Hi Rahul, I really like your proposal. Some time ago, I attempted to identify clients in Polaris by analyzing their User-Agent and X-Client-Version headers, but I quickly discovered that this approach was far more complicated and unreliable than I had expected [1]. I think the standardization that you are proposing here will finally make it doable.
Thanks, Alex [1]: https://github.com/apache/polaris/pull/4236 On Tue, Sep 1, 2026 at 6:26 PM rahul mahadev <[email protected]> wrote: > > Hi all, > > I'd like to propose a recommended User-Agent format for REST catalog > clients, so a client identifies itself to a catalog in a consistent, > parseable way. > > Today it's inconsistent: > - iceberg-java sends no User-Agent by default (it's opt-in via config). > - pyiceberg sends "PyIceberg/<version>", iceberg-rust "iceberg-rs/<version>", > iceberg-go "GoIceberg/<version>" basically all different naming. > - X-Client-Version carries the library version in Java/Python but the REST > spec version in Rust/Go. > - None of this is described in the spec. > > That makes it hard for a catalog operator to tell what's actually calling, > which matters for observability, debugging, and support. > > Proposal: use the standard User-Agent header (RFC 7231) i.e. whitespace- > separated product/version tokens, most specific first (engine -> > integration -> Iceberg library -> runtime), with an optional parenthesized > comment for extra context. Example: > > Spark/4.0.0 iceberg-spark/1.9.0 iceberg-java/1.9.0 (scala/2.13.16) > > The Iceberg library token is the one piece every client can supply, so it > should always be present. The header is optional and informational only: > servers must not reject on it, must not use it for auth or trust decisions, > and it is not a capability negotiation mechanism. > > Draft PR with the spec change: > https://github.com/apache/iceberg/pull/17727 > > If the format looks right, the follow-ups would be conforming the client > libraries (java / python / rust / go) to emit it. > > A few open questions: > 1. Should we also reconcile X-Client-Version (library vs spec version > across clients), or leave it and treat User-Agent as the canonical > identity going forward? > 2. Naming: keep the existing library names (PyIceberg, iceberg-rs, > GoIceberg) or converge on one scheme (iceberg-<lang>)? > 3. Is a separate environment token (e.g. prod/staging) useful, or does > that belong in the comment? > 4. Any other fields we want to capture ? > > Feedback welcome. > > Thanks, > Rahul Mahadev (github: rahulsmahadev)
