elasticsearch/docs/reference/ml/delayed-data-detection.asciidoc

[role="xpack"]
[[ml-delayed-data-detection]]
=== Handling delayed data

Delayed data are documents that are indexed late. That is to say, it is data
related to a time that the {dfeed} has already processed.

When you create a datafeed, you can specify a {ref}/ml-datafeed-resource.html[`query_delay`] setting.
This setting enables the datafeed to wait for some time past real-time, which means any "late" data in this period
is fully indexed before the datafeed tries to gather it. However, if the setting is set too low, the datafeed may query
for data before it has been indexed and consequently miss that document. Conversely, if it is set too high,
analysis drifts farther away from real-time. The balance that is struck depends upon each use case and
the environmental factors of the cluster.

==== Why worry about delayed data?

This is a particularly prescient question. If data are delayed randomly (and consequently missing from analysis),
the results of certain types of functions are not really affected. It all comes out ok in the end 
as the delayed data is distributed randomly. An example would be a `mean` metric for a field in a large collection of data.
In this case, checking for delayed data may not provide much benefit. If data are consistently delayed, however, jobs with a `low_count` function may
provide false positives. In this situation, it would be useful to see if data
comes in after an anomaly is recorded so that you can determine a next course of action.

==== How do we detect delayed data?

In addition to the `query_delay` field, there is a
{ref}/ml-datafeed-resource.html#ml-datafeed-delayed-data-check-config[delayed data check config], which enables you to
configure the datafeed to look in the past for delayed data. Every 15 minutes or every `check_window`,
whichever is smaller, the datafeed triggers a document search over the configured indices. This search looks over a
time span with a length of `check_window` ending with the latest finalized bucket. That time span is partitioned into buckets,
whose length equals the bucket span of the associated job. The `doc_count` of those buckets are then compared with the
job's finalized analysis buckets to see whether any data has arrived since the analysis. If there is indeed missing data
due to their ingest delay, the end user is notified.

==== What to do about delayed data?

The most common course of action is to simply to do nothing. For many functions and situations ignoring the data is
acceptable. However, if the amount of delayed data is too great or the situation calls for it, the next course
of action to consider is to increase the `query_delay` of the datafeed. This increased delay allows more time for data to be
indexed. If you have real-time constraints, however, an increased delay might not be desirable.
In which case, you would have to {ref}/tune-for-indexing-speed.html[tune for better indexing speed.]