fix(routing): retain adaptive feedback after failed flushes

litellm/router_strategy/adaptive_router/README.md

@@ -67,10 +67,22 @@ Callers may pass header `x-litellm-min-quality-tier: 3` (or metadata key
   (uptime, not quality). Skipped if conversation has fewer than
   `SIGNAL_GATE_MIN_MESSAGES` messages.
 - **Persistence.** Bandit cells: aggregated deltas, eventually consistent.
-  Session rows: last-write-wins snapshots.
+  Session rows: last-write-wins snapshots. Failed or cancelled flushes retain
+  unacknowledged rows for the next flush. Retried deltas merge with new feedback;
+  a newer queued session snapshot replaces a failed older snapshot. Flushes of
+  the same queue are serialized within one process, and their return values
+  count acknowledged writes only
 
 ## Known v0 limitations
 
+- **Retries are at-least-once.** If a state increment commits but its response
+  is lost, retrying can count that feedback twice. There is no durable receipt
+  to distinguish that outcome from a failed write. The queues are in memory,
+  so process termination still loses pending updates. Session write ordering
+  across workers, or server queries continuing after client cancellation, is
+  not guaranteed. A prolonged database outage retains pending session keys
+  in memory until a later flush succeeds
+
 - **Latency is not in the score.** Quality + cost only. A pathologically slow
   model can still be picked.
 - **Hard sample cap at 200.** Once `α + β > 200`, deltas are silently dropped.

litellm/router_strategy/adaptive_router/update_queue.py

@@ -42,6 +42,8 @@ class AdaptiveRouterUpdateQueue:
         self._state_agg: dict[StateKey, dict[str, float]] = {}
         self._session_agg: dict[SessionKey, Mapping[str, object]] = {}
         self._lock = asyncio.Lock()
+        self._state_flush_lock = asyncio.Lock()
+        self._session_flush_lock = asyncio.Lock()
         self._max_state_size_seen = 0
         self._max_session_size_seen = 0
 
@@ -93,6 +95,10 @@ class AdaptiveRouterUpdateQueue:
     # ---- Flushers (called by background task) ----------------------------
 
     async def flush_state_to_db(self, prisma_client: object) -> int:
+        async with self._state_flush_lock:
+            return await self._flush_state_to_db(prisma_client)
+
+    async def _flush_state_to_db(self, prisma_client: object) -> int:
         """
         Drain state aggregator and apply to LiteLLM_AdaptiveRouterState.
         Returns number of cells flushed.
@@ -106,49 +112,57 @@ class AdaptiveRouterUpdateQueue:
 
         # Sort keys to give deterministic write order across writers and
         # reduce the chance of cross-row deadlocks when other workers race us.
-        for key in sorted(batch.keys()):
-            router, rt, model = key
-            payload = batch[key]
-            try:
-                # Atomic increment: push the delta directly into the DB so
-                # concurrent flushers from multiple pods don't overwrite each
-                # other. The upsert creates the row with the delta as the
-                # initial value on first write, then increments on subsequent
-                # writes — no read-modify-write race.
-                await AdaptiveRouterStateRepository(prisma_client).table.upsert(
-                    where={
-                        "router_name_request_type_model_name": {
-                            "router_name": router,
-                            "request_type": rt,
-                            "model_name": model,
-                        }
-                    },
-                    data={
-                        "create": {
-                            "router_name": router,
-                            "request_type": rt,
-                            "model_name": model,
-                            "alpha": payload["delta_alpha"],
-                            "beta": payload["delta_beta"],
-                            "total_samples": int(payload["samples_added"]),
+        pending: Final = dict(batch)  # mutable-ok: tracks unacknowledged rows until the batch is restored
+        try:
+            for key, payload in sorted(batch.items()):
+                router, rt, model = key
+                try:
+                    # Atomic increment: push the delta directly into the DB so
+                    # concurrent flushers from multiple pods don't overwrite each
+                    # other. The upsert creates the row with the delta as the
+                    # initial value on first write, then increments on subsequent
+                    # writes — no read-modify-write race.
+                    await AdaptiveRouterStateRepository(prisma_client).table.upsert(
+                        where={
+                            "router_name_request_type_model_name": {
+                                "router_name": router,
+                                "request_type": rt,
+                                "model_name": model,
+                            }
                         },
-                        "update": {
-                            "alpha": {"increment": payload["delta_alpha"]},
-                            "beta": {"increment": payload["delta_beta"]},
-                            "total_samples": {"increment": int(payload["samples_added"])},
+                        data={
+                            "create": {
+                                "router_name": router,
+                                "request_type": rt,
+                                "model_name": model,
+                                "alpha": payload["delta_alpha"],
+                                "beta": payload["delta_beta"],
+                                "total_samples": int(payload["samples_added"]),
+                            },
+                            "update": {
+                                "alpha": {"increment": payload["delta_alpha"]},
+                                "beta": {"increment": payload["delta_beta"]},
+                                "total_samples": {"increment": int(payload["samples_added"])},
+                            },
                         },
-                    },
-                )
-            except Exception as e:
-                verbose_router_logger.exception(
-                    "AdaptiveRouterUpdateQueue: failed to flush state for %s: %s",
-                    key,
-                    e,
-                )
+                    )
+                    pending.pop(key)
+                except Exception as e:
+                    verbose_router_logger.exception(
+                        "AdaptiveRouterUpdateQueue: failed to flush state for %s: %s",
+                        key,
+                        e,
+                    )
+        finally:
+            await self._restore_state_batch(pending)
 
