perf(db): index LiteLLM_SpendLogs on (api_key, startTime)

The Logs tab filters LiteLLM_SpendLogs by api_key inside a startTime
window. The table only had indexes on startTime, (startTime, request_id),
end_user and session_id, so the api_key predicate degraded to a full scan
of the window on both the pagination count and the paged select, re-run on
every page turn. On a multi-month table that pinned a customer's Aurora
writer near 100% CPU for ~50 minutes.

Built CONCURRENTLY: this is the largest table on most deployments, and a
blocking build would stall spend-log inserts for the whole build.

litellm-proxy-extras/litellm_proxy_extras/migrations/20260818000000_add_spendlogs_api_key_start_time_idx/migration.sql

@@ -0,0 +1,11 @@
+-- CreateIndex (CONCURRENTLY)
+--
+-- Disclaimer:
+-- - CREATE INDEX CONCURRENTLY cannot run inside a transaction. This migration must stay a
+--   single statement so Prisma Migrate on PostgreSQL can apply it outside a transaction.
+-- - "LiteLLM_SpendLogs" is the largest table on most deployments, so a blocking build would
+--   stall spend-log inserts for the whole build; if the build is interrupted, Postgres may
+--   leave an INVALID index that must be dropped and recreated.
+-- - Do not edit this file after it has been applied to any database: Prisma checksums
+--   migrations; add a new migration instead.
+CREATE INDEX CONCURRENTLY IF NOT EXISTS "LiteLLM_SpendLogs_api_key_startTime_idx" ON "LiteLLM_SpendLogs"("api_key", "startTime" DESC);

litellm-proxy-extras/litellm_proxy_extras/schema.prisma

@@ -645,6 +645,7 @@ model LiteLLM_SpendLogs {
   @@index([startTime, request_id])
   @@index([end_user])
   @@index([session_id])
+  @@index([api_key, startTime(sort: Desc)], map: "LiteLLM_SpendLogs_api_key_startTime_idx")
 }
 
 // View spend, model, api_key per request

schema.prisma

@@ -645,6 +645,7 @@ model LiteLLM_SpendLogs {
   @@index([startTime, request_id])
   @@index([end_user])
   @@index([session_id])
+  @@index([api_key, startTime(sort: Desc)], map: "LiteLLM_SpendLogs_api_key_startTime_idx")
 }
 
 // View spend, model, api_key per request

tests/proxy_migration_tests/test_db_schema_migration.py

@@ -68,3 +68,43 @@ def test_schema_migration_in_sync():
         assert diff.returncode == 0, f"prisma migrate diff errored: {diff.stderr}"
     finally:
         shutil.rmtree(temp_base, ignore_errors=True)
+
+
+SPEND_LOGS_API_KEY_INDEX = "LiteLLM_SpendLogs_api_key_startTime_idx"
+
+SCHEMA_PATHS = (
+    Path("./schema.prisma"),
+    Path("./litellm-proxy-extras/litellm_proxy_extras/schema.prisma"),
+)
+
+
+def _spend_logs_model_block(schema: str) -> str:
+    start = schema.index("model LiteLLM_SpendLogs {")
+    return schema[start : schema.index("\n}", start)]
+
+
+@pytest.mark.parametrize("schema_path", SCHEMA_PATHS, ids=lambda p: str(p))
+def test_spend_logs_declares_api_key_index(schema_path: Path):
+    """The Logs tab filters ``LiteLLM_SpendLogs`` by ``api_key`` inside a ``startTime``
+    window. Without an index led by ``api_key`` that degrades to a full scan of the
+    window on every page turn, which pinned a customer's Aurora writer at ~99% CPU.
+    """
+    block = _spend_logs_model_block(schema_path.read_text())
+    assert f'@@index([api_key, startTime(sort: Desc)], map: "{SPEND_LOGS_API_KEY_INDEX}")' in block, (
+        f"{schema_path} must keep the api_key-led index on LiteLLM_SpendLogs"
+    )
+
+
+def test_spend_logs_api_key_index_has_a_migration():
+    """A schema-only index never reaches deployments that run ``prisma migrate deploy``."""
+    migrations = Path("./litellm-proxy-extras/litellm_proxy_extras/migrations")
+    creating = [
+        sql
+        for sql in (path.read_text() for path in migrations.glob("*/migration.sql"))
+        if SPEND_LOGS_API_KEY_INDEX in sql
+    ]
+    assert len(creating) == 1, f"expected exactly one migration creating {SPEND_LOGS_API_KEY_INDEX}, found {len(creating)}"
+    assert 'ON "LiteLLM_SpendLogs"("api_key", "startTime" DESC)' in creating[0]
+    assert "CONCURRENTLY" in creating[0], (
+        "LiteLLM_SpendLogs is the largest table on most deployments; a blocking build stalls spend-log inserts"
+    )