* [ML][Inference] Feature pre-processing objects and functions (#46777)
To support inference on pre-trained machine learning models, some basic feature encoding will be necessary. I am using a named object serialization approach so new encodings/pre-processing steps could be added in the future.
This PR lays down the ground work for 3 basic encodings:
* HotOne
* Target Mean
* Frequency
More feature encodings or pre-processings could be added in the future:
* Handling missing columns
* Standardization
* Label encoding
* etc....
* fixing compilation for namedxcontent tests
When using auto-generated IDs + the ingest drop processor (which looks to be used by filebeat
as well) + coordinating nodes that do not have the ingest processor functionality, this can lead
to a NullPointerException.
The issue is that markCurrentItemAsDropped() is creating an UpdateResponse with no id when
the request contains auto-generated IDs. The response serialization is lenient for our
REST/XContent format (i.e. we will send "id" : null) but the internal transport format (used for
communication between nodes) assumes for this field to be non-null, which means that it can't
be serialized between nodes. Bulk requests with ingest functionality are processed on the
coordinating node if the node has the ingest capability, and only otherwise sent to a different
node. This means that, in order to reproduce this, one needs two nodes, with the coordinating
node not having the ingest functionality.
Closes#46678
This commit reuses the same state processor that is used for autodetect
to parse state output from data frame analytics jobs. We then index the
state document into the state index.
Backport of #46804
It is possible for a running analytics job that its config is removed
from the '.ml-config' index (perhaps the user deleted the entire index,
etc.). In that case the task remains without a matching config. I have
raised #46781 to discuss how to deal with this issue.
This commit focuses on `MlMemoryTracker` and changes it so that when
we get the configs for the running tasks we leniently ignore missing ones.
This at least means memory tracking will keep working for other jobs
if one or more are missing.
In addition, this commit makes the cleanup code for native analytics
tests more robust by explicitly stopping all jobs and force-stopping
if an error occurs. This helps so that a single failing test does
not cause other tests fail due to pending tasks.
Backport of #46789
When the stop API is called while the task is running there is
a chance the task gets marked completed twice. This may cause
undesired side effects, like indexing the progress document a second
time after the stop API has returned (the cause for #46705).
This commit adds a check that the task has not been completed before
proceeding to mark it so. In addition, when we update the task's state
we could get some warnings that the task was missing if the stop API
has been called in the meantime. We now check the errors are
`ResourceNotFoundException` and ignore them if so.
Closes#46705
Backports #46721
This is fixing a bug where if an analytics job is started before any
anomaly detection job is opened, we create an index after the state
write alias.
Instead, we should create the state index and alias before starting
an analytics job and this commit makes sure this is the case.
Backport of #46602
After starting the analytics job and checking its state
the state can be any of "started", "reindexing" or
"analyzing" depending on how quickly the work is done.
Investigating the test failure reported in #45518 it appears that
the datafeed task was not found during a tast state update. There
are only two places where such an update is performed: when we set
the state to `started` and when we set it to `stopping`. We handle
`ResourceNotFoundException` in the latter but not in the former.
Thus the test reveals a rare race condition where the datafeed gets
requested to stop before we managed to update its state to `started`.
I could not reproduce this scenario but it would be my best guess.
This commit catches `ResourceNotFoundException` while updating the
state to `started` and lets the task terminate smoothly.
Closes#45518
Backport of #46495
ML users who upgrade from versions prior to 7.4 to 7.4 or later
will have ML results indices that do not have mappings for the
total_search_time_ms field. Therefore, when searching these
indices we must tolerate this field not having a mapping.
Fixes#46437
This refactors `DataFrameAnalyticsTask` into its own class.
The task has quite a lot of functionality now and I believe it would
make code more readable to have it live as its own class rather than
an inner class of the start action class.
Backport of #46402
* [ML] waiting for ml indices before waiting task assignment testFullClusterRestart
* waiting for a stable cluster after fullrestart
* removing unused imports
The test seems to have been failing due to a race condition between
stopping the task and refreshing the destination index. In particular,
we were going forward with refreshing the destination index even
though the task stopped in the meantime. This was fixed in
request.
Closes#43960
Backport of #46271
Though we allow CCS within datafeeds, users could prevent nodes from accessing remote clusters. This can cause mysterious errors and difficult to troubleshoot.
This commit adds a check to verify that `cluster.remote.connect` is enabled on the current node when a datafeed is configured with a remote index pattern.
* [ML] Regression dependent variable must be numeric
This adds a validation that the dependent variable of a regression
analysis must be numeric.
* Address review comments and fix some problems
In addition to addressing the review comments, this
commit fixes a few issues I found during testing.
In particular:
- if there were mappings for required fields but they were
not included we were not reporting the error
- if explicitly included fields had unsupported types we were
not reporting the error
Unfortunately, I couldn't get those fixed without refactoring
the code in `ExtractedFieldsDetector`.
This commit adds support for `boolean` fields in data frame
analytics (and currently both outlier detection and regression).
The analytics process expects `boolean` fields to be encoded as
integers with 0 or 1 value.
Adds a parameter `training_percent` to regression. The default
value is `100`. When the parameter is set to a value less than `100`,
from the rows that can be used for training (ie. those that have a
value for the dependent variable) we randomly choose whether to actually
use for training. This enables splitting the data into a training set and
the rest, usually called testing, validation or holdout set, which allows
for validating the model on data that have not been used for training.
Technically, the analytics process considers as training the data that
have a value for the dependent variable. Thus, when we decide a training
row is not going to be used for training, we simply clear the row's
dependent variable.
The native process requires that there be a non-zero number of rows to analyze. If the flag --rows 0 is passed to the executable, it throws and does not start.
When building the configuration for the process we should not start the native process if there are no rows.
Adding some logging to indicate what is occurring.
Previously, the stats API reports a progress percentage
for DF analytics tasks that are running and are in the
`reindexing` or `analyzing` state.
This means that when the task is `stopped` there is no progress
reported. Thus, one cannot distinguish between a task that never
run to one that completed.
In addition, there are blind spots in the progress reporting.
In particular, we do not account for when data is loaded into the
process. We also do not account for when results are written.
This commit addresses the above issues. It changes progress
to being a list of objects, each one describing the phase
and its progress as a percentage. We currently have 4 phases:
reindexing, loading_data, analyzing, writing_results.
