llamaIndex - 💡(How to fix) Fix [Bug]: ElasticsearchStore sync methods use run_until_complete — breaks in ASGI with pooled clients [1 participants]

Bug Description

ElasticsearchStore only accepts AsyncElasticsearch and bridges all sync methods via asyncio.get_event_loop().run_until_complete(). This causes two problems in ASGI environments (Gunicorn + Uvicorn):

1. Can't pool the async client across requests. Pooling avoids per-request aiohttp.ClientSession creation (which causes memory issues — see https://github.com/run-llama/workflows-py/issues/485). But a pooled AsyncElasticsearch client is bound to one event loop. When a sync method calls run_until_complete() from a different thread, it fails:

RuntimeError: Task <Task pending name='Task-3958' coro=<AsyncVectorStore.close()>>
got Future <Task pending> attached to a different loop

2. close() fails in ASGI. ElasticsearchStore.close() calls run_until_complete(self._store.close()) (line 286). In ASGI, this is called from a sync thread (e.g. VectorStoreService.__exit__() via sync_to_async), hitting the same "different loop" error.

Both problems stem from the same design: there is no sync code path. Every sync method delegates to async via run_until_complete, requiring nest_asyncio.apply() (line 225) and tying the store to one event loop.

OpensearchVectorStore doesn't have this problem

OpensearchVectorClient in the same repo holds separate sync and async clients:

# llama_index/vector_stores/opensearch/base.py:80-81
os_client: Optional[OSClient] = None,          # sync
os_async_client: Optional[OSClient] = None,    # async

Sync methods use the sync client directly:

# line 844-890 (OpensearchVectorClient.query)
res = self._os_client.search(index=self._index, body=search_query, params=params)

Async methods use the async client:

# line 892-938 (OpensearchVectorClient.aquery)
res = await self._os_async_client.search(index=self._index, body=search_query, params=params)

No run_until_complete, no nest_asyncio, no event loop conflicts.

Suggestion

Apply the same sync/async separation to ElasticsearchStore:

• Accept both Elasticsearch (sync) and AsyncElasticsearch (async) clients
• query(), add(), delete(), close() → use sync client directly
• aquery(), async_add(), adelete() → use async client directly
• Remove nest_asyncio.apply() and run_until_complete bridges

We ended up writing our own store with this pattern as a workaround. Happy to contribute a PR if there's interest.

Version

llama-index-vector-stores-elasticsearch 0.5.1

Steps to Reproduce

1. Deploy an ASGI application (Gunicorn + Uvicorn)
2. Pool AsyncElasticsearch clients across requests (to avoid per-request aiohttp session creation)
3. Call ElasticsearchStore.close() or any sync method from a sync thread

Relevant Logs/Tracebacks

File "llama_index/vector_stores/elasticsearch/base.py", line 286, in close
    return asyncio.get_event_loop().run_until_complete(self._store.close())
File "elasticsearch/helpers/vectorstore/_async/vectorstore.py", line 102, in close
    return await self.client.close()
File "aiohttp/client.py", line 1349, in close
    await self._connector.close()
RuntimeError: Task got Future attached to a different loop