Pathway

Pathway builds a Python framework for data pipelines that handle static and streaming input through the same code, using a Rust engine and incremental computation so a running pipeline updates its outputs as new data arrives rather than being re-run. It is documented for streaming ETL and for retrieval-augmented generation where the index should follow the source, and the vendor ships connectors for files, queues and databases plus tooling for language-model pipelines. That framework is still sold and documented.

The company behind it, however, now leads its public site with something else entirely: it presents itself as a frontier research laboratory working on a post-transformer architecture, with its own model family and benchmark claims about reasoning and cost. That is a real change in what a visitor to the main domain encounters, and anyone evaluating the framework should be aware that the corporate home page no longer describes it. Beyond that, the framework suits teams whose problem is staleness — a retrieval index that must reflect documents as they change — and suits less well a nightly batch job that nothing needs to see before morning, where a plain loader is less machinery. It also suits less well teams without the appetite to operate a long-running pipeline, since that is an operational commitment rather than a script.

Not verified: no pipeline was built or run, so the incremental computation, the ordering guarantees and the claim of exactly-once processing were not tested. The benchmark figures, the cost per task and the model results on the research landing page are the vendor's own results on its own page, and the page itself notes some of those results await contamination checks, independent validation and leaderboard review. The star count shown on the site was read off the page rather than from the hosting service's repository API, the licensing guide was not read, and the pricing page returned an error when fetched, so no price or edition information was obtained.