Skip to content

Commit

Permalink
[7.x] [ML] adding support for composite aggs in anomaly detection (#6…
Browse files Browse the repository at this point in the history 
…9970) (#71052)

* [ML] adding support for composite aggs in anomaly detection (#69970)

This commit allows for composite aggregations in datafeeds.

Composite aggs provide a much better solution for having influencers, partitions, etc. on high volume data. Instead of worrying about long scrolls in the datafeed, the calculation is distributed across cluster via the aggregations.

The restrictions for this support are as follows:

- The composite aggregation must have EXACTLY one `date_histogram` source
- The sub-aggs of the composite aggregation must have a `max` aggregation on the SAME timefield as the aforementioned `date_histogram` source
- The composite agg must be the ONLY top level agg and it cannot have a `composite` or `date_histogram` sub-agg
- If using a `date_histogram` to bucket time, it cannot have a `composite` sub-agg.
- The top-level `composite` agg cannot have a sibling pipeline agg. Pipeline aggregations are supported as a sub-agg (thus a pipeline agg INSIDE the bucket).

Some key user interaction differences:
- Speed + resources used by the cluster should be controlled by the `size` parameter in the `composite` aggregation. Previously, we said if you are using aggs, use a specific `chunking_config`. But, with composite, that is not necessary.
- Users really shouldn't use nested `terms` aggs anylonger. While this is still a "valid" configuration and MAY be desirable for some users (only wanting the top 10 of certain terms), typically when users want influencers, partition fields, etc. they want the ENTIRE population. Previously, this really wasn't possible with aggs, with `composite` it is.
- I cannot really think of a typical usecase that SHOULD ever use a multi-bucket aggregation that is NOT supported by composite.
  • Loading branch information
@benwtrent
benwtrent authored Mar 30, 2021
1 parent 1c74f2a commit abb182d
Show file tree
Hide file tree
Showing 31 changed files with 2,071 additions and 298 deletions.
1 change: 1 addition & 0 deletions docs/reference/aggregations/bucket/composite-aggregation.asciidoc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -848,6 +848,7 @@ GET /_search
--------------------------------------------------
// TESTRESPONSE[s/\.\.\.//]

[[search-aggregations-bucket-composite-aggregation-pipeline-aggregations]]
==== Pipeline aggregations

The composite agg is not currently compatible with pipeline aggregations, nor does it make sense in most cases.
Expand Down
265 changes: 210 additions & 55 deletions docs/reference/ml/anomaly-detection/ml-configuring-aggregations.asciidoc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -11,57 +11,61 @@ distributes these calculations across your cluster. You can then feed this
aggregated data into the {ml-features} instead of raw results, which
reduces the volume of data that must be considered while detecting anomalies.

TIP: If you use a terms aggregation and the cardinality of a term is high, the
aggregation might not be effective and you might want to just use the default
search and scroll behavior.
TIP: If you use a terms aggregation and the cardinality of a term is high but
still significantly less than your total number of documents, use
{ref}/search-aggregations-bucket-composite-aggregation.html[composite aggregations]
experimental:[Support for composite aggregations inside datafeeds is currently experimental].

[discrete]
[[aggs-limits-dfeeds]]
== Requirements and limitations

There are some limitations to using aggregations in {dfeeds}. Your aggregation
must include a `date_histogram` aggregation, which in turn must contain a `max`
aggregation on the time field. This requirement ensures that the aggregated data
is a time series and the timestamp of each bucket is the time of the last record
in the bucket.
There are some limitations to using aggregations in {dfeeds}.

IMPORTANT: The name of the aggregation and the name of the field that the agg
operates on need to match, otherwise the aggregation doesn't work. For example,
if you use a `max` aggregation on a time field called `responsetime`, the name
Your aggregation must include a `date_histogram` aggregation or a top level `composite` aggregation,
which in turn must contain a `max` aggregation on the time field.
This requirement ensures that the aggregated data is a time series and the timestamp
of each bucket is the time of the last record in the bucket.

IMPORTANT: The name of the aggregation and the name of the field that it
operates on need to match, otherwise the aggregation doesn't work. For example,
if you use a `max` aggregation on a time field called `responsetime`, the name
of the aggregation must be also `responsetime`.

