Query external datasets with ES|QL Data Federation

A dataset is a read source for the standard ES|QL pipeline. You query it with FROM like an index, and every processing command works the same way it does for an index.

FROM my_dataset
		

For a hands-on example, refer to get started with ES|QL Data Federation.

When you query a dataset, Elasticsearch reads data from object storage (such as Amazon S3) rather than from a local index. This means every column and every row that a query touches results in network I/O. The query engine applies several optimizations automatically, and there are things you can do to help it read less data.

Use KEEP or DROP to select only the columns your query needs. For Parquet files, column selection pushes down to the reader so that unrequested columns are never fetched from storage. For CSV and NDJSON, the full row is read but unrequested columns are discarded early.

In practice, this can make a significant difference. A filtered query over a Parquet dataset that selects three columns reads roughly a third of the bytes that the same query reads without column selection.

When a dataset's resource path uses Hive-style partitioning (for example, year=2024/month=3/), the engine detects partition keys automatically and promotes them to queryable columns. A WHERE condition on a partition column evaluates during file discovery, before any data is read. On a two-year monthly-partitioned dataset, WHERE year = 2024 AND month = 3 skips 23 out of 24 partitions at zero I/O cost.

A filter on a column inside the file can skip folders too, when the dataset sets partition_spec. WHERE ts > "2024-03-15T00:00:00Z" skips folders whose time range does not overlap that filter. ::datetime on the ISO string is optional: ES|QL casts a keyword literal to datetime on comparison with a date column. Path keys you do not list in the spec still prune on their own name.

Pruning applies when the partition filter comes before any LIMIT, SORT, or STATS in the query. If one of those commands sits between FROM and the WHERE on a partition column, pruning is silently skipped and every partition is read. Try to put partition filters first.

For details on partition detection modes, refer to partition_detection.

WHERE conditions and LIMIT reduce how much data a query reads, but how far they push down depends on the format. For Parquet files, filters push down into the reader itself: the engine uses row-group statistics and page indexes to skip data that cannot match the filter. Only row groups whose statistics overlap the filter condition are read, and within those row groups, late materialization reads predicate columns first and materializes other columns only for rows that survive the filter.

For CSV and NDJSON, filters do not reach the reader: every row must be read and parsed, but rows that fail the filter are discarded before further processing.

FROM access_logs
| WHERE status_code >= 500
| KEEP @timestamp, status_code, request_path
| LIMIT 100
		

The general query performance advice in optimize ES|QL query performance applies to datasets too. In particular, adding a WHERE, a KEEP, and a LIMIT are the three most effective ways to reduce how much data a query reads from storage.

Elasticsearch caches dataset metadata on each node so that repeated queries against the same dataset don't discover files and read file metadata again each time:

• File listings are reused for a fixed time before Elasticsearch lists the storage again. A file added to or removed from the bucket becomes visible once the listing expires.
• Inferred schemas are invalidated when the underlying file changes.
  • Additionally, a schema is reused for at most 20 minutes by default.
• Footers of columnar files, such as Parquet footers, are kept while queries keep reading them, and dropped 5 minutes after the last read by default. Footer entries are keyed by path and file length rather than by the object's version, so a file overwritten in place at the same byte length shares a cache entry with its predecessor. A write cap bounds how long a stale footer can be served: at the default TTL of 5 minutes, raw footer bytes are force-expired within 30 minutes regardless of how often they are read. Parsed footers can remain longer — up to approximately 60 minutes at the default — because a parsed-cache eviction may reload from raw bytes still within their own cap, starting a new parsed-entry clock.

The data in your files is not cached. Every query reads the rows it needs from object storage, so running the same query again or refreshing a dashboard reads the data again.

To change cache sizes and TTLs, refer to caching settings.

A dataset's resource path can use glob patterns to match many files. These cluster settings bound file discovery:

• esql.external.max_listed_objects (default 1,000,000): the maximum number of objects visited while listing a glob, including keys that do not match the pattern and keys dropped by exclusion. Applied independently to each glob listing. A comma-separated resource of N globs therefore does N listings; a rewrite-empty fallback can list the same glob again. The kept-files cap (esql.external.max_discovered_files) is shared across that list.
• esql.external.max_discovered_files (default 25,000): the maximum number of files a single dataset keeps after listing filters (_file.*).
• esql.external.max_glob_expansion (default 100): the maximum number of concrete paths a brace pattern ({a,b,c}) expands to. Past this cap, the engine falls back to listing the storage instead of failing.

If your dataset exceeds these limits, narrow the resource path or adjust the settings. Refer to cluster settings for details.

Datasets share the same namespace as indices, data streams, aliases, and ES|QL views, so FROM resolves each name independently.

