as hidden.
Relates #50251
Backport of #52962
When we test backwards compatibility we often end up in a situation
where we *sometimes* get a warning, and sometimes don't. Like, we won't
get the warning if we're testing against an older version, but we will
in a newer one. Or we won't get the warning if the request randomly
lands on a node with an old version of the code. But we wouldn't if it
randomed into a node with newer code.
This adds `allowed_warnings` to our yaml test runner for those cases:
warnings declared this way are "allowed" but not "required".
Blocks #52959
Co-authored-by: Benjamin Trent <ben.w.trent@gmail.com>
Implement an Exitable DirectoryReader that wraps the original
DirectoryReader so that when a search task is cancelled the
DirectoryReaders also stop their work fast. This is usuful for
expensive operations like wilcard/prefix queries where the
DirectoryReaders can spend lots of time and consume resources,
as previously their work wouldn't stop even though the original
search task was cancelled (e.g. because of timeout or dropped client
connection).
(cherry picked from commit 67acaf61f33bc5f54e26541514d07e375c202e03)
Tests in GoogleCloudStorageBlobStoreRepositoryTests are known
to be flaky on JDK 8 (#51446, #52430 ) and we suspect a JDK
bug (https://bugs.openjdk.java.net/browse/JDK-8180754) that triggers
some assertion on the server side logic that emulates the Google
Cloud Storage service.
Sadly we were not able to reproduce the failures, even when using
the same OS (Debian 9, Ubuntu 16.04) and JDK (Oracle Corporation
1.8.0_241 [Java HotSpot(TM) 64-Bit Server VM 25.241-b07]) of
almost all the test failures on CI. While we spent some time fixing
code (#51933, #52431) to circumvent the JDK bug they are still flaky
on JDK-8. This commit mute these tests for JDK-8 only.
Close ##52906
This commit introduces a module for Kibana that exposes REST APIs that
will be used by Kibana for access to its system indices. These APIs are wrapped
versions of the existing REST endpoints. A new setting is also introduced since
the Kibana system indices' names are allowed to be changed by a user in case
multiple instances of Kibana use the same instance of Elasticsearch.
Additionally, the ThreadContext has been extended to indicate that the use of
system indices may be allowed in a request. This will be built upon in the future
for the protection of system indices.
Backport of #52385
Generalize how queries on `_index` are handled at rewrite time (#52486)
Since this change refactors rewrites, I also took it as an opportunity to adrress #49254: instead of returning the same queries you would get on a keyword field when a field is unmapped, queries get rewritten to a MatchNoDocsQueryBuilder.
This change exposed a couple bugs, like the fact that the percolator doesn't rewrite queries at query time, or that the significant_terms aggregation doesn't rewrite its inner filter, which I fixed.
Closes#49254
In #42838 we moved the terms index of all fields off-heap except the
`_id` field because we were worried it might make indexing slower. In
general, the indexing rate is only affected if explicit IDs are used, as
otherwise Elasticsearch almost never performs lookups in the terms
dictionary for the purpose of indexing. So it's quite wasteful to
require the terms index of `_id` to be loaded on-heap for users who have
append-only workloads. Furthermore I've been conducting benchmarks when
indexing with explicit ids on the http_logs dataset that suggest that
the slowdown is low enough that it's probably not worth forcing the terms
index to be kept on-heap. Here are some numbers for the median indexing
rate in docs/s:
| Run | Master | Patch |
| --- | ------- | ------- |
| 1 | 45851.2 | 46401.4 |
| 2 | 45192.6 | 44561.0 |
| 3 | 45635.2 | 44137.0 |
| 4 | 46435.0 | 44692.8 |
| 5 | 45829.0 | 44949.0 |
And now heap usage in MB for segments:
| Run | Master | Patch |
| --- | ------- | -------- |
| 1 | 41.1720 | 0.352083 |
| 2 | 45.1545 | 0.382534 |
| 3 | 41.7746 | 0.381285 |
| 4 | 45.3673 | 0.412737 |
| 5 | 45.4616 | 0.375063 |
Indexing rate decreased by 1.8% on average, while memory usage decreased
by more than 100x.
The `http_logs` dataset contains small documents and has a simple
indexing chain. More complex indexing chains, e.g. with more fields,
ingest pipelines, etc. would see an even lower decrease of indexing rate.
We consider index level read_only_allow_delete blocks temporary since
the DiskThresholdMonitor can automatically release those when an index
is no longer allocated on nodes above high threshold.
The rest status has therefore been changed to 429 when encountering this
index block to signal retryability to clients.
Related to #49393
This commit renames ElasticsearchAssertions#assertThrows to
assertRequestBuilderThrows and assertFutureThrows to avoid a
naming clash with JUnit 4.13+ and static imports of these methods.
Additionally, these methods have been updated to make use of
expectThrows internally to avoid duplicating the logic there.
Relates #51787
Backport of #52582
Phase 1 of adding compilation limits per context.
* Refactor rate limiting and caching into separate class,
`ScriptCache`, which will be used per context.
* Disable compilation limit for certain tests.