You must also consider the interval of the date histogram aggregation carefully.
The bucket span of your {anomaly-job} must be divisible by the value of the
`calendar_interval` or `fixed_interval` in your aggregation (with no remainder).
If you specify a `frequency` for your {dfeed}, it must also be divisible by this
interval. {anomaly-jobs-cap} cannot use date histograms with an interval
measured in months because the length of the month is not fixed. {dfeeds-cap}
tolerate weeks or smaller units.
You must consider the interval of the `date_histogram` or `composite`
aggregation carefully. The bucket span of your {anomaly-job} must be divisible
by the value of the `calendar_interval` or `fixed_interval` in your aggregation
(with no remainder). If you specify a `frequency` for your {dfeed},
it must also be divisible by this interval. {anomaly-jobs-cap} cannot use
`date_histogram` or `composite` aggregations with an interval measured in months
because the length of the month is not fixed; they can use weeks or smaller units.

TIP: As a rule of thumb, if your detectors use <<ml-metric-functions,metric>> or
<<ml-sum-functions,sum>> analytical functions, set the date histogram
<<ml-sum-functions,sum>> analytical functions, set the `date_histogram` or `composite`
aggregation interval to a tenth of the bucket span. This suggestion creates
finer, more granular time buckets, which are ideal for this type of analysis. If
your detectors use <<ml-count-functions,count>> or <<ml-rare-functions,rare>>
functions, set the interval to the same value as the bucket span.

If your <<aggs-dfeeds,{dfeed} uses aggregations with nested `terms` aggs>> and
model plot is not enabled for the {anomaly-job}, neither the **Single Metric
Viewer** nor the **Anomaly Explorer** can plot and display an anomaly
chart for the job. In these cases, the charts are not visible and an explanatory
If your <<aggs-dfeeds,{dfeed} uses aggregations with nested `terms` aggs>> and
model plot is not enabled for the {anomaly-job}, neither the **Single Metric
Viewer** nor the **Anomaly Explorer** can plot and display an anomaly
chart for the job. In these cases, the charts are not visible and an explanatory
message is shown.

When the aggregation interval of the {dfeed} and the bucket span of the job
don't match, the values of the chart plotted in both the **Single Metric
Viewer** and the **Anomaly Explorer** differ from the actual values of the job.
To avoid this behavior, make sure that the aggregation interval in the {dfeed}
configuration and the bucket span in the {anomaly-job} configuration have the
When the aggregation interval of the {dfeed} and the bucket span of the job
don't match, the values of the chart plotted in both the **Single Metric
Viewer** and the **Anomaly Explorer** differ from the actual values of the job.
To avoid this behavior, make sure that the aggregation interval in the {dfeed}
configuration and the bucket span in the {anomaly-job} configuration have the
same values.

Your {dfeed} can contain multiple aggregations, but only the ones with names
that match values in the job configuration are fed to the job.

[discrete]
[[aggs-include-jobs]]
== Including aggregations in {anomaly-jobs}
[[aggs-using-date-histogram]]
=== Including aggregations in {anomaly-jobs}

When you create or update an {anomaly-job}, you can include the names of
aggregations, for example:
Expand All @@ -86,8 +90,8 @@ PUT _ml/anomaly_detectors/farequote
----------------------------------
// TEST[skip:setup:farequote_data]

<1> The `airline`, `responsetime`, and `time` fields are aggregations. Only the
aggregated fields defined in the `analysis_config` object are analyzed by the
<1> The `airline`, `responsetime`, and `time` fields are aggregations. Only the
aggregated fields defined in the `analysis_config` object are analyzed by the
{anomaly-job}.

NOTE: When the `summary_count_field_name` property is set to a non-null value,
Expand Down Expand Up @@ -134,25 +138,135 @@ PUT _ml/datafeeds/datafeed-farequote
----------------------------------
// TEST[skip:setup:farequote_job]

<1> The aggregations have names that match the fields that they operate on. The
<1> The aggregations have names that match the fields that they operate on. The
`max` aggregation is named `time` and its field also needs to be `time`.
<2> The `term` aggregation is named `airline` and its field is also named
<2> The `term` aggregation is named `airline` and its field is also named
`airline`.
<3> The `avg` aggregation is named `responsetime` and its field is also named
<3> The `avg` aggregation is named `responsetime` and its field is also named
`responsetime`.

Your {dfeed} can contain multiple aggregations, but only the ones with names
that match values in the job configuration are fed to the job.
TIP: If you are using a `term` aggregation to gather influencer or partition
field information, consider using a `composite` aggregation. It performs
better than a `date_histogram` with a nested `term` aggregation and also includes
all the values of the field instead of the top values per bucket.

