When a value is promoted in the LRU cache, its weight is removed and added.
The LocalModel object was recalculating the model size for ever weight check, which caused a polynomial runtime.
This commit changes the model size to only be calculated in the LocalModel ctor.
It appears that occasionally 30 seconds are not enough for CI workers
to complete DFA jobs. In order to eliminate such failures we increase
the time we wait for DFA jobs to complete in integration tests to
60 seconds.
Fixes#64926
Backport of #65126
This commit fixes two problems:
- When extracting a doc value, we allow boolean scalars to be used as input
- The output order of processed feature names is deterministic. Previous custom one hot fields used to be non-deterministic and thus could cause weird bugs.
If we encounter an exception during extracting data in a data
frame analytics job, we retry once. However, we were not catching
exceptions thrown from processing the search response. This may
result in an infinite loop that causes a stack overflow.
This commit fixes this problem.
Backport of #64947
When using custom processors, the field names extracted from the documents are not the
same as the feature names used for training.
Consequently, it is possible for the stratified sampler to have an incorrect view of the feature rows.
This can lead to the wrong column being read for the class label, and thus throw errors on training
row extraction.
This commit changes the training row feature names used by the stratified sampler so that it matches
the names (and their order) that are sent to the analytics process.
It is possible that a value mapped as a `keyword` has any scalar value type. This includes any numerical value, String, or boolean.
This commit allows `boolean` types to be considered as a part of the categorical feature collection when this is the case.
This field is added from version 7.11 onwards. We are
adding it to the list of ignored fields for data frame analytics
in 7.10 to avoid failing to start an outlier detection job
in a mixed cluster environment.
Relates #64503
ForecastIT.testOverflowToDisk has been observed to fail a few
times in FIPS JVMs because it takes longer than the permitted
30 seconds. This PR bumps the timeout up to 60 seconds.
Fixes#63793
Adds support for the unsigned_long type to data frame analytics.
This type is handled in the same way as the long type. Values
sent to the ML native processes are converted to floats and
hence will lose accuracy when outside the range where a float
can uniquely represent long values.
Backport of #64066
When calculating feature importance, the leaf values directly correlate the value of the importance.
Consequently, positive leaf values -> positive feature importance
negative leaf values -> negative feature importance.
It follows that for binary classification, this is done such that the importance relates to the leaf values, which relate directly to the "probability of class 1".
So, the feature importance calculated is always for the importance as it relates to class 1.
The inverse is the importance as it relates to class 0.
Now that deprecation logs get indexed to a data stream, if we
do not load the data stream plugin in our tests and any test
generates a deprecation log message then millions of exceptions
get logged, slowing down the tests to the extent that they can
fail.
This change loads the data streams plugin during the ML internal
cluster tests. (It should already be present in external cluster
tests.)
Fixes#63548
Adds validation that the dest pipeline exists when a transform
is updated. Refactors the pipeline check into the `SourceDestValidator`.
Fixes#59587
Backport of #63494
This PR adds deprecation warnings when accessing System Indices via the REST layer. At this time, these warnings are only enabled for Snapshot builds by default, to allow projects external to Elasticsearch additional time to adjust their access patterns.
Deprecation warnings will be triggered by all REST requests which access registered System Indices, except for purpose-specific APIs which access System Indices as an implementation detail a few specific APIs which will continue to allow access to system indices by default:
- `GET _cluster/health`
- `GET {index}/_recovery`
- `GET _cluster/allocation/explain`
- `GET _cluster/state`
- `POST _cluster/reroute`
- `GET {index}/_stats`
- `GET {index}/_segments`
- `GET {index}/_shard_stores`
- `GET _cat/[indices,aliases,health,recovery,shards,segments]`
Deprecation warnings for accessing system indices take the form:
```
this request accesses system indices: [.some_system_index], but in a future major version, direct access to system indices will be prevented by default
```
this adds the new field `feature_importance_baseline` and allows it to be optionally be included in the model's metadata.
Related to: https://github.com/elastic/ml-cpp/pull/1522
* [ML] optimize delete expired snapshots (#63134)
When deleting expired snapshots, we do an individual delete action per snapshot per job.
We should instead gather the expired snapshots and delete them in a single call.
This commit achieves this and a side-effect is there is less audit log spam on nightly cleanup
closes https://github.com/elastic/elasticsearch/issues/62875
* [ML] renames */inference* apis to */trained_models* (#63097)
This commit renames all `inference` CRUD APIs to `trained_models`.
This aligns with internal terminology, documentation, and use-cases.
