Many consumers of the field caps API need to do some post-processing of the
results before they can use them; for instance, Kibana would like to exclude
multifields from certain field selections, or would like to display only geo_point
fields in Maps. ML and QL consumers exclude nested fields in certain
circumstances. This post-processing is possible at the moment, but can be
hacky; and in all cases it involves sending the whole (possibly very large) field
caps response over the wire and then whittling it down in the client. It is also not
guaranteed to be accurate - runtime fields may be incorrectly classified as multifields,
for example.
This commit pushes filtering into elasticsearch itself, reducing the amount of data
that needs to be transported and ensuring better accuracy. The field caps API gets
two new parameters:
* filters - a comma-delimited list that may contain any combination of: `+metadata`,
`-metadata`, `-nested`, `-parent`, `-multifield`
* types - a comma-delimited list of field types; only fields that have a type in this set
will be returned
The API will make best-effort attempts to apply the filters post-hoc to responses from
older nodes, so this should still work in a mixed-cluster or cross-cluster situation.
Fixes#82966, #72174
8bc46ad959