Standalone Selection Service

Overview

The standalone selection service (python -m dynamo.select_service) exposes the KV router’s worker selection and active-load accounting over HTTP. It does not forward model requests or own response streams. External runtimes such as Ray register their worker catalog, request a selection, contact the selected worker, and report the reservation lifecycle.

The service combines:

• KV overlap indexing from worker ZMQ events.
• KV-aware and load-aware worker selection.
• Explicit or atomic selection and reservation.
• Best-effort active-load synchronization between selector replicas.
• Startup KV index recovery from another selector or standalone indexer.

Build And Launch

Build the Python bindings with the select-service feature:

$
cd lib/bindings/python
$
VIRTUAL_ENV=../../../.venv ../../../.venv/bin/maturin develop --uv --features select-service

Launch the service from the repository root:

$
.venv/bin/python -m dynamo.select_service --port 8092

The service binds to 0.0.0.0 and does not provide authentication. Run it on a trusted internal network or place it behind an appropriate network policy.

Embedded Rust API

Use SelectionServiceBuilder to embed selection without the HTTP server. The resulting SelectionService owns worker registration, KV-event listeners, indexer recovery, replica synchronization, and shutdown. It exposes the same worker, selection, bookkeeping, inspection, peer-membership, and recovery operations as the standalone HTTP service.

SelectionCore::new_local creates an intentionally unsynchronized core for tests and local-only use. Production integrations should use SelectionServiceBuilder so startup recovery, readiness, and background-task lifecycle remain consistent with the standalone service.

The C and Go bindings do not currently expose SelectionService. An EPP integration requires separate FFI lifecycle, error-mapping, worker, and peer APIs. Those bindings should wrap SelectionService rather than construct SelectionCore directly.

CLI

Router scheduling behavior continues to use the standard Dynamo router environment configuration.

Worker Registration

Every selector replica must receive the same worker catalog before it serves selection traffic. Replica traffic never creates workers.

1
POST /workers
2
Content-Type: application/json
3
4
{
5
  "worker_id": 1,
6
  "model_name": "model",
7
  "routing_group": "default",
8
  "endpoint": "http://worker:8000",
9
  "block_size": 16,
10
  "data_parallel_start_rank": 0,
11
  "data_parallel_size": 2,
12
  "kv_events_endpoints": {
13
    "0": "tcp://worker:5557",
14
    "1": "tcp://worker:5558"
15
  },
16
  "replay_endpoint": "tcp://worker:5560"
17
}

POST /workers returns 201. PATCH /workers/{worker_id} updates supplied fields, DELETE /workers/{worker_id} removes the worker, and GET /workers lists catalog state. model_name and routing_group scope all selection, indexer, and load state; both default to "default" when omitted.

GET /health is process liveness. GET /ready returns 200 only after at least one worker is schedulable, otherwise 503 with lifecycle details.

Selection API

POST /select

Select a worker without booking active load:

1
{
2
  "selection_id": "select-123",
3
  "model_name": "model",
4
  "routing_group": "default",
5
  "block_hashes": [11, 12, 13, 14, 15, 16, 17, 18],
6
  "sequence_hashes": [21, 22, 23, 24, 25, 26, 27, 28],
7
  "isl_tokens": 512
8
}

POST /select_and_reserve

Select and atomically book load in the receiving selector process. Supply a globally unique selection_id, or allow the service to generate one:

1
{
2
  "selection_id": "select-123",
3
  "model_name": "model",
4
  "routing_group": "default",
5
  "block_hashes": [11, 12, 13, 14, 15, 16, 17, 18],
6
  "sequence_hashes": [21, 22, 23, 24, 25, 26, 27, 28],
7
  "isl_tokens": 512
8
}

Both endpoints return the same selection shape:

1
{
2
  "selection_id": "select-123",
3
  "model_name": "model",
4
  "routing_group": "default",
5
  "worker_id": 1,
6
  "dp_rank": 0,
7
  "endpoint": "http://worker:8000",
8
  "block_size": 16,
9
  "overlap": {
10
    "longest_matched": 128,
11
    "gpu": 64,
12
    "dp": {"0": 64, "1": 32},
13
    "cpu": 96,
14
    "disk": 128
15
  },
16
  "effective_prefill_tokens": 384
17
}

selection_id is omitted when absent. All overlap values are matched token counts. gpu, cpu, and disk use the cumulative Mooncake tier semantics documented in the standalone indexer’s per-instance tier breakdown. A zero-overlap response includes the selected dp_rank with value 0.

The overlap summary is raw observability. effective_prefill_tokens is the authoritative weighted prefill-load value computed by the same cache-credit formula used for scheduler booking. It is not derived from longest_matched. When a request waits in the scheduler queue, both fields reflect the final overlap inputs after any dequeue-time refresh rather than the enqueue-time snapshot.

The previous public fields cached_tokens and effective_overlap_blocks have been removed. Their values remain internal scheduler inputs.

Ray Select-Then-Reserve Flow

Ray can keep model invocation separate from selector admission:

1. Call POST /select with a selection_id.
2. Send the request to the returned endpoint and dp_rank.
3. Call POST /reservations using the cached selection replay form, or the explicit form with the full worker identity and prompt.
4. Report prefill completion and request completion through the lifecycle API.