When an opened anomaly detection job is updated with a detection
rule that references a filter, apart from updating the c++ process
with the rule, we also need to update it with the referenced filter.
This commit fixes a bug which led to the job not applying such updates
on-the-fly.
Fixes#62948
Backport of #63057
Data frame analytics results format changed in version `7.10.0`.
If existing jobs that were not completed are restarted, it is
possible the destination index had already been created. That index's
mappings are not suitable for the new results format.
This commit checks the version of the destination index and deletes
it when the version is outdated. The job will then continue by
recreating the destination index and reindexing.
Backport of #62960
This PR adds timeouts to the named pipe connections of the
autodetect, normalize and data_frame_analyzer processes.
This argument requires the changes of elastic/ml-cpp#1514 in
order to work, so that PR will be merged before this one.
(The controller process already had a different mechanism,
tied to the ES JVM lifetime.)
Backport of #62993
* [ML] fixing testTwoJobsWithSameRandomizeSeedUseSameTrainingSet tests (#62976)
This fixes the two test failures.
The shard failure seems to be due to the .ml-stats index being in the middle of being created.
As we have decided top level importance for classification is not useful,
it has been removed from the results from the training job. This commit
also removes them from inference.
Backport of #62486
for get trained models include_model_definition is now deprecated.
This commit writes a deprecation warning if that parameter is used and suggests the caller to utilize the replacement
When state persistence was first implemented for data frame analytics
we had the assumption that state would always fit in a single document.
However this is not the case any more.
This commit adds handling of state that spreads over multiple documents.
Backport of #62564
This fixes reindexing progress in the scenario when a DFA job that had not finished
reindexing is resumed (either because the user called stop and start or because the
job was reassigned in the middle of reindexing). Before the fix reindexing progress
stays to the value it had reached before until it surpasses that value.
When we resume a data frame analytics job we want to preserve reindexing progress
and reset all other phases. Except for when reindexing was not completed.
In that case we are deleting the destination index and starting reindexing
from scratch. Thus we need to reset reindexing progress too.
Backport of #62772
* [ML] changing to not use global bulk indexing parameters in conjunction with add(object) calls (#62694)
* [ML] changing to not use global bulk indexing parameters in conjunction with add(object) calls
global parameters, outside of the global index, are ignored for internal callers in certain cases.
If the interal caller is adding requests via the following methods:
```
- BulkRequest#add(IndexRequest)
- BulkRequest#add(UpdateRequest)
- BulkRequest#add(DocWriteRequest)
- BulkRequest#add(DocWriteRequest[])
```
It is better to specifically set the desired parameters on the requests before they are added
to the bulk request object.
This commit addresses this issue for the ML plugin
* unmuting test
This reworks the code around grok's built-in patterns to name things
more like the rest of the code. Its not a big deal, but I'm just more
used to having `public static final` constants in SHOUTING_SNAKE_CASE.
This commit adjusts the following APIs so now they not only support an `_all` case, but wildcard patterned Ids as well.
- `GET _ml/calendars/<calendar_id>/events`
- `GET _ml/calendars/<calendar_id>`
- `GET _ml/anomaly_detectors/<job_id>/model_snapshots/<snapshot_id>`
- `DELETE _ml/anomaly_detectors/<job_id>/_forecast/<forecast_id>`
* [ML] Add new include flag to GET inference/<model_id> API for model training metadata (#61922)
Adds new flag include to the get trained models API
The flag initially has two valid values: definition, total_feature_importance.
Consequently, the old include_model_definition flag is now deprecated.
When total_feature_importance is included, the total_feature_importance field is included in the model metadata object.
Including definition is the same as previously setting include_model_definition=true.
* fixing test
* Update x-pack/plugin/core/src/test/java/org/elasticsearch/xpack/core/ml/action/GetTrainedModelsRequestTests.java
This commit address some build failures from the perspective of Intellij.
These changes include:
* changing an order of a dependency definition that seems to can cause Intellij build to fail.
* introduction of an abstract class out of the test source set (seems to be an issue sharing
classes cross projects with non-standard source sets.
* a couple of missing dependency definitions (not sure how the command line worked prior to this)
Removes methods that were no longer used regarding version 5.4 doc ids of ModelState.
Also adds clean up of 5.4 model state and quantile docs in the daily maintenance.
Backport of #62434
Benjamin Trent
Dimitris Athanasiou
Przemysław Witek
Daniel Mitterdorfer
David Roberts
Luca Cavanna
Tim Vernum
Gordon Brown
Igor Motov
David Kyle
Nik Everett
Andrei Dan
Ignacio Vera
Jake Landis