When the task stops, progress is persisted as a document in the
state index. The stats API now reports progress from in-memory
if the task is running, or returns the persisted document
(if there is one).
* [ML] Adding data frame analytics stats to _usage API (#45820)
* [ML] Adding data frame analytics stats to _usage API
* making the size of analytics stats 10k
* adjusting backport
Regression analysis support missing fields. Even more, it is expected
that the dependent variable has missing fields to the part of the
data frame that is not for training.
This commit allows to declare that an analysis supports missing values.
For such analysis, rows with missing values are not skipped. Instead,
they are written as normal with empty strings used for the missing values.
This also contains a fix to the integration test.
Closes#45425
* [ML] better handle empty results when evaluating regression
* adding new failure test to ml_security black list
* fixing equality check for regression results
We cannot know how long the analysis will take to complete thus we should not have
a timeout. Note that if the process crashes, the result processor will pick the
exception due to the stream closing.
Closes#45723
Changes the order of parameters in Geometries from lat, lon to lon, lat
and moves all Geometry classes are moved to the
org.elasticsearch.geomtery package.
Backport of #45332Closes#45048
* Reenable Integ Tests in native-multi-node-tests
* The tests broken here were likely fixed by #45463 => let's reenable them and see if things run fine again
* Relates #45405, #45455
This commit adds a first draft of a regression analysis
to data frame analytics. There is high probability that
the exact syntax might change.
This commit adds the new analysis type and its parameters as
well as appropriate validation. It also modifies the extractor
and the fields detector to be able to handle categorical fields
as regression analysis supports them.
In the FIPS JVM the JVM default locale seems to leak into places
where it should be overridden. This change skips assertions
in TimestampFormatFinderTests.testGuessIsDayFirstFromLocale
that may be impacted.
Fixes#45140
When doing a fieldwise Levenshtein distance comparison
between CSV rows, this change ignores all fields that
have long values, not just the longest field.
This approach works better for CSV formats that have
multiple freeform text fields rather than just a single
"message" field.
Fixes#45047
If one tries to start a DF analytics job that has already run,
the result will be that the task will fail after reindexing the
dest index from the source index. The results of the prior run
will be gone and the task state is not properly set to failed
with the failure reason.
This commit improves the behavior in this scenario. First, we
set the task state to `failed` in a set of failures that were
missed. Second, a validation is added that if the destination
index exists, it must be empty.
In case closing the process throws an exception we should be catching
it no matter its type. The process may have terminated because of a
fatal error in which case closing the process will throw a server
error, not an `IOException`. If this happens we fail to mark the
persistent task as failed and the task gets in limbo.
As data frame rows with missing values for analyzed fields are skipped,
we can be more efficient by including a query that only picks documents
that have values for all analyzed fields. Besides improving the number
of documents we go through, we also provide a more accurate measurement
of how many rows we need which reduces the memory requirements.
This also adds an integration test that runs outlier detection on data
with missing fields.
TaskListener accepts today Throwable in its onFailure method. Though
looking at where it is called (TransportAction), it can never be
notified of a Throwable.
This commit changes the signature of TaskListener#onFailure so that it
accepts an `Exception` rather than a `Throwable` as second argument.
* Switch from using docvalue_fields to extracting values from _source
where applicable. Doing this means parsing the _source and handling the
numbers parsing just like Elasticsearch is doing it when it's indexing
a document.
* This also introduces a minor limitation: aliases type of fields that
are NOT part of a tree of sub-fields will not be able to be retrieved
anymore. field_caps API doesn't shed any light into a field being an
alias or not and at _source parsing time there is no way to know if a
root field is an alias or not. Fields of the type "a.b.c.alias" can be
extracted from docvalue_fields, only if the field they point to can be
extracted from docvalue_fields. Also, not all fields in a hierarchy of
fields can be evaluated to being an alias.
(cherry picked from commit 8bf8a055e38f00df5f49c8d97f632f69d6e00c2c)
* Mute failing test
tracked in #44552
* mute EvilSecurityTests
tracking in #44558
* Fix line endings in ESJsonLayoutTests
* Mute failing ForecastIT test on windows
Tracking in #44609
* mute BasicRenormalizationIT.testDefaultRenormalization
tracked in #44613
* fix mute testDefaultRenormalization
* Increase busyWait timeout windows is slow
* Mute failure unconfigured node name
* mute x-pack internal cluster test windows
tracking #44610
* Mute JvmErgonomicsTests on windows
Tracking #44669
* mute SharedClusterSnapshotRestoreIT testParallelRestoreOperationsFromSingleSnapshot
Tracking #44671
* Mute NodeTests on Windows
Tracking #44256
Removes the warning suppression -Xlint:-deprecation,-rawtypes,-serial,-try,-unchecked.
Many warnings were unchecked warnings in the test code often because of the use of mocks.
These are suppressed with @SuppressWarning
many classes still use the Streamable constructors of HandledTransportAction,
this commit moves more of those classes to the new Writeable constructors.
relates #34389.
This commit adds constructors to AcknolwedgedRequest subclasses to
implement Writeable.Reader, and ensures all future subclasses implement
the same.
relates #34389
* Migrate ML Actions to use writeable ActionType (#44302)
This commit converts all the StreamableResponseActionType
actions in the ML core module to be ActionType and leverage
the Writeable infrastructure.
* Add Snapshot Lifecycle Management (#43934)
* Add SnapshotLifecycleService and related CRUD APIs
This commit adds `SnapshotLifecycleService` as a new service under the ilm
plugin. This service handles snapshot lifecycle policies by scheduling based on
the policies defined schedule.
This also includes the get, put, and delete APIs for these policies
Relates to #38461
* Make scheduledJobIds return an immutable set
* Use Object.equals for SnapshotLifecyclePolicy
* Remove unneeded TODO
* Implement ToXContentFragment on SnapshotLifecyclePolicyItem
* Copy contents of the scheduledJobIds
* Handle snapshot lifecycle policy updates and deletions (#40062)
(Note this is a PR against the `snapshot-lifecycle-management` feature branch)
This adds logic to `SnapshotLifecycleService` to handle updates and deletes for
snapshot policies. Policies with incremented versions have the old policy
cancelled and the new one scheduled. Deleted policies have their schedules
cancelled when they are no longer present in the cluster state metadata.
Relates to #38461
* Take a snapshot for the policy when the SLM policy is triggered (#40383)
(This is a PR for the `snapshot-lifecycle-management` branch)
This commit fills in `SnapshotLifecycleTask` to actually perform the
snapshotting when the policy is triggered. Currently there is no handling of the
results (other than logging) as that will be added in subsequent work.
