feat(BA-4330): Add configurable BatchSpanProcessor queue and batch size

Add max_queue_size and max_export_batch_size to OTELConfig and
OpenTelemetrySpec, defaulting to 65536 and 4096 respectively.
The SDK defaults (2048/512) are insufficient for production GraphQL
workloads and cause span drops during burst traffic.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: hhoikoo

src/ai/backend/agent/server.py

@@ -1521,6 +1521,8 @@ async def service_discovery_ctx(
             endpoint=local_config.otel.endpoint,
             service_instance_id=meta.id,
             service_instance_name=meta.display_name,
+            max_queue_size=local_config.otel.max_queue_size,
+            max_export_batch_size=local_config.otel.max_export_batch_size,
         )
         BraceStyleAdapter.apply_otel(otel_spec)
     try:

src/ai/backend/appproxy/coordinator/server.py

@@ -767,6 +767,8 @@ async def service_discovery_ctx(root_ctx: RootContext) -> AsyncIterator[None]:
             endpoint=root_ctx.local_config.otel.endpoint,
             service_instance_id=meta.id,
             service_instance_name=meta.display_name,
+            max_queue_size=root_ctx.local_config.otel.max_queue_size,
+            max_export_batch_size=root_ctx.local_config.otel.max_export_batch_size,
         )
         BraceStyleAdapter.apply_otel(otel_spec)
     try:

src/ai/backend/appproxy/worker/server.py

@@ -560,6 +560,8 @@ async def service_discovery_ctx(root_ctx: RootContext) -> AsyncIterator[None]:
             endpoint=root_ctx.local_config.otel.endpoint,
             service_instance_id=meta.id,
             service_instance_name=meta.display_name,
+            max_queue_size=root_ctx.local_config.otel.max_queue_size,
+            max_export_batch_size=root_ctx.local_config.otel.max_export_batch_size,
         )
         BraceStyleAdapter.apply_otel(otel_spec)
     try:

src/ai/backend/common/configs/otel.py

@@ -58,3 +58,36 @@ class OTELConfig(BaseConfigSchema):
             example=ConfigExample(local="http://localhost:4317", prod="http://otel-collector:4317"),
         ),
     ]
+    max_queue_size: Annotated[
+        int,
+        Field(
+            default=65536,
+            validation_alias=AliasChoices("max-queue-size", "max_queue_size"),
+            serialization_alias="max-queue-size",
+        ),
+        BackendAIConfigMeta(
+            description=(
+                "Maximum number of spans queued for export. "
+                "Spans are dropped when the queue is full. "
+                "The default (65536) accommodates burst traffic from GraphQL workloads."
+            ),
+            added_version="26.2.0",
+            example=ConfigExample(local="2048", prod="65536"),
+        ),
+    ]
+    max_export_batch_size: Annotated[
+        int,
+        Field(
+            default=4096,
+            validation_alias=AliasChoices("max-export-batch-size", "max_export_batch_size"),
+            serialization_alias="max-export-batch-size",
+        ),
+        BackendAIConfigMeta(
+            description=(
+                "Maximum number of spans exported in a single batch. "
+                "Larger batches reduce export overhead but increase memory usage."
+            ),
+            added_version="26.2.0",
+            example=ConfigExample(local="512", prod="4096"),
+        ),
+    ]

src/ai/backend/logging/otel.py

@@ -25,6 +25,8 @@ class OpenTelemetrySpec:
     endpoint: str
     service_instance_id: uuid.UUID
     service_instance_name: str
+    max_queue_size: int
+    max_export_batch_size: int
 
     def to_resource(self) -> Resource:
         attributes = {
@@ -62,7 +64,11 @@ def apply_otel_loggers(loggers: Iterable[logging.Logger], spec: OpenTelemetrySpe
 def apply_otel_tracer(spec: OpenTelemetrySpec) -> None:
     tracer_provider = TracerProvider(resource=spec.to_resource())
     span_exporter = OTLPSpanExporter(endpoint=spec.endpoint)
-    span_processor = BatchSpanProcessor(span_exporter)
+    span_processor = BatchSpanProcessor(
+        span_exporter,
+        max_queue_size=spec.max_queue_size,
+        max_export_batch_size=spec.max_export_batch_size,
+    )
     tracer_provider.add_span_processor(span_processor)
     trace.set_tracer_provider(tracer_provider)
     logging.info("OpenTelemetry tracing initialized successfully.")

src/ai/backend/manager/server.py

@@ -844,13 +844,16 @@ async def service_discovery_ctx(root_ctx: RootContext) -> AsyncIterator[None]:
 
     if root_ctx.config_provider.config.otel.enabled:
         meta = root_ctx.sd_loop.metadata
+        otel_config = root_ctx.config_provider.config.otel
         otel_spec = OpenTelemetrySpec(
             service_name=meta.service_group,
             service_version=meta.version,
-            log_level=root_ctx.config_provider.config.otel.log_level,
-            endpoint=root_ctx.config_provider.config.otel.endpoint,
+            log_level=otel_config.log_level,
+            endpoint=otel_config.endpoint,
             service_instance_id=meta.id,
             service_instance_name=meta.display_name,
+            max_queue_size=otel_config.max_queue_size,
+            max_export_batch_size=otel_config.max_export_batch_size,
         )
         BraceStyleAdapter.apply_otel(otel_spec)
     try:

src/ai/backend/storage/server.py

@@ -543,6 +543,8 @@ async def service_discovery_ctx(
             endpoint=local_config.otel.endpoint,
             service_instance_id=meta.id,
             service_instance_name=meta.display_name,
+            max_queue_size=local_config.otel.max_queue_size,
+            max_export_batch_size=local_config.otel.max_export_batch_size,
         )
         BraceStyleAdapter.apply_otel(otel_spec)
     try:

src/ai/backend/web/server.py

@@ -917,6 +917,8 @@ async def service_discovery_ctx(config: WebServerUnifiedConfig) -> AsyncGenerato
             endpoint=config.otel.endpoint,
             service_instance_id=uuid.uuid4(),
             service_instance_name=instance_name,
+            max_queue_size=config.otel.max_queue_size,
+            max_export_batch_size=config.otel.max_export_batch_size,
         )
         BraceStyleAdapter.apply_otel(otel_spec)
     yield