Description
Pylon strips client-supplied Dynamo priority headers (x-dynamo-request-priority, x-dynamo-request-strict-priority) before forwarding to the local inference server. Dynamo reads other request headers too, for example worker pinning and data-parallel rank hints, and those still pass through every hop verbatim. A client can set them and steer routing on the inference server.
Widen pylon's strip list to cover the remaining engine headers. Pylon is the last trusted hop, so it should own this; the gateway and Stargate need no Dynamo-specific handling.
Definition of Done
- Pylon strips the remaining Dynamo request headers (worker instance id, data-parallel rank hints, tenant id, their unprefixed aliases, and any future
x-dynamo-request-* names) before forwarding.
- The decision between a wider denylist and a prefix rule is written down next to the list.
- Unit and tunnel tests pin the strip so a refactor cannot silently reintroduce passthrough.
Resources
Description
Pylon strips client-supplied Dynamo priority headers (
x-dynamo-request-priority,x-dynamo-request-strict-priority) before forwarding to the local inference server. Dynamo reads other request headers too, for example worker pinning and data-parallel rank hints, and those still pass through every hop verbatim. A client can set them and steer routing on the inference server.Widen pylon's strip list to cover the remaining engine headers. Pylon is the last trusted hop, so it should own this; the gateway and Stargate need no Dynamo-specific handling.
Definition of Done
x-dynamo-request-*names) before forwarding.Resources
src/libraries/rust/stargate/crates/pylon-lib/src/quic_http_tunnel/backend.rsshould_forward_headerincrates/pylon-lib/src/quic_http_tunnel/core.rs