feat: implement async scheduling admission control

architecture/dataset-builders.md

@@ -35,7 +35,7 @@ Preparation (`_prepare_async_run`):
 4. Constructs `CompletionTracker`, `RowGroupBufferManager`, `AsyncTaskScheduler`
 5. Hooks `ProcessorRunner` for pre-batch and post-batch stages
 
-`AsyncTaskScheduler` runs on a dedicated async loop with frontier-driven dispatch, semaphore-based capacity limits, salvage rounds for failed tasks, and order-dependent locks for columns that must execute sequentially. Ready frontier tasks are admitted through a virtual-time fair queue so one hot column or model-backed generator cannot consume the whole submission window before peer work gets a turn.
+`AsyncTaskScheduler` runs on a dedicated async loop with frontier-driven dispatch, task-admission leases, salvage rounds for failed tasks, and order-dependent locks for columns that must execute sequentially. Ready frontier tasks enter `FairTaskQueue`, are selected through virtual-time ordering, and are committed only after `TaskAdmissionController` acquires the required scheduler resources.
 
 ### Execution Graph
 
@@ -121,7 +121,7 @@ DatasetBuilder.build()
   → _prepare_async_run()
       → ExecutionGraph.create()
       → CompletionTracker.with_graph()
-      → AsyncTaskScheduler(semaphores, salvage_rounds)
+      → AsyncTaskScheduler(task admission, fair queue, salvage_rounds)
   → scheduler.run()
       → for each row group, fairly admit ready tasks from frontier
       → tasks execute generators, update CompletionTracker
@@ -133,7 +133,7 @@ DatasetBuilder.build()
 
 - **Dual execution engines behind one API.** The sequential engine is simpler and easier to debug; the async engine adds row-group parallelism for throughput. Users switch via an environment variable without changing their code.
 - **DAG-driven ordering** ensures columns with dependencies (e.g., a judge column that depends on a text column) are generated in the correct order, regardless of the order they appear in the config.
-- **Fair async admission** keeps the scheduler flowing across ready columns and model groups. Global semaphores still bound memory/coroutine growth, while per-group virtual-time queues prevent a large ready frontier from degenerating into a column-by-column wave. LLM admission caps are peer-sensitive: a solo model group can fill available global capacity, but once another scheduling group has queued work the saturated group yields until peers get admission slots or admitted tasks complete.
+- **Fair async admission** keeps the scheduler flowing across ready columns and model groups. `FairTaskQueue.select_next(...)` chooses eligible ready work, `TaskAdmissionController` leases scheduler resources before spawn, and `FairTaskQueue.commit(...)` removes the selected task only after admission succeeds. Per-group virtual-time ordering prevents a large ready frontier from degenerating into a column-by-column wave, and scheduler-resource accounting remains separate from provider/model request admission.
 - **Salvage rounds in async mode** retry failed tasks after all other tasks in a round complete, improving resilience against transient LLM failures without blocking the entire generation.
 - **Unified DAG construction.** `topologically_sort_column_configs` (in `execution_graph.py`) determines column ordering using Kahn's algorithm; the runtime `ExecutionGraph` adds strategy-aware dependency tracking for the async scheduler.
 

architecture/models.md

@@ -1,6 +1,6 @@
 # Models
 
-The model subsystem provides a unified interface for LLM access: chat completions, embeddings, and image generation. It handles client creation, retry, rate-limit throttling, usage tracking, and MCP tool integration.
+The model subsystem provides a unified interface for LLM access: chat completions, embeddings, and image generation. It handles client creation, retry, request admission, usage tracking, and MCP tool integration.
 
 Source: `packages/data-designer-engine/src/data_designer/engine/models/`
 
@@ -11,12 +11,12 @@ The model subsystem is layered:
 ```
 ModelRegistry (lazy facade-per-alias)
   └── ModelFacade (completion, embeddings, image gen, MCP tool loops)
-        └── ThrottledModelClient (AIMD rate limiting)
+        └── ModelRequestExecutor (request admission + provider execution)
               └── ModelClient (OpenAI-compatible or Anthropic adapter)
                     └── RetryTransport (httpx-level retries)
 ```
 
-Generators never interact with HTTP clients directly. They request a `ModelFacade` by alias from the `ModelRegistry`, which handles lazy construction and shared throttle state.
+Generators never interact with HTTP clients directly. They request a `ModelFacade` by alias from the `ModelRegistry`, which handles lazy construction, request-resource canonicalization, and shared adaptive request admission state.
 
 ## Key Components
 
@@ -31,13 +31,13 @@ Defines the contract: sync/async chat, embeddings, image generation, `supports_*
 
 `create_model_client` routes by provider type to the appropriate adapter. Optionally wraps with:
 - **`RetryTransport`** — httpx-level retries via `httpx_retries.RetryTransport`. `HttpModelClient` sets `strip_rate_limit_codes=True` for the async client and `False` for the sync client (`http_model_client.py`), which controls whether 429 responses are eligible for transport-layer retries.
-- **`ThrottledModelClient`** — AIMD (Additive Increase, Multiplicative Decrease) concurrency control per throttle domain.
+- **`ModelRequestExecutor`** — maps model-call attempts to request-admission items, acquires request leases, invokes the provider client, and releases the exact lease on every terminal path.
 
