The ResourceWatcherService enables watching of files for modifications
and deletions. During startup various consumers register the files that
should be watched by this service. There is behavior that might be
unexpected in that the service may not start polling until later in the
startup process due to the use of lifecycle states to control when the
service actually starts the jobs to monitor resources. This change
removes this unexpected behavior so that upon construction the service
has already registered its tasks to poll resources for changes. In
making this modification, the service no longer extends
AbstractLifecycleComponent and instead implements the Closeable
interface so that the polling jobs can be terminated when the service
is no longer required.
Relates #54867
Backport of #54993
`auto_date_histogram`'s reduction behavior is fairly complex and we have
some fairly complex testing logic for it but it is super difficult to
look at that testing logic and say "ah, that is what it does in this
case". This adds some tests explicit (non-randomized) tests of the
reduction logic that *should* be easier to read.
I've noticed that a lot of our tests are using deprecated static methods
from the Hamcrest matchers. While this is not a big deal in any
objective sense, it seems like a small good thing to reduce compilation
warnings and be ready for a new release of the matcher library if we
need to upgrade. I've also switched a few other methods in tests that
have drop-in replacements.
Currently forbidden apis accounts for 800+ tasks in the build. These
tasks are aggressively created by the plugin. In forbidden apis 3.0, we
will get task avoidance
(https://github.com/policeman-tools/forbidden-apis/pull/162), but we
need to ourselves use the same task avoidance mechanisms to not trigger
these task creations. This commit does that for our foribdden apis
usages, in preparation for upgrading to 3.0 when it is released.
Upgrade to lucene 8.5.1 release that contains a bug fix for a bug that might introduce index corruption when deleting data from an index that was previously shrunk.
We can be a little more efficient when aborting a snapshot. Since we know the new repository
data after finalizing the aborted snapshot when can pass it down to the snapshot completion listeners.
This way, we don't have to fork off to the snapshot threadpool to get the repository data when the listener completes and can directly submit the delete task with high priority straight from the cluster state thread.
Snapshot deletes should first check the cluster state for an in-progress snapshot
and try to abort it before checking the repository contents. This allows for atomically
checking and aborting a snapshot in the same cluster state update, removing all possible
races where a snapshot that is in-progress could not be found if it finishes between
checking the repository contents and the cluster state.
Also removes confusing races, where checking the cluster state off of the cluster state thread
finds an in-progress snapshot that is then not found in the cluster state update to abort it.
Finally, the logic to use the repository generation of the in-progress snapshot + 1 was error
prone because it would always fail the delete when the repository had a pending generation different from its safe generation when a snapshot started (leading to the snapshot finalizing at a
higher generation).
These issues (particularly that last point) can easily be reproduced by running `SLMSnapshotBlockingIntegTests` in a loop with current `master` (see #54766).
The snapshot resiliency test for concurrent snapshot creation and deletion was made to more
aggressively start the delete operation so that the above races would become visible.
Previously, the fact that deletes would never coincide with initializing snapshots resulted
in a number of the above races not reproducing.
This PR is the most consistent I could get snapshot deletes without changes to the state machine. The fact that aborted deletes will not put the delete operation in the cluster state before waiting for the snapshot to abort still allows for some possible (though practically very unlikely) races. These will be fixed by a state-machine change in upcoming work in #54705 (which will have a much simpler and clearer diff after this change).
Closes#54766
* Remove Redundant Cluster State during Snapshot INIT + Master Failover (#54420)
Similar to #54395 we know that a snapshot in INIT state has not
written anything to the repository yet. If we see one from a master
failover, there is no point in moving it to ABORTED before removing it
from the cluster state in a subsequent CS update.
Instead, we can simply remove its job from the CS the first time
we see it on master failover and be done with it.
* Move Snapshot Status Related Method to Appropriate Places
Lots of things living in `SnapshotsService` for no reason other than
that `SnapshotsService` provides the `RepositoriesService`.
Cleaning this up to directly use `RepositoriesService` in the relevant
transport actions and by that shortening the already very complex `SnapshotsService`.
Just like in `AbstractCoordinatorTestCase` we can't just assume the cluster
is stable once all the cluster states align since stray follower/leader check
tasks could still hit us after a disconnect, causing future test operations to fail.
=> fixed by running all tasks in the possible time span of running into these
checks before validating that cluster states align on all nodes to prevent this
like we do in the coordinator tests.
Closes#55103
Provides basic repository-level stats that will allow us to get some insight into how many
requests are actually being made by the underlying SDK. Currently only tracks GET and LIST
calls for S3 repositories. Most of the code is unfortunately boiler plate to add a new endpoint
that will help us better understand some of the low-level dynamics of searchable snapshots.