This also adds unit tests and an integration test that schedules a policy and
ensures that a snapshot is correctly taken.
Relates to #38461
* Record most recent snapshot policy success/failure (#40619)
Keeping a record of the results of the successes and failures will aid
troubleshooting of policies and make users more confident that their
snapshots are being taken as expected.
This is the first step toward writing history in a more permanent
fashion.
* Validate snapshot lifecycle policies (#40654)
(This is a PR against the `snapshot-lifecycle-management` branch)
With the commit, we now validate the content of snapshot lifecycle policies when
the policy is being created or updated. This checks for the validity of the id,
name, schedule, and repository. Additionally, cluster state is checked to ensure
that the repository exists prior to the lifecycle being added to the cluster
state.
Part of #38461
* Hook SLM into ILM's start and stop APIs (#40871)
(This pull request is for the `snapshot-lifecycle-management` branch)
This change allows the existing `/_ilm/stop` and `/_ilm/start` APIs to also
manage snapshot lifecycle scheduling. When ILM is stopped all scheduled jobs are
cancelled.
Relates to #38461
* Add tests for SnapshotLifecyclePolicyItem (#40912)
Adds serialization tests for SnapshotLifecyclePolicyItem.
* Fix improper import in build.gradle after master merge
* Add human readable version of modified date for snapshot lifecycle policy (#41035)
* Add human readable version of modified date for snapshot lifecycle policy
This small change changes it from:
```
...
"modified_date": 1554843903242,
...
```
To
```
...
"modified_date" : "2019-04-09T21:05:03.242Z",
"modified_date_millis" : 1554843903242,
...
```
Including the `"modified_date"` field when the `?human` field is used.
Relates to #38461
* Fix test
* Add API to execute SLM policy on demand (#41038)
This commit adds the ability to perform a snapshot on demand for a policy. This
can be useful to take a snapshot immediately prior to performing some sort of
maintenance.
```json
PUT /_ilm/snapshot/<policy>/_execute
```
And it returns the response with the generated snapshot name:
```json
{
"snapshot_name" : "production-snap-2019.04.09-rfyv3j9qreixkdbnfuw0ug"
}
```
Note that this does not allow waiting for the snapshot, and the snapshot could
still fail. It *does* record this information into the cluster state similar to
a regularly trigged SLM job.
Relates to #38461
* Add next_execution to SLM policy metadata (#41221)
* Add next_execution to SLM policy metadata
This adds the next time a snapshot lifecycle policy will be executed when
retriving a policy's metadata, for example:
```json
GET /_ilm/snapshot?human
{
"production" : {
"version" : 1,
"modified_date" : "2019-04-15T21:16:21.865Z",
"modified_date_millis" : 1555362981865,
"policy" : {
"name" : "<production-snap-{now/d}>",
"schedule" : "*/30 * * * * ?",
"repository" : "repo",
"config" : {
"indices" : [
"foo-*",
"important"
],
"ignore_unavailable" : true,
"include_global_state" : false
}
},
"next_execution" : "2019-04-15T21:16:30.000Z",
"next_execution_millis" : 1555362990000
},
"other" : {
"version" : 1,
"modified_date" : "2019-04-15T21:12:19.959Z",
"modified_date_millis" : 1555362739959,
"policy" : {
"name" : "<other-snap-{now/d}>",
"schedule" : "0 30 2 * * ?",
"repository" : "repo",
"config" : {
"indices" : [
"other"
],
"ignore_unavailable" : false,
"include_global_state" : true
}
},
"next_execution" : "2019-04-16T02:30:00.000Z",
"next_execution_millis" : 1555381800000
}
}
```
Relates to #38461
* Fix and enhance tests
* Figured out how to Cron
* Change SLM endpoint from /_ilm/* to /_slm/* (#41320)
This commit changes the endpoint for snapshot lifecycle management from:
```
GET /_ilm/snapshot/<policy>
```
to:
```
GET /_slm/policy/<policy>
```
It mimics the ILM path only using `slm` instead of `ilm`.
Relates to #38461
* Add initial documentation for SLM (#41510)
* Add initial documentation for SLM
This adds the initial documentation for snapshot lifecycle management.
It also includes the REST spec API json files since they're sort of
documentation.
Relates to #38461
* Add `manage_slm` and `read_slm` roles (#41607)
* Add `manage_slm` and `read_slm` roles
This adds two more built in roles -
`manage_slm` which has permission to perform any of the SLM actions, as well as
stopping, starting, and retrieving the operation status of ILM.
`read_slm` which has permission to retrieve snapshot lifecycle policies as well
as retrieving the operation status of ILM.
Relates to #38461
* Add execute to the test
* Fix ilm -> slm typo in test
* Record SLM history into an index (#41707)
It is useful to have a record of the actions that Snapshot Lifecycle
Management takes, especially for the purposes of alerting when a
snapshot fails or has not been taken successfully for a certain amount of
time.
This adds the infrastructure to record SLM actions into an index that
can be queried at leisure, along with a lifecycle policy so that this
history does not grow without bound.
Additionally,
SLM automatically setting up an index + lifecycle policy leads to
`index_lifecycle` custom metadata in the cluster state, which some of
the ML tests don't know how to deal with due to setting up custom
`NamedXContentRegistry`s. Watcher would cause the same problem, but it
is already disabled (for the same reason).
* High Level Rest Client support for SLM (#41767)
* High Level Rest Client support for SLM
This commit add HLRC support for SLM.
Relates to #38461
* Fill out documentation tests with tags
* Add more callouts and asciidoc for HLRC
* Update javadoc links to real locations
* Add security test testing SLM cluster privileges (#42678)
* Add security test testing SLM cluster privileges
This adds a test to `PermissionsIT` that uses the `manage_slm` and `read_slm`
cluster privileges.
Relates to #38461
* Don't redefine vars
* Add Getting Started Guide for SLM (#42878)
This commit adds a basic Getting Started Guide for SLM.
* Include SLM policy name in Snapshot metadata (#43132)
Keep track of which SLM policy in the metadata field of the Snapshots
taken by SLM. This allows users to more easily understand where the
snapshot came from, and will enable future SLM features such as
retention policies.
* Fix compilation after master merge
* [TEST] Move exception wrapping for devious exception throwing
Fixes an issue where an exception was created from one line and thrown in another.