-        return len(batch)
+        return len(batch) - len(pending)
 
     async def flush_session_to_db(self, prisma_client: object) -> int:
+        async with self._session_flush_lock:
+            return await self._flush_session_to_db(prisma_client)
+
+    async def _flush_session_to_db(self, prisma_client: object) -> int:
         """
         Drain session aggregator and upsert into LiteLLM_AdaptiveRouterSession.
         Returns number of session rows flushed.
@@ -160,45 +174,63 @@ class AdaptiveRouterUpdateQueue:
         if not batch:
             return 0
 
-        for key in sorted(batch.keys()):
-            session_id, router, model = key
-            payload = batch[key]
-            try:
-                # NOTE: Prisma client lower-cases model names, so
-                # `LiteLLM_AdaptiveRouterSession` -> `litellm_adaptiveroutersession`
-                # (single 's', not 'litellm_adaptiverouterssession').
-                # Strip PK fields from the update payload — Prisma rejects
-                # writes to fields that are part of the @@id. asdict(state)
-                # always carries them, so build a separate update dict.
-                update_payload = {
-                    k: v for k, v in payload.items() if k not in ("session_id", "router_name", "model_name")
-                }
-                await AdaptiveRouterSessionRepository(prisma_client).table.upsert(
-                    where={
-                        "session_id_router_name_model_name": {
-                            "session_id": session_id,
-                            "router_name": router,
-                            "model_name": model,
-                        }
-                    },
-                    data={
-                        "create": {
-                            "session_id": session_id,
-                            "router_name": router,
-                            "model_name": model,
-                            **update_payload,
+        pending: Final = dict(batch)  # mutable-ok: tracks unacknowledged rows until the batch is restored
+        try:
+            for key, payload in sorted(batch.items()):
+                session_id, router, model = key
+                try:
+                    # NOTE: Prisma client lower-cases model names, so
+                    # `LiteLLM_AdaptiveRouterSession` -> `litellm_adaptiveroutersession`
+                    # (single 's', not 'litellm_adaptiverouterssession').
+                    # Strip PK fields from the update payload — Prisma rejects
+                    # writes to fields that are part of the @@id. asdict(state)
+                    # always carries them, so build a separate update dict.
+                    update_payload = {
+                        k: v for k, v in payload.items() if k not in ("session_id", "router_name", "model_name")
+                    }
+                    await AdaptiveRouterSessionRepository(prisma_client).table.upsert(
+                        where={
+                            "session_id_router_name_model_name": {
+                                "session_id": session_id,
+                                "router_name": router,
+                                "model_name": model,
+                            }
                         },
-                        "update": update_payload,
-                    },
-                )
-            except Exception as e:
-                verbose_router_logger.exception(
-                    "AdaptiveRouterUpdateQueue: failed to flush session for %s: %s",
-                    key,
-                    e,
-                )
+                        data={
+                            "create": {
+                                "session_id": session_id,
+                                "router_name": router,
+                                "model_name": model,
+                                **update_payload,
+                            },
+                            "update": update_payload,
+                        },
+                    )
+                    pending.pop(key)
+                except Exception as e:
+                    verbose_router_logger.exception(
+                        "AdaptiveRouterUpdateQueue: failed to flush session for %s: %s",
+                        key,
+                        e,
+                    )
+        finally:
+            await self._restore_session_batch(pending)
 