This is a first cut at giving NodeInfo the ability to carry a flexible
list of heterogeneous info responses. The trick is to be able to
serialize and deserialize an arbitrary list of blocks of information. It
is convenient to be able to deserialize into usable Java objects so that
we can aggregate nodes stats for the cluster stats endpoint.
In order to provide a little bit of clarity about which objects can and
can't be used as info blocks, I've introduced a new interface called
"ReportingService."
I have removed the hard-coded getters (e.g., getOs()) in favor of a
flexible method that can return heterogeneous kinds of info blocks
(e.g., getInfo(OsInfo.class)). Taking a class as an argument removes the
need to cast in the client code.
With this change, when a task is canceled, the task manager will cancel
not only its direct child tasks but all also its descendant tasks.
Closes#50990
Adds support for filters to T-Test aggregation. The filters can be used to
select populations based on some criteria and use values from the same or
different fields.
Closes#53692
The secure_settings_password was never taken into consideration in
the ReloadSecureSettings API. This commit fixes that and adds
necessary REST layer testing. Doing so, it also:
- Allows TestClusters to have a password protected keystore
so that it can be set for tests.
- Adds a parameter to the run task so that elastisearch can
be run with a password protected keystore from source.
The usage of local parameter for GetFieldMappingRequest has been removed from the underlying transport action since v2.0.
This PR deprecates the parameter from rest layer. It will be removed in next major version.
We added a fancy method to provide random realistic test data to the
reduction tests in #54910. This uses that to remove some of the more
esoteric machinations in the agg tests. This will marginally increase
the coverage of the serialiation tests and, more importantly, remove
some mysterious value generation code that only really made sense for
random reduction tests but was used all over the place. It doesn't, on
the other hand, make the tests shorter. Just *hopefully* more clear.
I only cleaned up a few tests this way. If we like this it'd probably be
worth grabbing others.
We found some problems during the test.
Data: 200Million docs, 1 shard, 0 replica
hits | avg | sum | value_count |
----------- | ------- | ------- | ----------- |
20,000 | .038s | .033s | .063s |
200,000 | .127s | .125s | .334s |
2,000,000 | .789s | .729s | 3.176s |
20,000,000 | 4.200s | 3.239s | 22.787s |
200,000,000 | 21.000s | 22.000s | 154.917s |
The performance of `avg`, `sum` and other is very close when performing
statistics, but the performance of `value_count` has always been poor,
even not on an order of magnitude. Based on some common-sense knowledge,
we think that `value_count` and sum are similar operations, and the time
consumed should be the same. Therefore, we have discussed the agg
of `value_count`.
The principle of counting in es is to traverse the field of each
document. If the field is an ordinary value, the count value is
increased by 1. If it is an array type, the count value is increased
by n. However, the problem lies in traversing each document and taking
out the field, which changes from disk to an object in the Java
language. We summarize its current problems with Elasticsearch as:
- Number cast to string overhead, and GC problems caused by a large
number of strings
- After the number type is converted to string, sorting and other
unnecessary operations are performed
Here is the proof of type conversion overhead.
```
// Java long to string source code, getChars is very time-consuming.
public static String toString(long i) {
int size = stringSize(i);
if (COMPACT_STRINGS) {
byte[] buf = new byte[size];
getChars(i, size, buf);
return new String(buf, LATIN1);
} else {
byte[] buf = new byte[size * 2];
StringUTF16.getChars(i, size, buf);
return new String(buf, UTF16);
}
}
```
test type | average | min | max | sum
------------ | ------- | ---- | ----------- | -------
double->long | 32.2ns | 28ns | 0.024ms | 3.22s
long->double | 31.9ns | 28ns | 0.036ms | 3.19s
long->String | 163.8ns | 93ns | 1921 ms | 16.3s
particularly serious.
Our optimization code is actually very simple. It is to manage different
types separately, instead of uniformly converting to string unified
processing. We added type identification in ValueCountAggregator, and
made special treatment for number and geopoint types to cancel their
type conversion. Because the string type is reduced and the string
constant is reduced, the improvement effect is very obvious.
hits | avg | sum | value_count | value_count | value_count | value_count | value_count | value_count |
| | | double | double | keyword | keyword | geo_point | geo_point |
| | | before | after | before | after | before | after |
----------- | ------- | ------- | ----------- | ----------- | ----------- | ----------- | ----------- | ----------- |
20,000 | 38s | .033s | .063s | .026s | .030s | .030s | .038s | .015s |
200,000 | 127s | .125s | .334s | .078s | .116s | .099s | .278s | .031s |
2,000,000 | 789s | .729s | 3.176s | .439s | .348s | .386s | 3.365s | .178s |
20,000,000 | 4.200s | 3.239s | 22.787s | 2.700s | 2.500s | 2.600s | 25.192s | 1.278s |
200,000,000 | 21.000s | 22.000s | 154.917s | 18.990s | 19.000s | 20.000s | 168.971s | 9.093s |
- The results are more in line with common sense. `value_count` is about
the same as `avg`, `sum`, etc., or even lower than these. Previously,
`value_count` was much larger than avg and sum, and it was not even an
order of magnitude when the amount of data was large.
