Skip to content

Switchyard lets LLM applications route traffic across models and providers while preserving native OpenAI and Anthropic API compatibility - enabling flexible model selection, benchmarking, and cost/performance optimization.

Rust

feat(relay): add schemas to routing marks (#604)

Alex Fournierd4247a2cd6d8
diff --git a/crates/switchyard-nemo-relay-plugin/README.md b/crates/switchyard-nemo-relay-plugin/README.md
index b87f25b16..e017f5893 100644
--- a/crates/switchyard-nemo-relay-plugin/README.md
+++ b/crates/switchyard-nemo-relay-plugin/README.md
@@ -187,6 +187,19 @@ attributes. `target_model` comes from the configured Switchyard target set,
rather than arbitrary caller input, keeping the metric cardinality bounded by
the deployment.
+Every non-metric mark sets `data_schema.name` to the mark name and
+`data_schema.version` to `1`. Consumers should tolerate unknown fields and
+values. Removing or renaming fields, changing their type, or changing their
+meaning requires a new schema version.
+
+| Mark | Data fields |
+| --- | --- |
+| `switchyard.routing.requested` | `algorithm` |
+| `switchyard.routing.llm_call` | `call_index`, `selected_model`, `call_role`, `outcome`, `latency_ms` |
+| `switchyard.routing.overhead` | `latency_ms` |
+| `switchyard.routing.decision` | `algorithm`, `selected_model` |
+| `switchyard.routing.error` | `failure_kind`; optional `category`, `phase`, `upstream_status`, and `target` |
+
## Failure policy
`switchyard-llm-client` owns provider retry and route-candidate fallback
diff --git a/crates/switchyard-nemo-relay-plugin/src/runtime.rs b/crates/switchyard-nemo-relay-plugin/src/runtime.rs
index 7b5a12fb9..136ec2a88 100644
--- a/crates/switchyard-nemo-relay-plugin/src/runtime.rs
+++ b/crates/switchyard-nemo-relay-plugin/src/runtime.rs
@@ -7,8 +7,8 @@ use std::sync::{Arc, Mutex};
use futures_util::{Stream, StreamExt};
use http::StatusCode;
use nemo_relay_plugin::{
- Json, LlmRequest as RelayRequest, LogSeverity, MetricKind, MetricMeasurement, MetricValueType,
- PluginRuntime,
+ DataSchema, Json, LlmRequest as RelayRequest, LogSeverity, MetricKind, MetricMeasurement,
+ MetricValueType, PluginRuntime,
};
use serde_json::{Map, json};
use switchyard_llm_client::{LlmCallObservation, RunObservation, RunObserver};
@@ -22,6 +22,8 @@ use switchyard_translation::{TranslationEngine, encode_stream_with_extensions};
use crate::config::SwitchyardConfig;
use crate::translation;
+const ROUTING_MARK_SCHEMA_VERSION: &str = "1";
+
#[derive(Debug)]
pub(crate) struct RoutingMark {
pub(crate) name: String,
@@ -30,6 +32,15 @@ pub(crate) struct RoutingMark {
pub(crate) severity: Option<LogSeverity>,
}
+impl RoutingMark {
+ fn data_schema(&self) -> DataSchema {
+ DataSchema {
+ name: self.name.clone(),
+ version: ROUTING_MARK_SCHEMA_VERSION.into(),
+ }
+ }
+}
+
#[derive(Debug)]
pub(crate) struct RoutingMetric {
pub(crate) name: String,
@@ -324,15 +335,18 @@ pub(crate) fn emit_events(runtime: &PluginRuntime, events: Vec<RoutingEvent>) {
pub(crate) fn emit_event(runtime: &PluginRuntime, event: RoutingEvent) {
let result = match event {
- RoutingEvent::Mark(mark) => runtime
- .emit_mark_with_options(
- &mark.name,
- Some(&mark.data),
- Some(&mark.metadata),
- None,
- mark.severity,
- )
- .map_err(|error| ("routing mark", mark.name, error)),
+ RoutingEvent::Mark(mark) => {
+ let data_schema = mark.data_schema();
+ runtime
+ .emit_mark_with_options(
+ &mark.name,
+ Some(&mark.data),
+ Some(&mark.metadata),
+ Some(&data_schema),
+ mark.severity,
+ )
+ .map_err(|error| ("routing mark", mark.name, error))
+ }
RoutingEvent::Metric(metric) => runtime
.emit_metric(&metric.name, metric.measurements, Some(&metric.metadata))
.map_err(|error| ("routing metric", metric.name, error)),
@@ -980,6 +994,8 @@ mod tests {
assert_eq!(mark.data["target"], "weak");
assert_eq!(mark.data["upstream_status"], Json::Null);
assert_eq!(mark.severity, Some(LogSeverity::Error));
+ assert_eq!(mark.data_schema().name, "switchyard.routing.error");
+ assert_eq!(mark.data_schema().version, "1");
assert!(!mark.data.to_string().contains(secret));
}