-        return len(batch)
+        return len(batch) - len(pending)
+
+    async def _restore_state_batch(self, pending: Mapping[StateKey, Mapping[str, float]]) -> None:
+        async with self._lock:
+            for key, payload in pending.items():
+                self._state_agg[key] = {  # mutable-ok: queue-owned accumulator merges unacknowledged deltas
+                    field: value + (self._state_agg[key][field] if key in self._state_agg else 0)
+                    for field, value in payload.items()
+                }
+            self._max_state_size_seen = max(self._max_state_size_seen, len(self._state_agg))
+
+    async def _restore_session_batch(self, pending: Mapping[SessionKey, Mapping[str, object]]) -> None:
+        async with self._lock:
+            self._session_agg = {**pending, **self._session_agg}  # mutable-ok: newer queued snapshots win over retries
+            self._max_session_size_seen = max(self._max_session_size_seen, len(self._session_agg))
 
     # ---- Observability ---------------------------------------------------
 

tests/unit/router_strategy/adaptive_router/test_update_queue.py

@@ -115,3 +115,193 @@ async def test_max_size_observability(queue):
     await queue.add_state_delta("r1", "code_generation", "gpt-4", 1.0, 0.0)
     sizes = await queue.queue_size()
     assert sizes["max_state_seen"] >= 3
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("kind", ["state", "session"])
+async def test_failed_flush_retains_only_failed_rows_and_reports_successes(queue, mock_prisma, kind):
+    for model in ("a", "b", "c"):
+        if kind == "state":
+            await queue.add_state_delta("router", "general", model, 1.0, 0.5)
+        else:
+            await queue.add_session_state("session", "router", model, {"turn_count": 1})
+    table = getattr(mock_prisma.db, f"litellm_adaptiverouter{kind}")
+    table.upsert.side_effect = [None, RuntimeError("database unavailable"), None, None]
+    flush = getattr(queue, f"flush_{kind}_to_db")
+
+    assert await flush(mock_prisma) == 2
+    assert (await queue.queue_size())[f"{kind}_pending"] == 1
+    assert await flush(mock_prisma) == 1
+    assert (await queue.queue_size())[f"{kind}_pending"] == 0
+    assert [call.kwargs["data"]["create"]["model_name"] for call in table.upsert.call_args_list] == ["a", "b", "c", "b"]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("kind", ["state", "session"])
+async def test_failed_flush_merges_concurrent_updates_without_replaying_successes(queue, mock_prisma, kind):
+    started = asyncio.Event()
+    release = asyncio.Event()
+
+    async def fail_write(**kwargs):
+        started.set()
+        await release.wait()
+        raise RuntimeError("database unavailable")
+
+    table = getattr(mock_prisma.db, f"litellm_adaptiverouter{kind}")
+    table.upsert.side_effect = fail_write
+    flush = getattr(queue, f"flush_{kind}_to_db")
+    if kind == "state":
+        await queue.add_state_delta("router", "general", "model", 1.0, 0.5)
+    else:
+        await queue.add_session_state("session", "router", "model", {"turn_count": 1})
+    task = asyncio.create_task(flush(mock_prisma))
+    await started.wait()
+    if kind == "state":
+        await queue.add_state_delta("router", "general", "model", 2.0, 1.0)
+    else:
+        await queue.add_session_state("session", "router", "model", {"turn_count": 2})
+    release.set()
+    assert await task == 0
+
+    table.upsert.side_effect = None
+    assert await flush(mock_prisma) == 1
+    update = table.upsert.call_args.kwargs["data"]["update"]
+    assert update == (
+        {"alpha": {"increment": 3.0}, "beta": {"increment": 1.5}, "total_samples": {"increment": 2}}
+        if kind == "state"
+        else {"turn_count": 2}
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("kind", ["state", "session"])
+async def test_cancelled_flush_retains_unfinished_rows_without_replaying_acknowledged_rows(queue, mock_prisma, kind):
+    started = asyncio.Event()
+    blocked = asyncio.Event()
+
+    async def write(**kwargs):
+        if kwargs["data"]["create"]["model_name"] == "b":
+            started.set()
+            await blocked.wait()
+
+    for model in ("a", "b", "c"):
+        if kind == "state":
+            await queue.add_state_delta("router", "general", model, 1.0, 0.5)
+        else:
+            await queue.add_session_state("session", "router", model, {"turn_count": 1})
+    table = getattr(mock_prisma.db, f"litellm_adaptiverouter{kind}")
+    table.upsert.side_effect = write
+    flush = getattr(queue, f"flush_{kind}_to_db")