- When calculating numeric types such as `double` and `long`, the
performance is improved by about 8 to 9 times; when calculating the
`geo_point` type, the performance is improved by 18 to 20 times.
Currently the remote cluster sniff connection process can succeed even
if no connections are opened. This commit fixes this by failing the
connection process if no connections are successfully opened.
This commit adds an explicit test of time zone rewrite on date nanos
field. Today this is working but we need tests to ensure that we don't
break it unintentionally.
Makes query result serialization more robust by propagating possible
IOExceptions that can occur during shard level result serialization to the
caller instead of throwing AssertionError that is not intercepted.
Fixes#54665
The use of available processors, the terminology, and the settings
around it have evolved over time. This commit cleans up some places in
the codes and in the docs to adjust to the current terminology.
* Prevent putting V2 index template when overlapping with existing template
This change prevents putting V2 index template when it would overlap with existing V2 template
of the same priority
Relates to #53101
This changes the behavior of aggregations when search is performed
against enough shards to enable "batch reduce" mode. In this case we
force always store aggregations in serialized form rather than a
traditional java reference. This should shrink the memory usage of large
aggregations at the cost of slightly slowing down aggregations where the
coordinating node is also a data node. Because we're only doing this
when there are many shards this is likely to be fairly rare.
As a side effect this lets us add logs for the memory usage of the aggs
buffer:
```
[2020-04-03T17:03:57,052][TRACE][o.e.a.s.SearchPhaseController] [runTask-0] aggs partial reduction [1320->448] max [1320]
[2020-04-03T17:03:57,089][TRACE][o.e.a.s.SearchPhaseController] [runTask-0] aggs partial reduction [1328->448] max [1328]
[2020-04-03T17:03:57,102][TRACE][o.e.a.s.SearchPhaseController] [runTask-0] aggs partial reduction [1328->448] max [1328]
[2020-04-03T17:03:57,103][TRACE][o.e.a.s.SearchPhaseController] [runTask-0] aggs partial reduction [1328->448] max [1328]
[2020-04-03T17:03:57,105][TRACE][o.e.a.s.SearchPhaseController] [runTask-0] aggs final reduction [888] max [1328]
```
These are useful, but you need to keep some things in mind before
trusting them:
1. The buffers are oversized ala Lucene's ArrayUtils. This means that we
are using more space than we need, but probably not much more.
2. Before they are merged the aggregations are inflated into their
traditional Java objects which *probably* take up a lot more space
than the serialized form. That is, after all, the reason why we store
them in serialized form in the first place.
And, just because I can, here is another example of the log:
```
[2020-04-03T17:06:18,731][TRACE][o.e.a.s.SearchPhaseController] [runTask-0] aggs partial reduction [147528->49176] max [147528]
[2020-04-03T17:06:18,750][TRACE][o.e.a.s.SearchPhaseController] [runTask-0] aggs partial reduction [147528->49176] max [147528]
[2020-04-03T17:06:18,809][TRACE][o.e.a.s.SearchPhaseController] [runTask-0] aggs partial reduction [147528->49176] max [147528]
[2020-04-03T17:06:18,827][TRACE][o.e.a.s.SearchPhaseController] [runTask-0] aggs partial reduction [147528->49176] max [147528]
[2020-04-03T17:06:18,829][TRACE][o.e.a.s.SearchPhaseController] [runTask-0] aggs final reduction [98352] max [147528]
```
I got that last one by building a ten shard index with a million docs in
it and running a `sum` in three layers of `terms` aggregations, all on
`long` fields, and with a `batched_reduce_size` of `3`.
`PipelineAggregator`s are only sent across the wire for backwards
compatibility with 7.7.0. `PipelineAggregator` needs to continue to
implement `NamedWriteable` for backwards compatibility but pipeline
aggregations created after 7.7.0 need not implement any of the methods
in that interface because we'll never attempt to call them. So this
creates implementations in `PipelineAggregator` (the base class) that
just throw exceptions.
* HLRC support for Index Templates V2 (#54838)
* HLRC support for Index Templates V2
This change adds High Level Rest Client support for Index Templates V2.
Relates to #53101
* fixed compilation error
Co-authored-by: Elastic Machine <elasticmachine@users.noreply.github.com>
When a new index is rolled over, we check to see whether there are any duplicate alias
configurations in the index template configuration. Additionally, when a new index is created from a
bulk action, we check the templates to see if there are any ingest pipelines that need to be applied
to the index that will be newly created.
