This commit instruments data frame analytics
with stats for the data that are being analyzed.
In particular, we count training docs, test docs,
and skipped docs.
In order to account docs with missing values as skipped
docs for analyses that do not support missing values,
this commit changes the extractor so that it only ignores
docs with missing values when it collects the data summary,
which is used to estimate memory usage.
Backport of #53998
add 2 additional stats: processing time and processing total which capture the
time spent for processing results and how often it ran. The 2 new stats
correspond to the existing indexing and search stats. Together with indexing
and search this now allows the user to see the full picture, all 3 stages.
This is the first in a series of commits that will introduce the
autoscaling deciders framework. This commit introduces the basic
framework for representing autoscaling decisions.
This commit switches to using an onlyIf to determine if a build Docker
image task execution should occur. This is preferred since it means that
the determination is performed at task execution time, rather than
during configuration.
This commit changes the pre_filter_shard_size default from 128 to unspecified.
This allows to apply heuristics based on the request and the target indices when deciding
whether the can match phase should run or not. When unspecified, this pr runs the can match phase
automatically if one of these conditions is met:
* The request targets more than 128 shards.
* The request contains read-only indices.
* The primary sort of the query targets an indexed field.
Users can opt-out from this behavior by setting the `pre_filter_shard_size` to a static value.
Closes#39835
I created this bug today in #53793. When a `DelayableWriteable` that
references an existing object serializes itself it wasn't taking the
version of the node on the other side of the wire into account. This
fixes that.
Adds multi-class feature importance calculation.
Feature importance objects are now mapped as follows
(logistic) Regression:
```
{
"feature_name": "feature_0",
"importance": -1.3
}
```
Multi-class [class names are `foo`, `bar`, `baz`]
```
{
“feature_name”: “feature_0”,
“importance”: 2.0, // sum(abs()) of class importances
“foo”: 1.0,
“bar”: 0.5,
“baz”: -0.5
},
```
For users to get the full benefit of aggregating and searching for feature importance, they should update their index mapping as follows (before turning this option on in their pipelines)
```
"ml.inference.feature_importance": {
"type": "nested",
"dynamic": true,
"properties": {
"feature_name": {
"type": "keyword"
},
"importance": {
"type": "double"
}
}
}
```
The mapping field name is as follows
`ml.<inference.target_field>.<inference.tag>.feature_importance`
if `inference.tag` is not provided in the processor definition, it is not part of the field path.
`inference.target_field` is defaulted to `ml.inference`.
//cc @lcawl ^ Where should we document this?
If this makes it in for 7.7, there shouldn't be any feature_importance at inference BWC worries as 7.7 is the first version to have it.
It seemed confusing for users that our top-level mapping page still had a
prominent section named 'Mapping Type'. This PR reworks the docs to remove this
reference and adds a note about types removal (similar to the note we added to
other APIs like put mapping).
Today when cluster.remote.connect is set to false, and some aspect of
the codebase tries to get a remote client, today we return a no such
remote cluster exception. This can be quite perplexing to users,
especially if the remote cluster is actually defined in their cluster
state, it is only that the local node is not a remote cluter
client. This commit addresses this by providing a dedicated error
message when a remote cluster is not available because the local node is
not a remote cluster client.
This commit removes the configuration time vs execution time distinction
with regards to certain BuildParms properties. Because of the cost of
determining Java versions for configuration JDK locations we deferred
this until execution time. This had two main downsides. First, we had
to implement all this build logic in tasks, which required a bunch of
additional plumbing and complexity. Second, because some information
wasn't known during configuration time, we had to nest any build logic
that depended on this in awkward callbacks.
We now defer to the JavaInstallationRegistry recently added in Gradle.
This utility uses a much more efficient method for probing Java
installations vs our jrunscript implementation. This, combined with some
optimizations to avoid probing the current JVM as well as deferring
some evaluation via Providers when probing installations for BWC builds
we can maintain effectively the same configuration time performance
while removing a bunch of complexity and runtime cost (snapshotting
inputs for the GenerateGlobalBuildInfoTask was very expensive). The end
result should be a much more responsive build execution in almost all
scenarios.
(cherry picked from commit ecdbd37f2e0f0447ed574b306adb64c19adc3ce1)
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.
If a setting is touched during bootstrap before logging is configured,
and that setting uses a byte size value, the deprecation logger for
ByteSizeValue will be initialized. However, this means a logger will be
configured before log4j is initialized, which we reject at startup. This
commit puts this deprecation logger in a holder pattern so that it is
not initialized until first use, which will happen after logging is
configured.