-### ThrottleManager
+### Request Admission
 
-Manages concurrency limits per `ThrottleDomain` (CHAT, EMBEDDING, IMAGE, HEALTHCHECK), keyed by `(provider_name, model_id)`. Thread-safe with a shared lock for sync/async access.
+`RequestAdmissionController` manages provider/model/domain request resources. `AdaptiveRequestAdmissionController` adds AIMD (Additive Increase, Multiplicative Decrease) adaptation per `RequestDomain` (`chat`, `embedding`, `image`, `healthcheck`) under the provider/model static cap.
 
-`ThrottledModelClient` wraps each API call in a context manager that acquires/releases throttle capacity and adjusts limits on success (additive increase) or rate-limit errors (multiplicative decrease).
+`ModelRequestExecutor` wraps each provider call with a request-admission lease and feeds success or rate-limit outcomes back to the controller. `RequestResourceResolver` owns canonical provider/model/domain identity so aliases that target the same endpoint share request capacity.
 
 ### ModelFacade
 
@@ -50,7 +50,7 @@ The primary interface for generators. Holds a `ModelConfig`, `ModelClient`, opti
 
 ### ModelRegistry
 
-Lazy `ModelFacade` construction per alias. Registers a shared `ThrottleManager` across all facades for coordinated rate limiting. Provides `get_model_usage_stats` and `log_model_usage` for post-build reporting.
+Lazy `ModelFacade` construction per alias. Registers shared request-admission state across all facades for coordinated provider/model/domain capacity. Provides `get_model_usage_stats` and `log_model_usage` for post-build reporting.
 
 ### Usage Tracking
 
@@ -59,18 +59,18 @@ Lazy `ModelFacade` construction per alias. Registers a shared `ThrottleManager`
 ## Data Flow
 
 1. Generator requests a model by alias from `ModelRegistry`
-2. Registry lazily creates `ModelFacade` with the appropriate client and throttle config
+2. Registry lazily creates `ModelFacade` with the appropriate client and request-admission executor
 3. Generator calls `completion()` with prompt/messages
-4. `ModelFacade` builds kwargs, calls `ThrottledModelClient`
-5. Throttle layer acquires capacity, delegates to `ModelClient`
+4. `ModelFacade` builds kwargs, calls `ModelRequestExecutor`
+5. Request admission acquires a provider/model/domain lease, delegates to `ModelClient`
 6. `ModelClient` makes the HTTP request through `RetryTransport`
 7. Response flows back; usage is tracked; if MCP tools are configured, tool calls are executed and results fed back for another completion round
 
 ## Design Decisions
 
-- **Facade pattern** hides HTTP, retry, throttle, and MCP complexity from generators. Generators see `completion()` and get back parsed results.
-- **AIMD throttling at the application layer** rather than relying solely on HTTP retries. This provides smoother throughput under rate limits — the transport layer still handles many transient failures, while the throttle manager adjusts concurrency to avoid sustained 429 storms.
-- **429 handling depends on sync vs async `HttpModelClient`** — The async client uses `strip_rate_limit_codes=True`, so 429s are not retried at the transport layer and rate-limit signals reach `ThrottledModelClient` / AIMD quickly. The sync client uses `strip_rate_limit_codes=False`, so 429s may still be retried transparently at the transport layer before surfacing to callers.
+- **Facade pattern** hides HTTP, retry, request admission, and MCP complexity from generators. Generators see `completion()` and get back parsed results.
+- **AIMD request admission at the application layer** rather than relying solely on HTTP retries. This provides smoother throughput under rate limits: the transport layer still handles many transient failures, while adaptive request admission adjusts concurrency to avoid sustained 429 storms.
+- **429 handling depends on sync vs async `HttpModelClient`** — The async client uses `strip_rate_limit_codes=True`, so 429s are not retried at the transport layer and rate-limit signals reach `ModelRequestExecutor` / request admission quickly. The sync client uses `strip_rate_limit_codes=False`, so 429s may still be retried transparently at the transport layer before surfacing to callers.
 - **Distribution-valued inference parameters** (`temperature`, `top_p` as `UniformDistribution` or `ManualDistribution`) enable controlled randomness across a dataset without per-row config changes.
 - **Lazy facade construction** avoids health-checking or connecting to models that are configured but never used in a particular generation run.
 

architecture/overview.md

@@ -30,7 +30,7 @@ Users declare what their data should look like through config objects (columns,
 | `DataDesigner` | `data-designer` | Public API — `create()`, `preview()`, `validate()` |
 | `DataDesignerConfigBuilder` | `data-designer-config` | Fluent builder for dataset configs |
 | `DatasetBuilder` | `data-designer-engine` | Orchestrates generation (sync or async) |
-| `ModelFacade` / `ModelRegistry` | `data-designer-engine` | LLM client abstraction with retry, throttle, usage tracking |
+| `ModelFacade` / `ModelRegistry` | `data-designer-engine` | LLM client abstraction with retry, request admission, usage tracking |
 | `MCPFacade` / `MCPRegistry` | `data-designer-engine` | Tool execution via Model Context Protocol |
 | `ColumnGeneratorRegistry` | `data-designer-engine` | Maps column types to generator implementations |
 | `PluginRegistry` | `data-designer-config` | Discovers and registers entry-point plugins |
@@ -44,7 +44,7 @@ Users declare what their data should look like through config objects (columns,
 
