Description
Bug Description
The ES Vector Store does lazy initialization of the backing index.
As the vector store is async capable, there can be multiple concurrent requests to the vector store.
Currently create_index_if_not_exists is used to create the index if it doesn't exist.
If two requests are made to add nodes and the index has not been initialized (or has recently been cleared), there is a race condition where both requests will branch off to create_index_if_not_exists, as this operation is several steps, both will run through all of the steps and the slower one will fail at the create index step:
await self.client.indices.create(
index=self.index, mappings=mappings, settings=settings
)
As the index already exists.
We should either make sure that the environment setup is idempotent or synchronous so that requests to add nodes to a new ES vector store, or a recently cleared ES vector store, do not result in failures that drop documents.
Version
12.43
Steps to Reproduce
Create an ES vector store and run two parallel requests to add nodes.