Backport of 0866031
Refs: #50152
Cache latest `RepositoryData` on heap when it's absolutely safe to do so (i.e. when the repository is in strictly consistent mode).
`RepositoryData` can safely be assumed to not grow to a size that would cause trouble because we often have at least two copies of it loaded at the same time when doing repository operations. Also, concurrent snapshot API status requests currently load it independently of each other and so on, making it safe to cache on heap and assume as "small" IMO.
The benefits of this move are:
* Much faster repository status API calls
* listing all snapshot names becomes instant
* Other operations are sped up massively too because they mostly operate in two steps: load repository data then load multiple other blobs to get the additional data
* Additional cloud cost savings
* Better resiliency, saving another spot where an IO issue could break the snapshot
* We can simplify a number of spots in the current code that currently pass around the repository data in tricky ways to avoid loading it multiple times in follow ups.
* Refactor Inflexible Snapshot Repository BwC (#52365)
Transport the version to use for a snapshot instead of whether to use shard generations in the snapshots in progress entry. This allows making upcoming repository metadata changes in a flexible manner in an analogous way to how we handle serialization BwC elsewhere.
Also, exposing the version at the repository API level will make it easier to do BwC relevant changes in derived repositories like source only or encrypted.
Currently we have three different implementations representing a
`ConnectionManager`. There is the basic `ConnectionManager` which
holds all connections for a cluster. And a remote connection manager
which support proxy behavior. And a stubbable connection manager for
tests. The remote and stubbable instances use the delegate pattern,
so this commit extracts an interface for them all to implement.
It looks like #52000 is caused by a slowdown in cluster state application
(maybe due to #50907) but I would like to understand the details to ensure that
there's nothing else going on here too before simply increasing the timeout.
This commit enables some relevant `DEBUG` loggers and also captures stack
traces from all threads rather than just the three hottest ones.
Closes#52450. `setRandomIndexSettings(...)` in `ESIntegTestCase` was
using both a thread-local and a supplied source of randomness. Fix this
method to only use the supplied source.
The `top_metrics` agg is kind of like `top_hits` but it only works on
doc values so it *should* be faster.
At this point it is fairly limited in that it only supports a single,
numeric sort and a single, numeric metric. And it only fetches the "very
topest" document worth of metric. We plan to support returning a
configurable number of top metrics, requesting more than one metric and
more than one sort. And, eventually, non-numeric sorts and metrics. The
trick is doing those things fairly efficiently.
Co-Authored by: Zachary Tong <zach@elastic.co>
This commit makes the names of fetch subphases more consistent:
* Now the names end in just 'Phase', whereas before some ended in
'FetchSubPhase'. This matches the query subphases like AggregationPhase.
* Some names include 'fetch' like FetchScorePhase to avoid ambiguity about what
they do.
This adds a builder and parsed results for the `string_stats`
aggregation directly to the high level rest client. Without this the
HLRC can't access the `string_stats` API without the elastic licensed
`analytics` module.
While I'm in there this adds a few of our usual unit tests and
modernizes the parsing.
Add a new cluster setting `search.allow_expensive_queries` which by
default is `true`. If set to `false`, certain queries that have
usually slow performance cannot be executed and an error message
is returned.
- Queries that need to do linear scans to identify matches:
- Script queries
- Queries that have a high up-front cost:
- Fuzzy queries
- Regexp queries
- Prefix queries (without index_prefixes enabled
- Wildcard queries
- Range queries on text and keyword fields
- Joining queries
- HasParent queries
- HasChild queries
- ParentId queries
- Nested queries
- Queries on deprecated 6.x geo shapes (using PrefixTree implementation)
- Queries that may have a high per-document cost:
- Script score queries
- Percolate queries
Closes: #29050
(cherry picked from commit a8b39ed842c7770bd9275958c9f747502fd9a3ea)
Modifies SLM's and ILM's history indices to be hidden indices for added
protection against accidental querying and deletion, and improves
IndexTemplateRegistry to handle upgrading index templates.
Also modifies the REST test cleanup to delete hidden indices.
Today we use `cluster.join.timeout` to prevent nodes from waiting indefinitely
if joining a faulty master that is too slow to respond, and
`cluster.publish.timeout` to allow a faulty master to detect that it is unable
to publish its cluster state updates in a timely fashion. If these timeouts
occur then the node restarts the discovery process in an attempt to find a
healthier master.
In the special case of `discovery.type: single-node` there is no point in
looking for another healthier master since the single node in the cluster is
all we've got. This commit suppresses these timeouts and instead lets the node
wait for joins and publications to succeed no matter how long this might take.
We can just put the `IndexId` instead of just the index name into the recovery soruce and
save one load of `RepositoryData` on each shard restore that way.
Nik Everett
Yannick Welsch
Tanguy Leroux
Marios Trivyzas
Ryan Ernst
Armin Braun
Tim Vernum
David Turner
Alan Woodward
Stuart Tettemer
Jason Tedor
Jim Ferenczi
Nhat Nguyen
Gordon Brown
Jay Modi
Adrien Grand
bellengao
Przemysław Witek
Tim Brooks
Rory Hunter
Julie Tibshirani