* Fix SLM for the change to AcknowledgedResponse
* Add Snapshot Lifecycle Management Package Docs (#43535)
* Fix compilation for transport actions now that task is required
* Add a note mentioning the privileges needed for SLM (#43708)
* Add a note mentioning the privileges needed for SLM
This adds a note to the top of the "getting started with SLM"
documentation mentioning that there are two built-in privileges to
assist with creating roles for SLM users and administrators.
Relates to #38461
* Mention that you can create snapshots for indices you can't read
* Fix REST tests for new number of cluster privileges
* Mute testThatNonExistingTemplatesAreAddedImmediately (#43951)
* Fix SnapshotHistoryStoreTests after merge
* Remove overridden newResponse functions that have been removed
* Fix compilation for backport
* Fix get snapshot output parsing in test
* [DOCS] Add redirects for removed autogen anchors (#44380)
* Switch <tt>...</tt> in javadocs for {@code ...}
Test clusters currently has its own set of logic for dealing with
finding different versions of Elasticsearch, downloading them, and
extracting them. This commit converts testclusters to use the
DistributionDownloadPlugin.
This commit creates new base classes for master node actions whose
response types still implement Streamable. This simplifies both finding
remaining classes to convert, as well as creating new master node
actions that use Writeable for their responses.
relates #34389
* HLRC: Fix '+' Not Correctly Encoded in GET Req.
* Encode `+` correctly as `%2B` in URL paths
* Keep encoding `+` as space in URL parameters
* Closes#33077
This commit moves the Supplier variant of HandledTransportAction to have
a different ordering than the Writeable.Reader variant. The Supplier
version is used for the legacy Streamable, and currently having the
location of the Writeable.Reader vs Supplier in the same place forces
using casts of Writeable.Reader to select the correct super constructor.
This change in ordering allows easier migration to Writeable.Reader.
relates #34389
Now that ML job configs are stored in an index rather than
cluster state, availability of the .ml-config index is very
important to the operation of ML. When a cluster starts up
the ML persistent tasks will be considered for node
assignment very early on. It is best in this case if
assignment is deferred until after the .ml-config index is
available.
The introduction of data frame analytics jobs has made this
problem worse, because anomaly detection jobs already waited
for the primary shards of the .ml-state, .ml-anomalies-shared
and .ml-meta indices to be available before doing node
assignment, and by coincidence this would probably lead to
the primary shards of .ml-config also being searchable. But
data frame analytics jobs had no other index checks prior to
this change.
This fixes problem 2 of #44156
By default, we don't check ranges while indexing geo_shapes. As a
result, it is possible to index geoshapes that contain contain
coordinates outside of -90 +90 and -180 +180 ranges. Such geoshapes
will currently break SQL and ML retrieval mechanism. This commit removes
these restriction from the validator is used in SQL and ML retrieval.
When the ML memory tracker is refreshed and a refresh is
already in progress the idea is that the second and
subsequent refresh requests receive the same response as
the currently in progress refresh.
There was a bug that if a refresh failed then the ML
memory tracker's view of whether a refresh was in progress
was not reset, leading to every subsequent request being
registered to receive a response that would never come.
This change makes the ML memory tracker pass on failures
as well as successes to all interested parties and reset
the list of interested parties so that further refresh
attempts are possible after either a success or failure.
This fixes problem 1 of #44156
Custom timestamp overrides provided to the find_file_structure
endpoint produced an invalid Grok pattern if the fractional
seconds separator was a dot rather than a comma or colon.
This commit fixes that problem and adds tests for this sort
of timestamp override.
Fixes#44110
The count should match the number of all df-analytics that
matched the id in the request. However, we set the count
to the number of df-analytics returned which was bound to the
`size` parameter. This commit fixes this by setting the count
to the count of the `get` response.
A bug was introduced in 6.6.0 when we added support for
rollup indices. Rollup caps does NOT support looking at
remote indices, consequently, since we always look up rollup
caps, the datafeed fails with an error if its config
includes a concrete remote index. (When all remote indices
in a datafeed config are wildcards the problem did not
occur.)
The rollups feature does not support remote indices, so if
there is any remote index in a datafeed config (wildcarded
or not), we can skip the rollup cap checks. This PR
implements that change.
This brings TokenizerFactory into line with CharFilterFactory and TokenFilterFactory,
and removes the need to pass around tokenizer names when building custom analyzers.
As this means that TokenizerFactory is no longer a functional interface, the commit also
adds a factory method to TokenizerFactory to make construction simpler.
This introduces a `failed` state to which the data frame analytics
persistent task is set to when something unexpected fails. It could
be the process crashing, the results processor hitting some error,
etc. The failure message is then captured and set on the task state.
From there, it becomes available via the _stats API as `failure_reason`.
The df-analytics stop API now has a `force` boolean parameter. This allows
the user to call it for a failed task in order to reset it to `stopped` after
we have ensured the failure has been communicated to the user.
This commit also adds the analytics version in the persistent task
params as this allows us to prevent tasks to run on unsuitable nodes in
the future.
Renames `_id_copy` to `ml__id_copy` as field names starting with
underscore are deprecated. The new field name `ml__id_copy` was
chosen as an obscure enough field that users won't have in their data.
Otherwise, this field is only intented to be used by df-analytics.
If a job is opened and then closed and does nothing in
between then it should not persist any results or state
documents. This change adapts the no-op job test to
assert no results in addition to no state, and to log
any documents that cause this assertion to fail.
Relates elastic/ml-cpp#512
Relates #43680
The Action base class currently works for both Streamable and Writeable
response types. This commit intorduces StreamableResponseAction, for
which only the legacy Action implementions which provide newResponse()
will extend. This eliminates the need for overriding newResponse() with
an UnsupportedOperationException.
relates #34389
Since #41817 was merged the ml-cpp zip file for any
given version has been cached indefinitely by Gradle.
This is problematic, particularly in the case of the
master branch where the version 8.0.0-SNAPSHOT will
be in use for more than a year.
This change tells Gradle that the ml-cpp zip file is
a "changing" dependency, and to check whether it has
changed every two hours. Two hours is a compromise
between checking on every build and annoying developers
with slow internet connections and checking rarely
causing bug fixes in the ml-cpp code to take a long
time to propagate through to elasticsearch PRs that
rely on them.
This commit adds support for multiple source indices.
