OpenSearch Clients are available in multiple programming languages. The biggest challenge with this is keeping the clients up to date with the latest changes in OpenSearch. To solve this problem, we're automating the process of generating clients for OpenSearch using the OpenAPI specification. While OpenAPI comes with many well established off-the-shelf client generators for most languages, the OpenSearch APIs come with a lot of quirkiness that makes it near impossible to use these off-the-shelf generators. For this reason, we've opted to write our own client generators that is specifically tailored to the OpenSearch APIs. This document will walk you through the process of generating clients from the published OpenSearch OpenAPI spec, more specifically client API methods.
The OpenSearch clients, though written in different languages for different frameworks, share one distinct characteristic: API Operations are grouped into actions and namespaces. Operations serving the same functionality are often grouped together into a single API Action, which is represented in the client as an API method.
The search action, for example, is consisted of 4 operations:
GET /_searchPOST /_searchGET /{index}/_searchPOST /{index}/_search
The indices.get_field_mapping, is consisted of 2 operations:
GET /_mapping/field/{fields}GET /{index}/_mapping/field/{fields}
In a client, the search operations are grouped in to a single API method, client.search(...), and the indices.get_field_mapping operations are grouped into a single API method, get_field_mapping of the indices namespace, client.indices.get_field_mapping(...)
In the published OpenAPI spec, this grouping is denoted by x-operation-group vendor extension in every operation definition. The value of this extension is the name of the API action (like search or indices.get_field_mapping). Operations with the same x-operation-group value are guaranteed to have the same query string parameters, response body, and request body (for PUT/POST/DELETE operations). Common path parameters are also guaranteed to be the same. The only differences between operations are the HTTP method and the path. With that in mind, below are rules on how to combine operations of different HTTP methods and path compositions.
- If an operation is marked with
x-ignorable: true, then ignore the operation. Such an operation has been deprecated and has been superseded by a newer one. As far as the clients are concerned, ignorable operations do not exist. - If two operations have identical HTTP methods, but different paths: use the path that best matches the path parameters provided.
- If two operations have identical path, but different HTTP methods:
- GET/POST: if the request body is provided then use POST, otherwise use GET
- PUT/POST: Either works, but PUT is preferred when an optional path parameter is provided.
The pseudocode that combines the search operations into a single API method is as follows:
def search(self, index=None, body=None):
if index is None:
path = "/_search"
else:
path = f"/{index}/_search"
if body is None:
method = "GET"
else:
method = "POST"
return self.perform_request(method, path, body=body)You will also encounter x-overloaded-param: metric for the node_id path parameter of the GET /_nodes/{node_id} operation in nodes.info action. This is a special case where the path parameter is overloaded to accept either a node ID or a metric name. When the user evokes the client.nodes.info method with either metric or node_id (but not both), the method will use the GET /_nodes/{node_id} operation. When evoked with both metric and node_id, it will use the GET /_nodes/{node_id}/{metric} operation.
Some operations accept a bulk of data in the request body. For example, the bulk action accepts a bulk of index, update, and delete operations on multiple documents. Unlike other operations where the request body is a JSON object, the request body for bulk operations is an NDJSON (i.e a Newline-delimited JSON).
In the spec, these request bodies have the content of application/x-ndjson and their schemas are of type array. In this situation, the client must serialize the request body accordingly, and set the Content-Type header to application/x-ndjson.
As of right now, most clients only validate whether required parameters are present. The clients do not validate the values of parameters against the enum values or regex patterns. This is to reduce performance overhead for the clients as the validation is already done on the server. However, the list of enum values and regex patterns are often written into the parameter description.
Some clients also check for the validity of query string parameter names to guard the users from typos. If you decide to implement this feature, make sure that it's performant. Scripting languages like Python and Ruby require the code to be loaded into memory at runtime, and constructs used for this feature can be expensive to load, as far as microservices are concerned.
All operations in the spec contain a set of parameters that are common across all operations. These parameters are denoted with x-global: true vendor extension. The generated clients should find a way to DRY these parameters in type definitions and method documentation.
Parameters can have default values either through schema or the x-default vendor extension. When both are present, x-default will take precedence.
OpenAPI does not allow duplicate paths, which makes it challenging to express certain OpenSearch APIs, such as /_nodes/{node_id} and /_nodes/{metric}. In this example the server expects either a node_id or a metric. The API specification handles this by defining a /_nodes/{node_id_or_metric} path, and decorating the node_id_or_metric type with a title field.
name: node_id_or_metric
schema:
anyOf:
- title: node_id
$ref: '../schemas/_common.yaml#/components/schemas/NodeIds'
- title: metric
type: array
items:
$ref: '../schemas/nodes.info.yaml#/components/schemas/Metric'See #416 for a complete example.
Clients should generate multiple methods or a method that accepts multiple parameters in this case.