[discrete]
[[aggs-using-composite]]
=== Using composite aggregations in {anomaly-jobs}

experimental::[]

For `composite` aggregation support, there must be exactly one `date_histogram` value
source. That value source must not be sorted in descending order. Additional
`composite` aggregation value sources are allowed, such as `terms`.

NOTE: A {dfeed} that uses composite aggregations may not be as performant as datafeeds that use scrolling or
date histogram aggregations. Composite aggregations are optimized
for queries that are either `match_all` or `range` filters. Other types of
queries may cause the `composite` aggregation to be ineffecient.

Here is an example that uses a `composite` aggregation instead of a
`date_histogram`.

Assuming the same job configuration as above.

[source,console]
----------------------------------
PUT _ml/anomaly_detectors/farequote-composite
{
"analysis_config": {
"bucket_span": "60m",
"detectors": [{
"function": "mean",
"field_name": "responsetime",
"by_field_name": "airline"
}],
"summary_count_field_name": "doc_count"
},
"data_description": {
"time_field":"time"
}
}
----------------------------------
// TEST[skip:setup:farequote_data]

This is an example of a datafeed that uses a `composite` aggregation to bucket
the metrics based on time and terms:

[source,console]
----------------------------------
PUT _ml/datafeeds/datafeed-farequote-composite
{
"job_id": "farequote-composite",
"indices": [
"farequote"
],
"aggregations": {
"buckets": {
"composite": {
"size": 1000, <1>
"sources": [
{
"time_bucket": { <2>
"date_histogram": {
"field": "time",
"fixed_interval": "360s",
"time_zone": "UTC"
}
}
},
{
"airline": { <3>
"terms": {
"field": "airline"
}
}
}
]
},
"aggregations": {
"time": { <4>
"max": {
"field": "time"
}
},
"responsetime": { <5>
"avg": {
"field": "responsetime"
}
}
}
}
}
}
----------------------------------
// TEST[skip:setup:farequote_job]

<1> Provide the `size` to the composite agg to control how many resources
are used when aggregating the data. A larger `size` means a faster datafeed but
more cluster resources are used when searching.
<2> The required `date_histogram` composite aggregation source. Make sure it
is named differently than your desired time field.
<3> Instead of using a regular `term` aggregation, adding a composite
aggregation `term` source with the name `airline` works. Note its name
is the same as the field.
<4> The required `max` aggregation whose name is the time field in the
job analysis config.
<5> The `avg` aggregation is named `responsetime` and its field is also named
`responsetime`.

[discrete]
[[aggs-dfeeds]]
== Nested aggregations in {dfeeds}

{dfeeds-cap} support complex nested aggregations. This example uses the
`derivative` pipeline aggregation to find the first order derivative of the
{dfeeds-cap} support complex nested aggregations. This example uses the
`derivative` pipeline aggregation to find the first order derivative of the
counter `system.network.out.bytes` for each value of the field `beat.name`.

NOTE: `derivative` or other pipeline aggregations may not work within `composite`
aggregations. See
{ref}/search-aggregations-bucket-composite-aggregation.html#search-aggregations-bucket-composite-aggregation-pipeline-aggregations[composite aggregations and pipeline aggregations].

[source,js]
----------------------------------
"aggregations": {
Expand Down Expand Up @@ -247,8 +361,9 @@ number of unique entries for the `error` field.
[[aggs-define-dfeeds]]
== Defining aggregations in {dfeeds}

When you define an aggregation in a {dfeed}, it must have the following form:
When you define an aggregation in a {dfeed}, it must have one of the following forms:

When using a `date_histogram` aggregation to bucket by time:
[source,js]
----------------------------------
"aggregations": {
Expand Down Expand Up @@ -282,36 +397,75 @@ When you define an aggregation in a {dfeed}, it must have the following form:
----------------------------------
// NOTCONSOLE

The top level aggregation must be either a
{ref}/search-aggregations-bucket.html[bucket aggregation] containing as single
sub-aggregation that is a `date_histogram` or the top level aggregation is the
required `date_histogram`. There must be exactly one `date_histogram`
aggregation. For more information, see
{ref}/search-aggregations-bucket-datehistogram-aggregation.html[Date histogram aggregation].
When using a `composite` aggregation:

