pangwangshu commented on issue #40837: URL: https://github.com/apache/beam/issues/40837#issuecomment-6491720673
@LibenHailu I've opened a draft PR with a working prototype: #40881. It follows the [design doc](https://docs.google.com/document/d/1cC5wezcXugmyHDbt-xuP6zrlKia337dJObWNKnZZJvM/edit), which is still under discussion on dev@. The writer uses a thin `/document/v1` client on `requests` (no new dependency). Unit tests and a testcontainers-based IT both pass locally. Here is what is still open and where your help would be most useful. **Waiting on dev@ (we shouldn't settle these on our own)** 1. Thin `/document/v1` client vs. pyvespa as an optional extra. 2. Which test suite the container-based IT runs in. 3. Fail the bundle vs. emit failed documents to a dead-letter output. **Good places to jump in** - **Throughput benchmark** (feeds into 1): compare the prototype with pyvespa's feed client, using the IT container or Vespa Cloud. Numbers would make the dev@ decision much easier. - **429 handling**: 429 retries are currently unbounded. Should there be a total time limit per document, so an overloaded cluster can't hang a batch job? - **Session per thread**: one `requests.Session` is shared across worker threads. Is it worth switching to one session per thread? - **Server-side timeout**: should we also send Vespa's `timeout` query parameter alongside `timeout_s`? - **`update` semantics**: partial updates keep fields they don't write, while `put` clears them. Should we document this or change it? - **Docs PR**: a section in the RAG ingestion docs (or an example notebook) plus the CHANGES.md entry, once the API is settled. - **Follow-up**: a Vespa enrichment handler for retrieval (nearestNeighbor and hybrid YQL), as a separate proposal. The PR description has the full list. Pick whatever interests you and comment on the PR or here, so we don't duplicate work. Happy to review anything you push. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
