› › ›

Mapping explosion

Elasticsearch’s search and Kibana’s discover Javascript rendering are dependent on the search’s backing indices total amount of mapped fields, of all mapping depths. When this total amount is too high or is exponentially climbing, we refer to it as experiencing mapping explosion. Field counts going this high are uncommon and usually suggest an upstream document formatting issue as shown in this blog.

Mapping explosion may surface as the following performance symptoms:

CAT nodes reporting high heap or CPU on the main node and/or nodes hosting the indices shards. This may potentially escalate to temporary node unresponsiveness and/or main overwhelm.
CAT tasks reporting long search durations only related to this index or indices, even on simple searches.
CAT tasks reporting long index durations only related to this index or indices. This usually relates to pending tasks reporting that the coordinating node is waiting for all other nodes to confirm they are on mapping update request.
Discover’s Fields for wildcard page-loading API command or Dev Tools page-refreshing Autocomplete API commands are taking a long time (more than 10 seconds) or timing out in the browser’s Developer Tools Network tab. For more information, refer to our walkthrough on troubleshooting Discover.
Discover’s Available fields taking a long time to compile Javascript in the browser’s Developer Tools Performance tab. This may potentially escalate to temporary browser page unresponsiveness.
Kibana’s alerting or security rules may error The content length (X) is bigger than the maximum allowed string (Y) where X is attempted payload and Y is Kibana’s server-maxPayload.
Long Elasticsearch start-up durations.

Prevent or prepare

edit

Mappings cannot be field-reduced once initialized. Elasticsearch indices default to dynamic mappings which doesn’t normally cause problems unless it’s combined with overriding index.mapping.total_fields.limit. The default 1000 limit is considered generous, though overriding to 10000 doesn’t cause noticeable impact depending on use case. However, to give a bad example, overriding to 100000 and this limit being hit by mapping totals would usually have strong performance implications.

If your index mapped fields expect to contain a large, arbitrary set of keys, you may instead consider:

Setting index.mapping.total_fields.ignore_dynamic_beyond_limit to true. Instead of rejecting documents that exceed the field limit, this will ignore dynamic fields once the limit is reached.
Using the flattened data type. Please note, however, that flattened objects is not fully supported in Kibana yet. For example, this could apply to sub-mappings like { host.name , host.os, host.version }. Desired fields are still accessed by runtime fields.
Disable dynamic mappings. This cannot effect current index mapping, but can apply going forward via an index template.

Modifying to the nested data type would not resolve the core issue.

Check for issue

edit

To confirm the field totals of an index to check for mapping explosion:

Check Elasticsearch cluster logs for errors Limit of total fields [X] in index [Y] has been exceeded where X is the value of index.mapping.total_fields.limit and Y is your index. The correlated ingesting source log error would be Limit of total fields [X] has been exceeded while adding new fields [Z] where Z is attempted new fields.
For top-level fields, poll field capabilities for fields=*.
Search the output of get mapping for "type".

If you’re inclined to use the third-party tool JQ, you can process the get mapping mapping.json output.

$ cat mapping.json | jq -c 'to_entries[]| .key as $index| [.value.mappings| to_entries[]|select(.key=="properties") | {(.key):([.value|..|.type?|select(.!=null)]|length)}]| map(to_entries)| flatten| from_entries| ([to_entries[].value]|add)| {index: $index, field_count: .}'

You can use analyze index disk usage to find fields which are never or rarely populated as easy wins.

Complex explosions

edit

Mapping explosions also covers when an individual index field totals are within limits but combined indices fields totals are very high. It’s very common for symptoms to first be noticed on a data view and be traced back to an individual index or a subset of indices via the resolve index API.

However, though less common, it is possible to only experience mapping explosions on the combination of backing indices. For example, if a data stream's backing indices are all at field total limit but each contain unique fields from one another.

This situation most easily surfaces by adding a data view and checking its Fields tab for its total fields count. This statistic does tells you overall fields and not only where index:true, but serves as a good baseline.

If your issue only surfaces via a data view, you may consider this menu’s Field filters if you’re not using multi-fields. Alternatively, you may consider a more targeted index pattern or using a negative pattern to filter-out problematic indices. For example, if logs-* has too high a field count because of problematic backing indices logs-lotsOfFields-*, then you could update to either logs-*,-logs-lotsOfFields-* or logs-iMeantThisAnyway-*.

Resolve

edit

Mapping explosion is not easily resolved, so it is better prevented via the above. Encountering it usually indicates unexpected upstream data changes or planning failures. If encountered, we recommend reviewing your data architecture. The following options are additional to the ones discussed earlier on this page; they should be applied as best use-case applicable:

Disable dynamic mappings.
Reindex into an index with a corrected mapping, either via index template or explicitly set.
If index is unneeded and/or historical, consider deleting.
Export and re-import data into a mapping-corrected index after pruning problematic fields via Logstash.

Splitting index would not resolve the core issue.

« Task queue backlog Hot spotting »