 3. **Generation** — `DatasetBuilder` instantiates column generators from the registry, then executes one of two paths:
    - **Sequential** (default): batch loop over columns in topological order. Each generator produces its column via `CELL_BY_CELL` (threaded fan-out) or `FULL_COLUMN` strategy.
-   - **Async** (`DATA_DESIGNER_ASYNC_ENGINE=1`): builds an `ExecutionGraph`, partitions rows into groups, and dispatches tasks via `AsyncTaskScheduler` with semaphore-based concurrency, salvage rounds, and per-row-group checkpointing.
+   - **Async** (`DATA_DESIGNER_ASYNC_ENGINE=1`): builds an `ExecutionGraph`, partitions rows into groups, and dispatches tasks via `AsyncTaskScheduler` with `FairTaskQueue` selection, `TaskAdmissionController` scheduler-resource leases, salvage rounds, and per-row-group checkpointing.
 
 4. **Post-processing** — `ProcessorRunner` applies transformations (pre-batch, post-batch, after-generation). Profilers analyze the generated dataset.
 
@@ -61,7 +61,7 @@ Users declare what their data should look like through config objects (columns,
 
 - [Config Layer](config.md) — builder API, column types, model configs, plugin system
 - [Engine Layer](engine.md) — compilation, generators, registries
-- [Models](models.md) — model facade, adapters, retry/throttle
+- [Models](models.md) — model facade, adapters, retry, request admission
 - [Dataset Builders](dataset-builders.md) — sync/async orchestration, DAG, batching
 - [MCP](mcp.md) — tool execution, session pooling
 - [Sampling](sampling.md) — statistical generators, person/entity data

artifacts/645-live-bench-nano/agent-aimd-scale/README.md

@@ -0,0 +1,48 @@
+# 645 live bench nano AIMD scale lane
+
+Ran in `/Users/etramel/src/DataDesigner` on branch `scheduling-yolo` with only `/tmp/dd_live_aimd_bench.py` and artifacts under this directory. No tracked repo files were edited.
+
+## Configuration
+
+- Model: `openai/openai/gpt-5-nano`
+- Provider: `nvidia-internal`
+- Temperature: omitted
+- `skip_health_check=True`
+- `DATA_DESIGNER_ASYNC_ENGINE=1`
+- `DATA_DESIGNER_ASYNC_TRACE=1`
+- `max_parallel_requests=16`
+- AIMD initial limit: 1
+- AIMD `increase_after_successes=16`
+- Shape for final scenarios: 512 rows x 2 independent `LLMTextColumnConfig` columns = 1024 model generations
+- Event instrumentation: `InMemoryAdmissionEventSink`, patched request admission init, model-client factory `request_event_sink`, and scheduler init `scheduler_event_sink`
+
+## Scenarios
+
+| Scenario | Buffer | Requests | Success | Failures | Wall s | Time to cap s | Max in-flight | Max waiters | Request wait p50 / p95 / max s |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| diagnostic-4x2-buffer32 | 32 | 8 | 8 | 0 | 5.586 | n/a | 2 | 7 | 2.638 / 3.871 / 3.887 |
+| final-512x2-buffer32 | 32 | 1024 | 1024 | 0 | 104.823 | 50.332 | 16 | 63 | 3.180 / 19.371 / 29.990 |
+| final-512x2-buffer512 | 512 | 1024 | 1024 | 0 | 114.522 | 57.062 | 16 | 63 | 3.283 / 22.949 / 35.324 |
+
+## Observations
+
+- Both final scenarios completed exactly 1024 `model_request_started` and 1024 `model_request_completed` events, with 0 failed model requests and no fallback model.
+- AIMD limit increased monotonically from 1 through 16 in both final scenarios. There were 15 `request_limit_increased` events, 0 decreases, and 0 rate-limit events in each final scenario.
+- Cap enforcement held: observed request in-flight max was 16 in both final scenarios, matching `max_parallel_requests=16`.
+- `buffer_size=32` reached cap faster (50.332s) and completed faster (104.823s) than `buffer_size=512` (57.062s to cap, 114.522s wall).
+- Request wait p95 was lower for `buffer_size=32` (19.371s) than `buffer_size=512` (22.949s).
+- Traffic became steady after the initial AIMD ramp in both final scenarios; see each `flow_buckets.json` for per-second starts/completions and `monitor_samples.jsonl` for sampled pressure snapshots.
+
+## Artifacts
+
+Each scenario directory contains:
+
+- `timeline.jsonl`
+- `request_events.jsonl`
+- `monitor_samples.jsonl`
+- `task_traces.csv`
+- `task_traces.json`
+- `flow_buckets.json`
+- `summary.json`
+
+Combined summary: `combined_summary.json`.