[nnshah1]

Overview

An advisory selection (SchedulerQueue::select_without_admission) no longer goes through the per-model admission actor. It runs on the caller's task, with the same selection code and the same state.

Stack (DEP #17638): #17597 → #17598 → #17665 → #17669 → this PR. This is the only PR in the stack that changes lib/kv-router.

Why: the fleet benchmark of the Global Frontend's core path (inbox/global-view-relay/gf-max-2026-11-03) showed that the remaining cost is the hop through the admission actor. Round-robin cost about 15 µs more per request than the current path. At saturation, selection took milliseconds, because each advisory pick waited behind the actor's other commands. Any advisory caller gets this fix, including the local router.

Key Changes

• SelectionView holds what one selection reads: worker configurations, the selector, the availability and overload providers, the selection trace and observer, the block size and the full-rank counts. The actor and the inline path both build one, so there is one copy of the selection logic.
• Shared helpers: project_worker_loads (load projection and reject thresholds) and advisory_response.
• Selector lock: the selector moves behind Arc<parking_lot::Mutex<_>>, shared by the actor and the queue. The lock keeps picks and runtime policy updates serialized, as on the actor. No trait bound changes: pickers stay Send and do not need Sync.
• Inline path: each call uses its own scratch buffers for loads and full ranks.
• Semantics: an advisory selection reads the same shared state as before. It no longer waits behind actor commands that the caller did not wait for (for example, an unacknowledged cleanup). Acknowledged bookings and releases are already applied, so the caller sees them. Cleanup still runs: it has its own wake-up, and every booking is an actor command.
• One yield before the pick: select_without_admission yields once (tokio::task::yield_now) before the inline pick, as the round trip to the actor did. Without it the caller's task runs on without a break, and the runtime loses the chance to batch other connections' I/O. Nothing is picked before the yield, so a caller dropped there leaves nothing behind.
• Test reference: the actor path stays, test-only, as select_without_admission_on_actor.
• Global Router: the abandoned-select test now accepts a select that answers on its first poll. A new test runs 64 least-in-flight picks on 4 threads: they split evenly and leave nothing pending.

Validation

• cargo test --release -p dynamo-kv-router -p dynamo-global-router -p dynamo-global-frontend -p dynamo-global-common: all pass (kv-router 1427, global-router 79, global-frontend 12 + 9 + 16, discovery peers 2, global-common 10).
• New kv-router tests: the inline path picks what the actor picks from the same state; 256 concurrent advisory selections next to bookings all succeed.
• The parallel least-in-flight test passed 20 times in a row.
• cargo fmt, cargo clippy --all-targets -D warnings: clean. cargo check -p dynamo-llm: builds.
• Fleet (4 regions, 8 Global Frontends at 0.5 CPU, stub pools), first pass, before the yield: least-in-flight on the core went from 43,173 req/s at 0.093 ms CPU per request (R2) to 37,634 at 0.106 ms with this change, while selection time fell from 1.9 ms to 31 µs. A local profile showed the cause: no lock contention (the lock's slow path is 0.01% of samples); the request no longer yields, so tokio batches less I/O and the whole forwarding path costs more. Adding the yield brought the local cost back to R2's level (least-in-flight, 64 connections: R2 50.9, without yield 66.2, with yield 54.8 µs CPU per request). A fleet rerun with the yield is next.
• Local benchmark (advisory_selection_cost, #[ignore]; 4 workers, 4 runtime threads, 20,000 selections):

Callers	Actor p50 / p99	Inline p50 / p99	Calls/s actor → inline
1	0.7 / 2.8 µs	0.3 / 0.3 µs	0.75M → 2.3M
16	17.3 / 29.0 µs	0.7 / 4.8 µs	0.89M → 3.0M
256	287 / 352 µs	0.6 / 4.1 µs	0.88M → 3.2M

Issues and Ongoing Work

• Part of DEP DEP: Global Router selects pools through the embedded router core #17638.
• Draft: the fleet rerun with the yield is not done yet.
• Open question: if a selector's pick gets expensive, the lock serializes picks across callers, as the actor did. A read path without the lock would need Sync pickers.

Pull Request Checklist

• I used Dynamo's /pr-checklist skill.

🤖 Generated with Claude Code

[devin-ai-integration]

✅ Dynamo PR CI passed — run 41711101971 (attempt 1) on 3f1a6e1194

Gate checks: ✅ backend-status-check · ✅ deploy-status-check · ✅ dynamo-status-check

_{Posted automatically by Devin for run 41711101971. Updated on every full-CI run of this PR.}