+    task = asyncio.create_task(flush(mock_prisma))
+    await started.wait()
+    task.cancel()
+    with pytest.raises(asyncio.CancelledError):
+        await task
+
+    assert (await queue.queue_size())[f"{kind}_pending"] == 2
+    table.upsert.side_effect = None
+    assert await flush(mock_prisma) == 2
+    assert [call.kwargs["data"]["create"]["model_name"] for call in table.upsert.call_args_list] == ["a", "b", "b", "c"]
+
+
+@pytest.mark.asyncio
+async def test_overlapping_session_flushes_cannot_persist_an_older_snapshot_last(queue, mock_prisma):
+    started = asyncio.Event()
+    release = asyncio.Event()
+    stored = []
+
+    async def write(**kwargs):
+        turn_count = kwargs["data"]["update"]["turn_count"]
+        if turn_count == 1:
+            started.set()
+            await release.wait()
+        stored.append(turn_count)
+
+    mock_prisma.db.litellm_adaptiveroutersession.upsert.side_effect = write
+    await queue.add_session_state("session", "router", "model", {"turn_count": 1})
+    first = asyncio.create_task(queue.flush_session_to_db(mock_prisma))
+    await started.wait()
+    await queue.add_session_state("session", "router", "model", {"turn_count": 2})
+    second = asyncio.create_task(queue.flush_session_to_db(mock_prisma))
+    await asyncio.sleep(0)
+    release.set()
+
+    assert await asyncio.gather(first, second) == [1, 1]
+    assert stored == [1, 2]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("kind", ["state", "session"])
+async def test_repeated_write_failures_keep_the_queue_until_recovery(queue, mock_prisma, kind):
+    if kind == "state":
+        await queue.add_state_delta("router", "general", "model", 1.0, 0.5)
+    else:
+        await queue.add_session_state("session", "router", "model", {"turn_count": 1})
+    table = getattr(mock_prisma.db, f"litellm_adaptiverouter{kind}")
+    table.upsert.side_effect = RuntimeError("database unavailable")
+    flush = getattr(queue, f"flush_{kind}_to_db")
+    for _ in range(3):
+        assert await flush(mock_prisma) == 0
+        assert (await queue.queue_size())[f"{kind}_pending"] == 1
+    table.upsert.side_effect = None
+
+    assert await flush(mock_prisma) == 1
+    assert await flush(mock_prisma) == 0
+    update = table.upsert.call_args.kwargs["data"]["update"]
+    assert update == (
+        {"alpha": {"increment": 1.0}, "beta": {"increment": 0.5}, "total_samples": {"increment": 1}}
+        if kind == "state"
+        else {"turn_count": 1}
+    )
+
+
+@pytest.mark.asyncio
+async def test_unacknowledged_state_commit_is_retried_with_at_least_once_semantics(queue, mock_prisma):
+    persisted = []
+
+    async def commit_then_lose_response(**kwargs):
+        persisted.append(kwargs["data"]["update"])
+        if len(persisted) == 1:
+            raise ConnectionError("commit response lost")
+
+    mock_prisma.db.litellm_adaptiverouterstate.upsert.side_effect = commit_then_lose_response
+    await queue.add_state_delta("router", "general", "model", 1.0, 0.5)
+
+    assert await queue.flush_state_to_db(mock_prisma) == 0
+    assert (await queue.queue_size())["state_pending"] == 1
+    assert await queue.flush_state_to_db(mock_prisma) == 1
+    assert (
+        persisted == [{"alpha": {"increment": 1.0}, "beta": {"increment": 0.5}, "total_samples": {"increment": 1}}] * 2
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("kind", ["state", "session"])
+async def test_queue_high_water_mark_includes_restored_rows_and_new_keys(queue, mock_prisma, kind):
+    started = asyncio.Event()
+    release = asyncio.Event()
+
+    async def fail_write(**kwargs):
+        started.set()
+        await release.wait()
+        raise RuntimeError("database unavailable")
+
+    table = getattr(mock_prisma.db, f"litellm_adaptiverouter{kind}")
+    table.upsert.side_effect = fail_write
+    if kind == "state":
+        await queue.add_state_delta("router", "general", "a", 1.0, 0.5)
+    else:
+        await queue.add_session_state("session", "router", "a", {"turn_count": 1})
+    task = asyncio.create_task(getattr(queue, f"flush_{kind}_to_db")(mock_prisma))
+    await started.wait()
+    for model in ("b", "c"):
+        if kind == "state":
+            await queue.add_state_delta("router", "general", model, 1.0, 0.5)
+        else:
+            await queue.add_session_state("session", "router", model, {"turn_count": 1})
+    release.set()
+
+    assert await task == 0
+    sizes = await queue.queue_size()
+    assert sizes[f"{kind}_pending"] == 3
+    assert sizes[f"max_{kind}_seen"] == 3