General recommendations

Elasticsearch is designed as a search engine, which makes it very good at getting back the top documents that match a query. However, it is not as good for workloads that fall into the database domain, such as retrieving all documents that match a particular query. If you need to do this, make sure to use the Scroll API.

Given that the default http.max_context_length is set to 100MB, Elasticsearch will refuse to index any document that is larger than that. You might decide to increase that particular setting, but Lucene still has a limit of about 2GB.

Even without considering hard limits, large documents are usually not practical. Large documents put more stress on network, memory usage and disk, even for search requests that do not request the _source since Elasticsearch needs to fetch the _id of the document in all cases, and the cost of getting this field is bigger for large documents due to how the filesystem cache works. Indexing this document can use an amount of memory that is a multiplier of the original size of the document. Proximity search (phrase queries for instance) and highlighting also become more expensive since their cost directly depends on the size of the original document.

It is sometimes useful to reconsider what the unit of information should be. For instance, the fact you want to make books searchable doesn’t necesarily mean that a document should consist of a whole book. It might be a better idea to use chapters or even paragraphs as documents, and then have a property in these documents that identifies which book they belong to. This does not only avoid the issues with large documents, it also makes the search experience better. For instance if a user searches for two words foo and bar, a match across different chapters is probably very poor, while a match within the same paragraph is likely good.

The data-structures behind Lucene, which Elasticsearch relies on in order to index and store data, work best with dense data, ie. when all documents have the same fields. This is especially true for fields that have norms enabled (which is the case for text fields by default) or doc values enabled (which is the case for numerics, date, ip and keyword by default).

The reason is that Lucene internally identifies documents with so-called doc ids, which are integers between 0 and the total number of documents in the index. These doc ids are used for communication between the internal APIs of Lucene: for instance searching on a term with a match query produces an iterator of doc ids, and these doc ids are then used to retrieve the value of the norm in order to compute a score for these documents. The way this norm lookup is implemented currently is by reserving one byte for each document. The norm value for a given doc id can then be retrieved by reading the byte at index doc_id. While this is very efficient and helps Lucene quickly have access to the norm values of every document, this has the drawback that documents that do not have a value will also require one byte of storage.

In practice, this means that if an index has M documents, norms will require M bytes of storage per field, even for fields that only appear in a small fraction of the documents of the index. Although slightly more complex with doc values due to the fact that doc values have multiple ways that they can be encoded depending on the type of field and on the actual data that the field stores, the problem is very similar. In case you wonder: fielddata, which was used in Elasticsearch pre-2.0 before being replaced with doc values, also suffered from this issue, except that the impact was only on the memory footprint since fielddata was not explicitly materialized on disk.

Note that even though the most notable impact of sparsity is on storage requirements, it also has an impact on indexing speed and search speed since these bytes for documents that do not have a field still need to be written at index time and skipped over at search time.

It is totally fine to have a minority of sparse fields in an index. But beware that if sparsity becomes the rule rather than the exception, then the index will not be as efficient as it could be.

This section mostly focused on norms and doc values because those are the two features that are most affected by sparsity. Sparsity also affect the efficiency of the inverted index (used to index text/keyword fields) and dimensional points (used to index geo_point and numerics) but to a lesser extent.

Here are some recommendations that can help avoid sparsity:

You should avoid putting documents that have totally different structures into the same index in order to avoid sparsity. It is often better to put these documents into different indices, you could also consider giving fewer shards to these smaller indices since they will contain fewer documents overall.

Note that this advice does not apply in the case that you need to use parent/child relations between your documents since this feature is only supported on documents that live in the same index.

Even if you really need to put different kinds of documents in the same index, maybe there are opportunities to reduce sparsity. For instance if all documents in the index have a timestamp field but some call it timestamp and others call it creation_date, it would help to rename it so that all documents have the same field name for the same data.

Types might sound like a good way to store multiple tenants in a single index. They are not: given that types store everything in a single index, having multiple types that have different fields in a single index will also cause problems due to sparsity as described above. If your types do not have very similar mappings, you might want to consider moving them to a dedicated index.

If none of the above recommendations apply in your case, you might want to check whether you actually need norms and doc_values on your sparse fields. norms can be disabled if producing scores is not necessary on a field, this is typically true for fields that are only used for filtering. doc_values can be disabled on fields that are neither used for sorting nor for aggregations. Beware that this decision should not be made lightly since these parameters cannot be changed on a live index, so you would have to reindex if you realize that you need norms or doc_values.