In order to deal with multiple indices having different mappings,
it attempts a best-effort approach to merge the mappings assuming
there are no conflicts. In case conflicts exists an error will be
returned.
To allow users creating custom mappings for special use cases,
the destination index is now allowed to exist before the analytics
job runs. In addition, settings are no longer copied except for
the `index.number_of_shards` and `index.number_of_replicas`.
* Deduplicate org.elasticsearch.xpack.core.dataframe.utils.TimeUtils and org.elasticsearch.xpack.core.ml.utils.time.TimeUtils into a common class: org.elasticsearch.xpack.core.common.time.TimeUtils.
* Add unit tests for parseTimeField and parseTimeFieldToInstant methods
This change introduces a new setting,
xpack.ml.process_connect_timeout, to enable
the timeout for one of the external ML processes
to connect to the ES JVM to be increased.
The timeout may need to be increased if many
processes are being started simultaneously on
the same machine. This is unlikely in clusters
with many ML nodes, as we balance the processes
across the ML nodes, but can happen in clusters
with a single ML node and a high value for
xpack.ml.node_concurrent_job_allocations.
This merges the initial work that adds a framework for performing
machine learning analytics on data frames. The feature is currently experimental
and requires a platinum license. Note that the original commits can be
found in the `feature-ml-data-frame-analytics` branch.
A new set of APIs is added which allows the creation of data frame analytics
jobs. Configuration allows specifying different types of analysis to be performed
on a data frame. At first there is support for outlier detection.
The APIs are:
- PUT _ml/data_frame/analysis/{id}
- GET _ml/data_frame/analysis/{id}
- GET _ml/data_frame/analysis/{id}/_stats
- POST _ml/data_frame/analysis/{id}/_start
- POST _ml/data_frame/analysis/{id}/_stop
- DELETE _ml/data_frame/analysis/{id}
When a data frame analytics job is started a persistent task is created and started.
The main steps of the task are:
1. reindex the source index into the dest index
2. analyze the data through the data_frame_analyzer c++ process
3. merge the results of the process back into the destination index
In addition, an evaluation API is added which packages commonly used metrics
that provide evaluation of various analysis:
- POST _ml/data_frame/_evaluate
The error message if the native controller failed to run
(for example due to running Elasticsearch on an unsupported
platform) was not easy to understand. This change removes
pointless detail from the message and adds some hints about
likely causes.
Fixes#42341
This commit replaces usages of Streamable with Writeable for the
AcknowledgedResponse and its subclasses, plus associated actions.
Note that where possible response fields were made final and default
constructors were removed.
This is a large PR, but the change is mostly mechanical.
Relates to #34389
Backport of #43414
After the network disruption a partition is created,
one side of which can form a cluster the other can't.
Ensure requests are sent to a node on the correct side
of the cluster
This commit removes some very old test logging annotations that appeared
to be added to investigate test failures that are long since closed. If
these are needed, they can be added back on a case-by-case basis with a
comment associating them to a test failure.
* Return 0 for negative "free" and "total" memory reported by the OS
We've had a situation where the MX bean reported negative values for the
free memory of the OS, in those rare cases we want to return a value of
0 rather than blowing up later down the pipeline.
In the event that there is a serialization or creation error with regard
to memory use, this adds asserts so the failure will occur as soon as
possible and give us a better location for investigation.
Resolves#42157
* Fix test passing in invalid memory value
* Fix another test passing in invalid memory value
* Also change mem check in MachineLearning.machineMemoryFromStats
* Add background documentation for why we prevent negative return values
* Clarify comment a bit more
This trace logging looks like it was copy/pasted from another test,
where the logging in that test was only added to investigate a test
failure. This commit removes the trace logging.
The ML failover tests sometimes need to wait for jobs to be
assigned to new nodes following a node failure. They wait
10 seconds for this to happen. However, if the node that
failed was the master node and a new master was elected then
this 10 seconds might not be long enough as a refresh of the
memory stats will delay job assignment. Once the memory
refresh completes the persistent task will be assigned when
the next cluster state update occurs or after the periodic
recheck interval, which defaults to 30 seconds. Rather than
increase the length of the wait for assignment to 31 seconds,
this change decreases the periodic recheck interval to 1
second.
Fixes#43289
We were stopping a node in the cluster at a time when
the replica shards of the .ml-state index might not
have been created. This change moves the wait for
green status to a point where the .ml-state index
exists.
Fixes#40546Fixes#41742
Forward port of #43111
A static code analysis revealed that we are not closing
the input stream in the post_data endpoint. This
actually makes no difference in practice, as the
particular InputStream implementation in this case is
org.elasticsearch.common.bytes.BytesReferenceStreamInput
and its close() method is a no-op. However, it is good
practice to close the stream anyway.
The machine learning feature of xpack has native binaries with a
different commit id than the rest of code. It is currently exposed in
the xpack info api. This commit adds that commit information to the ML
info api, so that it may be removed from the info api.
Previously 10 digit numbers were considered candidates to be
timestamps recorded as seconds since the epoch and 13 digit
numbers as timestamps recorded as milliseconds since the epoch.
However, this meant that we could detect these formats for
numbers that would represent times far in the future. As an
example ISBN numbers starting with 9 were detected as milliseconds
since the epoch since they had 13 digits.
This change tweaks the logic for detecting such timestamps to
require that they begin with 1 or 2. This means that numbers
that would represent times beyond about 2065 are no longer
detected as epoch timestamps. (We can add 3 to the definition
as we get closer to the cutoff date.)
The description field of xpack featuresets is optionally part of the
xpack info api, when using the verbose flag. However, this information
is unnecessary, as it is better left for documentation (and the existing
descriptions describe anything meaningful). This commit removes the
description field from feature sets.
The tests for the ML TimeoutChecker rely on threads
not being interrupted after the TimeoutChecker is
closed. This change ensures this by making the
close() and setTimeoutExceeded() methods synchronized
so that the code inside them cannot execute
simultaneously.
Fixes#43097
* [ML] Adding support for geo_shape, geo_centroid, geo_point in datafeeds
* only supporting doc_values for geo_point fields
* moving validation into GeoPointField ctor
Get resources action sorts on the resource id. When there are no resources at
all, then it is possible the index does not contain a mapping for the resource
id field. In that case, the search api fails by default.
This commit adjusts the search request to ignore unmapped fields.