FROM speedtest_data, network_incidents METADATA _class, _name
| KEEP _class, _name, category, severity, avg_d_kbps, avg_lat_ms
| LIMIT 10
		

When sources have different schemas, columns that do not exist in a given source return null for rows from that source. To see which source each row came from, request a metadata column. For details, refer to Use metadata columns.

• Use METADATA _name and METADATA _class. _name returns the dataset name for dataset rows and the index name for index rows. _class returns index or dataset, so a query can tell the two apart without knowing the names in advance. _index returns null for dataset rows, because a dataset is not an index.
• Use METADATA _index, which returns the dataset name for dataset rows. _class and _name are not available.

By default, a wildcard does not match datasets. FROM speedtest_data reads the dataset, while FROM speedtest* resolves to indices, data streams, aliases, and views only. Registering a dataset therefore does not change what an existing wildcard query reads.

To let wildcards discover datasets, enable the wildcards_match_datasets query setting:

SET wildcards_match_datasets = true;
FROM speedtest*
		

You can also send it in the _query request body as "settings": {"wildcards_match_datasets": true}, or change the cluster-wide default with esql.query.settings.wildcards_match_datasets. A value set in the query overrides the request body, which overrides the cluster default.

Metadata columns are available using the METADATA directive:

_index, _id, _version, and _source return null on a dataset. A dataset is not an index, and files carry no document identity, version, or stored source.

For example, this query returns file-level metadata for each matching row:

FROM access_logs METADATA _file.path, _file.name, _file.size
| KEEP _file.path, _file.name, _file.size, status_code
| LIMIT 10
		

If a field named in the METADATA clause matches a file column, the engine-generated value takes precedence and the file column is dropped. A warning identifies the file column that was dropped. To keep both values, rename the file column in the dataset mapping. If the query omits METADATA, the file column remains available as a regular data column.

Search functions can filter dataset rows by evaluating the query against values read from the files. This runtime search does not use an inverted index.

Because there is no inverted index, search functions on a dataset evaluate by scanning values row by row. For large datasets where search is the primary access pattern, consider ingesting the data into Elasticsearch for indexed search performance.

How dataset rows contribute to METADATA _score depends on the version:

• MATCH and MATCH_PHRASE on dataset rows contribute to the relevance score based on the boost option and the query terms matched. Scoring does not use BM25, because a dataset has no index statistics.
• Dataset rows do not contribute to _score.

Runtime MATCH on a dataset requires the query value's type to match the field's type. How it handles options and analyzers depends on the version:

• Text is analyzed with the standard analyzer unless a values analyzer is declared through TO_TEXT's analyzer option. MATCH's own analyzer option applies to the query string only, and defaults to the values analyzer.
• MATCH does not accept function options, and text is analyzed with a fixed standard analyzer.

The following search functions are available for datasets:

The limitations below include operations that require structures available only in an Elasticsearch index, such as the inverted index, doc values, or time series metadata, as well as unsupported data shapes. Unsupported operations fail with a clear error; representation limitations are described in the table.

If a query against a dataset returns unexpected results or errors, check the following common causes.

Unexpected nulls in query results

If you query a dataset and an index together with FROM, columns that do not exist in one source return null for rows from that source. To check which source each row came from, refer to Query across datasets and indices.

Complex Parquet types MAP, nested LIST, and VARIANT also return null, because they are not currently supported.

Slow queries

Add KEEP to select only the columns you need, add a WHERE filter, and add a LIMIT. For Parquet datasets, these push down to the reader and can significantly reduce the amount of data read from storage. Check the number of files your dataset's resource path resolves to. Large file counts increase query planning time.

503 error when creating a data source with credentials

Elasticsearch encrypts credentials before storing them. If the cluster state encryption key is not available, the request returns 503 SERVICE_UNAVAILABLE. Refer to credential encryption for details.

New files not appearing in query results

Elasticsearch caches file listings for each dataset. A file added to or removed from the bucket might not show up until the listing cache expires. Lower the listing cache TTL when new or removed files must be visible sooner. Refer to caching settings.

Columns with unexpected types or missing values

When Elasticsearch infers a dataset's schema from its files, it might infer types differently than you expect. For example, a date column might appear as a keyword if the values do not match the default datetime format. To inspect the inferred field mappings, refer to check field mappings in the quickstart. Declare the schema explicitly to control column types, or adjust the datetime_format setting for CSV and TSV or NDJSON. If some rows have null values for a column that exists in other files, review schema inference and resolution.

Access denied or connection errors

Credential and permission errors appear at query time, not when the data source or dataset is created. For each access denied message and the permission it needs, refer to troubleshoot access errors for Amazon S3, and verify that the region is correct.

Use the following resources to tune queries and dataset behavior:

• To adjust caching TTLs, file-discovery limits, or request concurrency, refer to cluster settings.
• To control column types or rename columns, declare the dataset schema explicitly.
• For general ES|QL tuning advice that also applies to datasets, refer to optimize ES|QL query performance.