Cached selection replay form

A /select that carries a selection_id caches its booking inputs (the chosen worker, the normalized prompt, effective_prefill_tokens, expected_output_tokens, and the prefill-tracking decision) on the selector that served it. A reservation that passes the same selection_id, model_name, and routing_group replays the cached selection, booked under that selection_id, without re-sending the prompt:

1
POST /reservations
2
Content-Type: application/json
3
4
{
5
  "selection_id": "select-123",
6
  "model_name": "model",
7
  "routing_group": "default"
8
}

• Id namespace: selection_id is client-chosen and scoped per (model_name, routing_group); use a distinct id per in-flight select. A new select reusing a pending id replaces it (latest wins), and an explicit booking discards the cached selection for its selection_id.
• Required id: selection_id is always required (the replay key and the booking id); a request without it is rejected.
• Single-use: The first successful booking consumes the entry; a repeat replay returns 404 (no pending selection). Concurrent replays of the same id collide at the scheduler, so only one books.
• Retryable on failure: A booking that fails before landing (worker no longer schedulable, service not ready) leaves the entry in place, so the same call can be retried once the condition clears.
• Bounded window: By default, entries expire after 120 seconds and each selector retains at most 4096 pending selections within a 256 MiB budget, evicting oldest first. All three limits are configurable.
• Replica-local: The cache lives in the selector process that served the /select. With multiple selector replicas, route the reservation to the same replica or use the explicit form.
• Pure replay: The booking uses exactly what select captured; other request fields are ignored. Supplying worker_id switches to the explicit form.

On any miss (expired, already consumed, wrong model or routing group, or a different replica) the call returns 404; fall back to the explicit form.

Explicit form

The self-contained form carries the worker identity and prompt and needs no cached selection; it wins whenever worker_id is present. It discards the cached selection for its selection_id, so a later replay of the same id cannot book stale state:

1
POST /reservations
2
Content-Type: application/json
3
4
{
5
  "selection_id": "request-123",
6
  "model_name": "model",
7
  "routing_group": "default",
8
  "worker_id": 1,
9
  "dp_rank": 0,
10
  "sequence_hashes": [21, 22, 23, 24, 25, 26, 27, 28],
11
  "isl_tokens": 512,
12
  "effective_prefill_tokens": 384
13
}

When supplied, effective_prefill_tokens is authoritative and directly enables prefill-load tracking. It must not exceed the normalized input sequence length. When omitted, existing router configuration controls prefill tracking. The reservation API does not accept or derive accounting from overlap fields.

Reservation Lifecycle

1
POST /reservations/{selection_id}/prefill_complete
2
POST /reservations/{selection_id}/output_block
3
DELETE /reservations/{selection_id}

prefill_complete clears active prefill load. output_block updates only the receiving selector’s local decode-block accounting and accepts an optional decay_fraction in [0.0, 1.0]. DELETE frees the reservation.

NOTE: Output-block updates are intentionally not replica-synchronized. They can occur at high frequency, and broadcasting them would consume disproportionate network bandwidth.

Peer Planes

The selector has two independent peer configurations:

Example:

$
.venv/bin/python -m dynamo.select_service \
>
  --port 8092 \
>
  --indexer-peers http://selector-b:8092 \
>
  --replica-sync-port 9092 \
>
  --replica-sync-peers 'tcp://selector-b:9092'

Configure the reverse peer direction on selector B for bidirectional lifecycle synchronization. GET /dump exposes the selector’s current indexer snapshot in the same recovery format as the standalone indexer.

Replica-sync peers may also be changed without restarting the selector:

1
POST /replica_sync/register_peer
2
Content-Type: application/json
3
4
{"endpoint":"tcp://selector-b:9092"}

The same body is accepted by POST /replica_sync/deregister_peer. GET /replica_sync/peers returns the sorted configured endpoints. Dynamic membership is in-memory; after restart, only peers supplied through --replica-sync-peers are restored. These routes only manage live ZMQ replica-sync peers. They do not alter the HTTP indexer-recovery peers.

Consistency Invariants

• Replica synchronization is bounded and best-effort. Delays, reordering, dropped events, and temporary active-load divergence are accepted.
• There is no sequencing, acknowledgement, replay, backpressure, or resynchronization for replica lifecycle events.
• Unknown worker, model, routing-group, DP-rank, and block-size events are dropped. Register the same worker catalog on every selector before routing traffic.
• The v1 replica envelope uses routing_group and is incompatible with binaries that send tenant_id. Drain active reservations and lifecycle traffic, upgrade all connected selectors together, re-register worker catalogs, and then resume traffic. Active advisory state is not migrated.
• Admission, prefill-complete, and free are synchronized. Output-block growth remains local to avoid excessive network bandwidth.
• Startup recovery waits for recovered events to be submitted to the indexer, not for complete processing. Early selections may temporarily miss recovered KV state.
• /select followed by /reservations provides eventual, not atomic, cross-replica admission, and the pending-selection cache behind the minimal reservation form is local to the selector that served the /select. Use /select_and_reserve for atomic local booking.
• Reservation IDs must be globally unique. Duplicate bookings for the same ID conflict (409); no idempotency ledger is added. An explicit booking that carries a selection_id also discards that cached selection.

Inspection APIs

• GET /loads returns active-load snapshots, optionally filtered by model_name and routing_group.
• POST /potential_loads estimates worker load for a prompt without selection.
• POST /overlap_scores returns per-worker/per-rank tiered overlap rows.
• GET /dump returns the compatible indexer recovery snapshot.