[source,js]
----------------------------------
"aggregations": {
"composite_agg": {
"sources": [
{
"date_histogram_agg": {
"field": "time",
...settings...
}
},
...other valid sources...
],
...composite agg settings...,
"aggregations": {
"timestamp": {
"max": {
"field": "time"
}
},
...other aggregations...
[
[,"aggregations" : {
[<sub_aggregation>]+
} ]
}]
}
}
}
----------------------------------
// NOTCONSOLE

The top level aggregation must be exclusively one of the following:
* A {ref}/search-aggregations-bucket.html[bucket aggregation] containing a single
sub-aggregation that is a `date_histogram`
* A top level aggregation that is a `date_histogram`
* A top level aggregation is a `composite` aggregation.

There must be exactly one `date_histogram`, `composite` aggregation. For more information, see
{ref}/search-aggregations-bucket-datehistogram-aggregation.html[Date histogram aggregation] and
{ref}/search-aggregations-bucket-composite-aggregation.html[Composite aggregation].

NOTE: The `time_zone` parameter in the date histogram aggregation must be set to
`UTC`, which is the default value.

Each histogram bucket has a key, which is the bucket start time. This key cannot
be used for aggregations in {dfeeds}, however, because they need to know the
time of the latest record within a bucket. Otherwise, when you restart a
{dfeed}, it continues from the start time of the histogram bucket and possibly
fetches the same data twice. The max aggregation for the time field is therefore
necessary to provide the time of the latest record within a bucket.
Each histogram or composite bucket has a key, which is the bucket start time.
This key cannot be used for aggregations in {dfeeds}, however, because
they need to know the time of the latest record within a bucket.
Otherwise, when you restart a {dfeed}, it continues from the start time of the
histogram or composite bucket and possibly fetches the same data twice.
The max aggregation for the time field is therefore necessary to provide
the time of the latest record within a bucket.

You can optionally specify a terms aggregation, which creates buckets for
different values of a field.

IMPORTANT: If you use a terms aggregation, by default it returns buckets for
the top ten terms. Thus if the cardinality of the term is greater than 10, not
all terms are analyzed.
all terms are analyzed. In this case, consider using `composite` aggregations
experimental:[Support for composite aggregations inside datafeeds is currently experimental].

You can change this behavior by setting the `size` parameter. To
determine the cardinality of your data, you can run searches such as:

[source,js]
--------------------------------------------------
GET .../_search
GET .../_search
{
"aggs": {
"service_cardinality": {
Expand All @@ -324,10 +478,11 @@ GET .../_search
--------------------------------------------------
// NOTCONSOLE


By default, {es} limits the maximum number of terms returned to 10000. For high
cardinality fields, the query might not run. It might return errors related to
circuit breaking exceptions that indicate that the data is too large. In such
cases, do not use aggregations in your {dfeed}. For more information, see
cases, use `composite` aggregations in your {dfeed}. For more information, see
{ref}/search-aggregations-bucket-terms-aggregation.html[Terms aggregation].

You can also optionally specify multiple sub-aggregations. The sub-aggregations
Expand Down
10 changes: 8 additions & 2 deletions .../elasticsearch/search/aggregations/bucket/composite/DateHistogramValuesSourceBuilder.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -211,15 +211,21 @@ public DateHistogramValuesSourceBuilder fixedInterval(DateHistogramInterval inte
* {@code null} then it means that the interval is expressed as a fixed
* {@link TimeValue} and may be accessed via {@link #getIntervalAsFixed()} ()}. */
public DateHistogramInterval getIntervalAsCalendar() {
return dateHistogramInterval.getAsCalendarInterval();
if (dateHistogramInterval.getIntervalType().equals(DateIntervalWrapper.IntervalTypeEnum.CALENDAR)) {
return dateHistogramInterval.getAsCalendarInterval();
}
return null;
}

/**
* Get the interval as a {@link TimeValue}, regardless of how it was configured. Returns null if
* the interval cannot be parsed as a fixed time.
*/
public DateHistogramInterval getIntervalAsFixed() {
return dateHistogramInterval.getAsFixedInterval();
if (dateHistogramInterval.getIntervalType().equals(DateIntervalWrapper.IntervalTypeEnum.FIXED)) {
return dateHistogramInterval.getAsFixedInterval();
}
return null;
}

/**
Expand Down
Loading

0 comments on commit abb182d

Please sign in to comment.