Closeselastic/kibana#37870
Both TransportAnalyzeAction and CategorizationAnalyzer have logic to build
custom analyzers for index-independent analysis. A lot of this code is duplicated,
and it requires the AnalysisRegistry to expose a number of internal provider
classes, as well as making some assumptions about when analysis components are
constructed.
This commit moves the build logic directly into AnalysisRegistry, reducing the
registry's API surface considerably.
Previously, a reindex request had two different size specifications in the body:
* Outer level, determining the maximum documents to process
* Inside the source element, determining the scroll/batch size.
The outer level size has now been renamed to max_docs to
avoid confusion and clarify its semantics, with backwards compatibility and
deprecation warnings for using size.
Similarly, the size parameter has been renamed to max_docs for
update/delete-by-query to keep the 3 interfaces consistent.
Finally, all 3 endpoints now support max_docs in both body and URL.
Relates #24344
A static code analysis revealed that we are not closing
the input stream in the find_file_structure endpoint.
This actually makes no difference in practice, as the
particular InputStream implementation in this case is
org.elasticsearch.common.bytes.BytesReferenceStreamInput
and its close() method is a no-op. However, it is good
practice to close the stream anyway.
This change adds the earliest and latest timestamps into
the field stats for fields of type "date" in the output of
the ML find_file_structure endpoint. This will enable the
cards for date fields in the file data visualizer in the UI
to be made to look more similar to the cards for date
fields in the index data visualizer in the UI.
Dots in the column names cause an error in the ingest
pipeline, as dots are special characters in ingest pipeline.
This PR changes dots into underscores in CSV field names
suggested by the ML find_file_structure endpoint _unless_
the field names are specifically overridden. The reason for
allowing them in overrides is that fields that are not
mentioned in the ingest pipeline can contain dots. But it's
more consistent that the default behaviour is to replace
them all.
Fixeselastic/kibana#26800
When analysing a semi-structured text file the
find_file_structure endpoint merges lines to form
multi-line messages using the assumption that the
first line in each message contains the timestamp.
However, if the timestamp is misdetected then this
can lead to excessive numbers of lines being merged
to form massive messages.
This commit adds a line_merge_size_limit setting
(default 10000 characters) that halts the analysis
if a message bigger than this is created. This
prevents significant CPU time being spent subsequently
trying to determine the internal structure of the
huge bogus messages.
This change helps to prevent the situation where a binary
file uploaded to the find_file_structure endpoint is
detected as being text in the UTF-16 character set, and
then causes a large amount of CPU to be spent analysing
the bogus text structure.
The approach is to check the distribution of zero bytes
between odd and even file positions, on the grounds that
UTF-16BE or UTF16-LE would have a very skewed distribution.
This change fixes a race condition that would result in an
in-memory data structure becoming out-of-sync with persistent
tasks in cluster state.
If repeated often enough this could result in it being
impossible to open any ML jobs on the affected node, as the
master node would think the node had capacity to open another
job but the chosen node would error during the open sequence
due to its in-memory data structure being full.
The race could be triggered by opening a job and then closing
it a tiny fraction of a second later. It is unlikely a user
of the UI could open and close the job that fast, but a script
or program calling the REST API could.
The nasty thing is, from the externally observable states and
stats everything would appear to be fine - the fast open then
close sequence would appear to leave the job in the closed
state. It's only later that the leftovers in the in-memory
data structure might build up and cause a problem.
This change contains a major refactoring of the timestamp
format determination code used by the ML find file structure
endpoint.
Previously timestamp format determination was done separately
for each piece of text supplied to the timestamp format finder.
This had the drawback that it was not possible to distinguish
dd/MM and MM/dd in the case where both numbers were 12 or less.
In order to do this sensibly it is best to look across all the
available timestamps and see if one of the numbers is greater
than 12 in any of them. This necessitates making the timestamp
format finder an instantiable class that can accumulate evidence
over time.
Another problem with the previous approach was that it was only
possible to override the timestamp format to one of a limited
set of timestamp formats. There was no way out if a file to be
analysed had a timestamp that was sane yet not in the supported
set. This is now changed to allow any timestamp format that can
be parsed by a combination of these Java date/time formats:
yy, yyyy, M, MM, MMM, MMMM, d, dd, EEE, EEEE, H, HH, h, mm, ss,
a, XX, XXX, zzz
Additionally S letter groups (fractional seconds) are supported
providing they occur after ss and separated from the ss by a dot,
comma or colon. Spacing and punctuation is also permitted with
the exception of the question mark, newline and carriage return
characters, together with literal text enclosed in single quotes.
The full list of changes/improvements in this refactor is:
- Make TimestampFormatFinder an instantiable class
- Overrides must be specified in Java date/time format - Joda
format is no longer accepted
- Joda timestamp formats in outputs are now derived from the
determined or overridden Java timestamp formats, not stored
separately
- Functionality for determining the "best" timestamp format in
a set of lines has been moved from TextLogFileStructureFinder
to TimestampFormatFinder, taking advantage of the fact that
TimestampFormatFinder is now an instantiable class with state
- The functionality to quickly rule out some possible Grok
patterns when looking for timestamp formats has been changed
from using simple regular expressions to the much faster
approach of using the Shift-And method of sub-string search,
but using an "alphabet" consisting of just 1 (representing any
digit) and 0 (representing non-digits)
- Timestamp format overrides are now much more flexible
- Timestamp format overrides that do not correspond to a built-in
Grok pattern are mapped to a %{CUSTOM_TIMESTAMP} Grok pattern
whose definition is included within the date processor in the
ingest pipeline
- Grok patterns that correspond to multiple Java date/time
patterns are now handled better - the Grok pattern is accepted
as matching broadly, and the required set of Java date/time
patterns is built up considering all observed samples
- As a result of the more flexible acceptance of Grok patterns,
when looking for the "best" timestamp in a set of lines
timestamps are considered different if they are preceded by
a different sequence of punctuation characters (to prevent
timestamps far into some lines being considered similar to
timestamps near the beginning of other lines)
- Out-of-the-box Grok patterns that are considered now include
%{DATE} and %{DATESTAMP}, which have indeterminate day/month
ordering
- The order of day/month in formats with indeterminate day/month
order is determined by considering all observed samples (plus
the server locale if the observed samples still do not suggest
an ordering)
Relates #38086Closes#35137Closes#35132
This adds the node name where we fail to start a process via the native
controller to facilitate debugging as otherwise it might not be known
to which node the job was allocated.
