Skip to content

feat: optimize router config guidelines for agentic workloads #1958

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Madhumasa84 wants to merge 1 commit into
base: main
Choose a base branch
from
+32 −0
Open

feat: optimize router config guidelines for agentic workloads #1958

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 17 additions & 0 deletions pkg/epp/framework/plugins/scheduling/scorer/prefix/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,6 +31,23 @@ The attribute is typically produced by the approximate prefix cache data produce

This plugin does not define any plugin-specific parameters.

### Agentic Workload Example

For agentic workloads, sessions typically consist of long-running, multi-turn interactions where the system prompt is shared but session-specific context diverges rapidly.

To optimize routing for these workloads, configure both the `prefix-cache-scorer` (with length awareness enabled) and the `session-affinity-scorer`. Length awareness allows shorter sessions to migrate freely when pods are overloaded, while session affinity keeps the rapidly diverging chat history sticky to the original pod:

```yaml
- type: prefix-cache-scorer
parameters:
matchLengthWeight: 0.8
matchLengthScaleTokens: 16000

- type: session-affinity-scorer
parameters:
headerName: x-session-token
```

## Operational notes

- The scorer itself does not hash prompts or maintain cache state.
Expand Down
15 changes: 15 additions & 0 deletions pkg/epp/framework/plugins/scheduling/scorer/sessionaffinity/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,6 +42,21 @@ To support session affinity with PD disaggregation, configure two separate insta

The decode instance uses the default behavior (writing the decode pod to `x-session-token`). The prefill instance uses `profileName: prefill` to look up the prefill pod from the scheduling results and write it to `x-session-token-prefill`. This ensures that subsequent requests in the same session target both the same prefill pod and the same decode pod.

### Agentic Workload Example

For agentic workloads (long-running, multi-turn interactions), it is highly recommended to pair this scorer with the `prefix-cache-scorer` (with length awareness enabled) to ensure rapidly diverging chat history remains sticky to the original pod, while allowing shorter new sessions to migrate to less loaded pods:

```yaml
- type: prefix-cache-scorer
parameters:
matchLengthWeight: 0.8
matchLengthScaleTokens: 16000

- type: session-affinity-scorer
parameters:
headerName: x-session-token
```

## Relationship to the session affinity filter

The [session affinity filter](../../filter/sessionaffinity/README.md) (`session-affinity-filter`) provides the same affinity behavior as a hard constraint and writes the same response header. Configuring both alongside the scorer is unnecessary and can be misleading; see [Relationship to the session affinity scorer](../../filter/sessionaffinity/README.md#relationship-to-the-session-affinity-scorer) for details. Use the scorer for a soft preference that can be outweighed by other scorers, or the filter for a hard pin.