This is a "hack" to allow usage of the legacy Elasticsearch 1.x/2.x completion suggester in Elasticsearch 5.x. This legacy plugin supports deduplication and is not document based.
This fixes the following Elasticsearch issue by a workaround: elastic/elasticsearch#22912
To restore the old behaviour of the old completion suggester for new indexes created with Elasticsearch 5.x use the following process:
The plugin compiles with Elasticsearch 5.4.1, but should also work with
earlier or later versions. Just edit the POM file, compile, and create
the package: mvn clean install
After that copy the es-legacy-completion-plugin-5.X.Y.jar file and the
plugin-descriptor.properties file from the target directory to a new
subdirectory (e.g., legacy-completion) of Elasticsearch's plugin
folder.
After installataion you can change your ES 2.x mapping to use the new
field type legacy_completion instead of completion (which
unfortunately triggers new behaviour). When you then query the index
using the usual suggestion endpoint, it will return results in the same
way as Elasticsearch 2.x did.
Elasticsearch 5.x still has the old suggester code inside to support Elasticsearch 2.x indexes, but for new indexes it uses the new suggester. The completion suggester endpoint just checks the Lucene Codec that was used to create the suggestion field and then uses the old or new one.
This plugin just adds a new field type that internally enforces the use of the old Lucene Codec. This way the suggester endpoint automatically falls back to old behaviour. It does this by some hacks in the field mapper and (unfortunately), lots of code duplication, but this was kept minimal.
Caution: This is likely to break with Elasticsearch 6!