Moves the test infrastructure away from using node.max_local_storage_nodes, allowing us in a
follow-up PR to deprecate this setting in 7.x and to remove it in 8.0.
This also changes the behavior of InternalTestCluster so that starting up nodes will not automatically
reuse data folders of previously stopped nodes. If this behavior is desired, it needs to be explicitly
done by passing the data path from the stopped node to the new node that is started.
Re-enable muted tests and accommodate recent backend changes
that result in higher memory usage being reported for a job
at the start of its life-cycle
This corrects what appears to have been a copy-paste error
where the logger for `MachineLearning` and `DataFrame` was wrongly
set to be that of `XPackPlugin`.
The date_histogram accepts an interval which can be either a calendar
interval (DST-aware, leap seconds, arbitrary length of months, etc) or
fixed interval (strict multiples of SI units). Unfortunately this is inferred
by first trying to parse as a calendar interval, then falling back to fixed
if that fails.
This leads to confusing arrangement where `1d` == calendar, but
`2d` == fixed. And if you want a day of fixed time, you have to
specify `24h` (e.g. the next smallest unit). This arrangement is very
error-prone for users.
This PR adds `calendar_interval` and `fixed_interval` parameters to any
code that uses intervals (date_histogram, rollup, composite, datafeed, etc).
Calendar only accepts calendar intervals, fixed accepts any combination of
units (meaning `1d` can be used to specify `24h` in fixed time), and both
are mutually exclusive.
The old interval behavior is deprecated and will throw a deprecation warning.
It is also mutually exclusive with the two new parameters. In the future the
old dual-purpose interval will be removed.
The change applies to both REST and java clients.
Muting a number of AutoDetectMemoryLimitIT tests to give CI a chance to
settle before easing in required backend changes.
relates elastic/ml-cpp#486
relates #42086
Improve the hard_limit memory audit message by reporting how many bytes
over the configured memory limit the job was at the point of the last
allocation failure.
Previously the model memory usage was reported, however this was
inaccurate and hence of limited use - primarily because the total
memory used by the model can decrease significantly after the models
status is changed to hard_limit but before the model size stats are
reported from autodetect to ES.
While this PR contains the changes to the format of the hard_limit audit
message it is dependent on modifications to the ml-cpp backend to
send additional data fields in the model size stats message. These
changes will follow in a subsequent PR. It is worth noting that this PR
must be merged prior to the ml-cpp one, to keep CI tests happy.
This change replaces the extremely unfriendly message
"Number of messages analyzed must be positive" in the
case where the sample lines were incorrectly grouped
into just one message to an error that more helpfully
explains the likely root cause of the problem.
The run task is supposed to run elasticsearch with the given plugin or
module. However, for modules, this is most realistic if using the full
distribution. This commit changes the run setup to use the default or
oss as appropriate.
This switches the strategy used to download machine learning artifacts
from a manual download through S3 to using an Ivy repository on top of
S3. This gives us all the benefits of Gradle dependency resolution
including local caching.
* [ML] Refactor NativeStorageProvider to enable reuse
Moves `NativeStorageProvider` as a machine learning component
so that it can be reused for other job types. Also, we now
pass the persistent task description as unique identifier which
avoids conflicts between jobs of different type but with same ids.
* Adding nativeStorageProvider as component
Since `TransportForecastJobAction` is expected to get injected a `NativeStorageProvider` class, we need to make sure that it is a constructed component, as it does not have a zero parametered, public ctor.
The date_histogram internally converts obsolete timezones (such as
"Canada/Mountain") into their modern equivalent ("America/Edmonton").
But rollup just stored the TZ as provided by the user.
When checking the TZ for query validation we used a string comparison,
which would fail due to the date_histo's upgrading behavior.
Instead, we should convert both to a TimeZone object and check if their
rules are compatible.
Values higher than 100% are now allowed to accommodate use
cases where swapping has been determined to be acceptable.
Anomaly detector jobs only use their full model memory
during background persistence, and this is deliberately
staggered, so with large numbers of jobs few will generally
be persisting state at the same time. Settings higher than
available memory are only recommended for OEM type
situations where a wrapper tightly controls the types of
jobs that can be created, and each job alone is considerably
smaller than what each node can handle.
* [ML] Add validation that rejects duplicate detectors in PutJobAction
Closes#39704
* Add YML integration test for duplicate detectors fix.
* Use "== false" comparison rather than "!" operator.
* Refine error message to sound more natural.
* Put job description in square brackets in the error message.
* Use the new validation in ValidateJobConfigAction.
* Exclude YML tests for new validation from permission tests.
* Replace usages RandomizedTestingTask with built-in Gradle Test (#40978)
This commit replaces the existing RandomizedTestingTask and supporting code with Gradle's built-in JUnit support via the Test task type. Additionally, the previous workaround to disable all tasks named "test" and create new unit testing tasks named "unitTest" has been removed such that the "test" task now runs unit tests as per the normal Gradle Java plugin conventions.
(cherry picked from commit 323f312bbc829a63056a79ebe45adced5099f6e6)
* Fix forking JVM runner
* Don't bump shadow plugin version
The invalid license enforced is exposed to the cluster state update
thread (via the license state listener) before the constructor has
finished. This violates the JLS for safe publication of an object, and
means there is a concurrency bug lurking here. This commit addresses
this by avoiding publication of the invalid license enforcer before the
constructor has returned.
This change adds information about which UI path
(if any) created ML anomaly detector jobs to the
stats returned by the _xpack/usage endpoint.
Counts for the following possibilities are expected:
* ml_module_apache_access
* ml_module_apm_transaction
* ml_module_auditbeat_process_docker
* ml_module_auditbeat_process_hosts
* ml_module_nginx_access
* ml_module_sample
* multi_metric_wizard
* population_wizard
* single_metric_wizard
* unknown
The "unknown" count is for jobs that do not have a
created_by setting in their custom_settings.
Closes#38403
Ensure that there is at least a 1s delay between the time that state
is persisted by each of the two jobs in the test.
Model snapshot IDs use the current time in epoch seconds to
distinguish themselves, hence snapshots will be overwritten
by another if it occurs in the same 1s window.
Closes#40347
If multiple jobs are created together and the anomaly
results index does not exist then some of the jobs could
fail to update the mappings of the results index. This
lead them to fail to write their results correctly later.
Although this scenario sounds rare, it is exactly what
happens if the user creates their first jobs using the
Nginx module in the ML UI.