Both of these actions previously checked the v1 templates for their settings, they now also check
the v2 index templates, with the v2 index templates taking precendence similar to the way they do
when creating an index.
Relates to #53101
This change reintroduces the system index APIs for Kibana without the
changes made for marking what system indices could be accessed using
these APIs. In essence, this is a partial revert of #53912. The changes
for marking what system indices should be allowed access will be
handled in a separate change.
The APIs introduced here 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.
Relates #52385
Backport of #54858
Today we construct the node environment relatively early in the node
construction process, before we have even constructed the final
environment, which means before the final settings are
available. Rather, we should defer constructing the node environment
until the final environment is available. This commit does that. This
helps delay node environment construction until after the node roles are
properly determined, which is important since the node environment does
some checks on the basis of whether or not the node is neither a data
nor a master node (such nodes should not have index metadata nor shard
data on disk). Note that a consequence of this is that the initial log
line that displays the node name, node ID, and cluster name does not
appear until later in startup (after we have loaded plugins). This seems
okay.
Guava was removed from Elasticsearch many years ago, but remnants of it
remain due to transitive dependencies. When a dependency pulls guava
into the compile classpath, devs can inadvertently begin using methods
from guava without realizing it. This commit moves guava to a runtime
dependency in the modules that it is needed.
Note that one special case is the html sanitizer in watcher. The third
party dep uses guava in the PolicyFactory class signature. However, only
calling a method on the PolicyFactory actually causes the class to be
loaded, a reference alone does not trigger compilation to look at the
class implementation. There we utilize a MethodHandle for invoking the
relevant method at runtime, where guava will continue to exist.
IndexShardIT#testMaybeFlush relies on the assumption that the safe commit
and translog deletion policy have advanced after IndexShard#sync returns .
This assumption does not hold if there's a race with the global checkpoint sync.
Closes#52223
This commit introduces a new `geo` module that is intended
to be contain all the geo-spatial-specific features in server.
As a first step, the responsibility of registering the geo_shape
field mapper is moved to this module.
Co-authored-by: Nicholas Knize <nknize@gmail.com>
This commit moves the action name validation and circuit breaking into
the InboundAggregator. This work is valuable because it lays the
groundwork for incrementally circuit breaking as data is received.
This PR includes the follow behavioral change:
Handshakes contribute to circuit breaking, but cannot be broken. They
currently do not contribute nor are they broken.
In 7.x, an index template will fail to apply if it contains a `_default_`
mapping. Several users have expressed confusion over the fact that loading the
template doesn't show any default mappings. This docs change clarifies that in
order to see all mappings in the template, you must pass `include_type_name`.
Currently the TransportHandshaker has a specialized codepath for sending
a response. In other work, we are going to start having handshakes
contribute to circuit breaking (while not being breakable). This commit
moves in that direction by allowing the handshaker to responding using a
standard TcpTransportChannel similar to other requests.
This removes pipeline aggregators from the aggregation result tree
except for a single field used for backwards compatibility with pre-7.8
versions of Elasticsearch. That field isn't populated unless we are
serializing to pre-7.8 Elasticsearch. So, good news! We no longer build
pipeline aggregators on the data node. Most of the time.
If you didn't explictly set `global_ordinals` execution mode we were
never collecting the information that we needed to select `depth_first`
based on the request so we were always defaulting to `breadth_first`.
This fixes it so we collect the information.
* Remove Unused Snapshot Status Values
This is a left-over from before #41940 when we used the same status enum for the shards
and the snapshots overall. The two removed values were never used on the shard level
so we can simply remove them here.
* Only allow retrieving a single index or component template
This changes the Index Template v2 APIs to only allow retrieving a single "named" entity, where the
named entity can be nothing (return everything), a wildcard (return the ones that match), or the
name of a template.
Relates to #53101
* Throw exception when resource is not found
Co-authored-by: Elastic Machine <elasticmachine@users.noreply.github.com>
Co-authored-by: Elastic Machine <elasticmachine@users.noreply.github.com>
`scripted_metric` did not work with cross cluster search because it
assumed that you'd never perform a partial reduction, serialize the
results, and then perform a final reduction. That
serialized-after-partial-reduction step was broken.
This is also required to support #54758.
Ignacio Vera
Jay Modi
Julie Tibshirani
Nik Everett
William Brafford
Stéphane Campinas
Ryan Ernst
Armin Braun
Mark Vieira
Yannick Welsch
Alan Woodward
Nhat Nguyen
Igor Motov
Ioannis Kakavas
Yang Wang
Tim Brooks
Jim Ferenczi
Jason Tedor
Przemko Robakowski
Dan Hermann
Lee Hinman
Tal Levy