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)
Feature importance storage format is changing to encompass multi-class.
Feature importance objects are now mapped as follows
(logistic) Regression:
```
{
"feature_name": "feature_0",
"importance": -1.3
}
```
Multi-class [class names are `foo`, `bar`, `baz`]
```
{
“feature_name”: “feature_0”,
“importance”: 2.0, // sum(abs()) of class importances
“foo”: 1.0,
“bar”: 0.5,
“baz”: -0.5
},
```
This change adjusts the mapping creation for analytics so that the field is mapped as a `nested` type.
Native side change: https://github.com/elastic/ml-cpp/pull/1071
This change adds the `nori_number` token filter.
It also adds a `discard_punctuation` option in nori_tokenizer that should be used in conjunction with the new filter.
* Get Async Search: omit _clusters section when empty (#53907)
The _clusters section is omitted by the search API whenever no remote clusters are searched. Async search should do the same, but Get Async Search returns a deserialized response, hence a weird `_clusters` section with all values set to `0` gets returned instead. In fact the recreated Clusters object is not the same object as the EMPTY constant, yet it has the same content.
This commit addresses this by changing the comparison in the `toXContent` method to not print out the section if the number of total clusters is `0`.
* Async search: remove version from response (#53960)
The goal of the version field was to quickly show when you can expect to find something new in the search response, compared to when nothing has changed. This can also be done by looking at the `_shards` section and `num_reduce_phases` returned with the search response. In fact when there has been one or more additional reduction of the results, you can expect new results in the search response. Otherwise, the `_shards` section could notify of additional failures of shards that have completed the query, but that is not a guarantee that their results will be exposed (only when the following partial reduction is performed their results will be available).
That said this commit clarifies this in the docs and removes the version field from the async search response
* Async Search: replicas to auto expand from 0 to 1 (#53964)
This way single node clusters that are green don't go yellow once async search is used, while
all the others still have one replica.
* [DOCS] address timing issue in async search docs tests (#53910)
The docs snippets for submit async search have proven difficult to test as it is not possible to guarantee that you get a response that is not final, even when providing `wait_for_completion=0`. In the docs we want to show though a proper long-running query, and its first response should be partial rather than final.
With this commit we adapt the docs snippets to show a partial response, and replace under the hood all that's needed to make the snippets tests succeed when we get a final response. Also, increased the timeout so we always get a final response.
Closes#53887Closes#53891
Since a data frame analytics job may have associated docs
in the .ml-stats-* indices, when the job is deleted we
should delete those docs too.
Backport of #53933
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.
The test in CloseWhileRelocatingShardsIT failed recently
multiple times (3) when waiting for initial indices to be
become green. Looking at the execution logs from #53544
it appears at the very beginning of the test and when
the WindowsFS file system is picked up (which is known
to slow down tests).
This commit simply increases the timeout for the first
ensureGreen() to 60 seconds. If the test continues to fail,
we might want to test a larger timeout or disable
WindowsFS for this test.
Closes#53544
While `CustomProcessor` is generic and allows for flexibility, there
are new requirements that make cross validation a concept it's hard
to abstract behind custom processor. In particular, we would like to
add data_counts to the DFA jobs stats. Counting training VS. test
docs would be a useful statistic. We would also want to add a
different cross validation strategy for multiclass classification.
This commit renames custom processors to cross validation splitters
which allows for those enhancements without cryptically doing
things as a side effect of the abstract custom processing.
Backport of #53915
DoubleValuesSource is the type-safe replacement for ValueSource in the lucene
core. Most of elasticsearch has moved to use these, but lang-expressions is still
using the old version. This commit migrates lang-expressions as well.
Upgrading AWS SDK to v1.11.749.
Required building clients inside privileged contexts because some class loading that requires privileges now happens there and working around a new SDK bug in the S3 client builder.
Closes#53191
Dimitris Athanasiou
Hendrik Muhs
Mark Vieira
muachilin
Jason Tedor
Tim Vernum
Jim Ferenczi
Nik Everett
Benjamin Trent
Gordon Brown
Julie Tibshirani
James Rodewig
Marios Trivyzas
Christoph Büscher
Namgyu Kim
Przemysław Witek
Dan Hermann
Paweł Krześniak
Luca Cavanna
Ryan Ernst
Lisa Cawley
C.J. Jameson
Armin Braun
Tanguy Leroux
Alan Woodward