This change fixes the problem by updating the mappings
of the results index if it is found to exist during a
creation attempt.
Fixes#38785
* [ML] Refactor common utils out of ML plugin to XPack.Core
* implementing GET filters with abstract transport
* removing added rest param
* adjusting how defaults can be supplied
The problem here was that `DatafeedJob` was updating the last end time searched
based on the `now` even though when there are aggregations, the extactor will
only search up to the floor of `now` against the histogram interval.
This commit fixes the issue by using the end time as calculated by the extractor.
It also adds an integration test that uses aggregations. This test would fail
before this fix. Unfortunately the test is slow as we need to wait for the
datafeed to work in real time.
Closes#39842
* [ML] refactoring lazy query and agg parsing
* Clean up and addressing PR comments
* removing unnecessary try/catch block
* removing bad call to logger
* removing unused import
* fixing bwc test failure due to serialization and config migrator test
* fixing style issues
* Adjusting DafafeedUpdate class serialization
* Adding todo for refactor in v8
* Making query non-optional so it does not write a boolean byte
This change does the following:
1. Makes the per-node setting xpack.ml.max_open_jobs
into a cluster-wide dynamic setting
2. Changes the job node selection to continue to use the
per-node attributes storing the maximum number of open
jobs if any node in the cluster is older than 7.1, and
use the dynamic cluster-wide setting if all nodes are on
7.1 or later
3. Changes the docs to reflect this
4. Changes the thread pools for native process communication
from fixed size to scaling, to support the dynamic nature
of xpack.ml.max_open_jobs
5. Renames the autodetect thread pool to the job comms
thread pool to make clear that it will be used for other
types of ML jobs (data frame analytics in particular)
Backport of #39320
Investigating how to make DeleteExpiredDataIT faster, it was
revealed that the security audit trail threads were quite hot.
Disabling that seems to be helping quite a bit with making this
test faster. This commit also unmutes the test to see how it goes
with the audit trail disabled.
Relates #39658Closes#39575
ML has historically used doc as the single mapping type but reindex in 7.x
will change the mapping to _doc. Switching to the typeless APIs handles
case where the mapping type is either doc or _doc. This change removes
deprecated typed usages.
This commit parallelizes some parts of the test
and its remove an unnecessary refresh call.
On my local machine it shaves off about 15 seconds
for a test execution time of ~64s (down from ~80s).
This test is still slow but progress over perfection.
Relates #37339
The ScheduledEvent class has never preserved the time
zone so it makes more sense for it to store the start and
end time using Instant rather than ZonedDateTime.
Closes#38620
The assertBusy() that waits the default 10 seconds for a
datafeed to complete very occasionally times out on slow
machines. This commit increases the timeout to 60 seconds.
It will almost never actually take this long, but it's
better to have a timeout that will prevent time being
wasted looking at spurious test failures.
This test failed once in a very long time with the assertion
that there is no document for the `non_existing_job` in the
state index. I could not see how that is possible and I cannot
reproduce. With this commit the failure message will reveal
some examples of the left behind docs which might shed a light
about what could go wrong.
The ML memory tracker does searches against ML results
and config indices. These searches can be asynchronous,
and if they are running while the node is closing then
they can cause problems for other components.
This change adds a stop() method to the MlMemoryTracker
that waits for in-flight searches to complete. Once
stop() has returned the MlMemoryTracker will not kick
off any new searches.
The MlLifeCycleService now calls MlMemoryTracker.stop()
before stopping stopping the node.
Fixes#37117
These two changes are interlinked.
Before this change unsetting ML upgrade mode would wait for all
datafeeds to be assigned and not waiting for their corresponding
jobs to initialise. However, this could be inappropriate, if
there was a reason other that upgrade mode why one job was unable
to be assigned or slow to start up. Unsetting of upgrade mode
would hang in this case.
This change relaxes the condition for considering upgrade mode to
be unset to simply that an assignment attempt has been made for
each ML persistent task that did not fail because upgrade mode
was enabled. Thus after unsetting upgrade mode there is no
guarantee that every ML persistent task is assigned, just that
each is not unassigned due to upgrade mode.
In order to make setting upgrade mode work immediately after
unsetting upgrade mode it was then also necessary to make it
possible to stop a datafeed that was not assigned. There was
no particularly good reason why this was not allowed in the past.
It is trivial to stop an unassigned datafeed because it just
involves removing the persistent task.
The .ml-annotations index is created asynchronously when
some other ML index exists. This can interfere with the
post-test index deletion, as the .ml-annotations index
can be created after all other indices have been deleted.
This change adds an ML specific post-test cleanup step
that runs before the main cleanup and:
1. Checks if any ML indices exist
2. If so, waits for the .ml-annotations index to exist
3. Deletes the other ML indices found in step 1.
4. Calls the super class cleanup
This means that by the time the main post-test index
cleanup code runs:
1. The only ML index it has to delete will be the
.ml-annotations index
2. No other ML indices will exist that could trigger
recreation of the .ml-annotations index
Fixes#38952
* Fix#38623 remove xpack namespace REST API
Except for xpack.usage and xpack.info API's, this moves the last remaining API's out of the xpack namespace
* rename xpack api's inside inside the files as well
* updated yaml tests references to xpack namespaces api's
* update callsApi calls in the IT subclasses
* make sure docs testing does not use xpack namespaced api's
* fix leftover xpack namespaced method names in docs/build.gradle
* found another leftover reference
(cherry picked from commit ccb5d934363c37506b76119ac050a254fa80b5e7)
* ML: update set_upgrade_mode, add logging
* Attempt to fix datafeed isolation
Also renamed a few methods/variables for clarity and added
some comments
Tanguy Leroux
Benjamin Trent
Yannick Welsch
Dimitris Athanasiou
Przemysław Witek
David Roberts
Igor Motov
Armin Braun
Mark Vieira
Przemyslaw Gomulka
Jason Tedor
Luca Cavanna
David Kyle
Andrei Stefan
Alpar Torok
Tal Levy
Ryan Ernst
Lee Hinman
Alan Woodward
Christoph Büscher
Martijn van Groningen
Ryan Ernst
David Turner
Henning Andersen
Ed Savage
Zachary Tong
Ed Savage
Tom Veasey
Yogesh Gaikwad
Zachary Tong
Iana Bondarska
Mehran Koushkebaghi
Martijn Laarman