WIP: Add overhead breakdown infrastructure for UI logs

[INCOMPLETE] This commit adds the infrastructure to display overhead breakdown
in UI logs but does NOT yet collect all the timing data.

What's implemented:
- Added overhead_breakdown field to SpendLogsMetadata
- Added extraction logic for cache_read_time_ms and retry_count (from existing data)
- Added UI component to display overhead breakdown in LogDetailsDrawer

What's NOT implemented (data collection missing):
- Auth time tracking (auth_time_ms)
- Request translation time tracking (request_translation_time_ms)
- Response translation time tracking (response_translation_time_ms)

The infrastructure follows existing patterns (same as cost_breakdown) and is
backward compatible. Currently only displays cache timing and retry count.
Auth and translation timing need to be added in follow-up commits.

litellm/proxy/_types.py

@@ -2845,6 +2845,9 @@ class SpendLogsMetadata(TypedDict):
     cost_breakdown: Optional[
         CostBreakdown
     ]  # Detailed cost breakdown (input_cost, output_cost, margin, discount, etc.)
+    overhead_breakdown: Optional[
+        dict
+    ]  # Detailed overhead breakdown: auth_time_ms, cache_read_time_ms, request_translation_time_ms, response_translation_time_ms, retry_count
 
 
 class SpendLogsPayload(TypedDict):

litellm/proxy/spend_tracking/spend_tracking_utils.py

@@ -62,6 +62,7 @@ def _get_spend_logs_metadata(
     cold_storage_object_key: Optional[str] = None,
     litellm_overhead_time_ms: Optional[float] = None,
     cost_breakdown: Optional[CostBreakdown] = None,
+    overhead_breakdown: Optional[dict] = None,
 ) -> SpendLogsMetadata:
     if metadata is None:
         return SpendLogsMetadata(
@@ -87,6 +88,7 @@ def _get_spend_logs_metadata(
             cold_storage_object_key=cold_storage_object_key,
             litellm_overhead_time_ms=None,
             cost_breakdown=None,
+            overhead_breakdown=None,
         )
     verbose_proxy_logger.debug(
         "getting payload for SpendLogs, available keys in metadata: "
@@ -113,6 +115,7 @@ def _get_spend_logs_metadata(
     clean_metadata["cold_storage_object_key"] = cold_storage_object_key
     clean_metadata["litellm_overhead_time_ms"] = litellm_overhead_time_ms
     clean_metadata["cost_breakdown"] = cost_breakdown
+    clean_metadata["overhead_breakdown"] = overhead_breakdown
 
     return clean_metadata
 
@@ -311,9 +314,50 @@ def get_logging_payload(  # noqa: PLR0915
 
     # Extract overhead from hidden_params if available
     litellm_overhead_time_ms = None
+    overhead_breakdown = None
     if standard_logging_payload is not None:
         hidden_params = standard_logging_payload.get("hidden_params", {})
         litellm_overhead_time_ms = hidden_params.get("litellm_overhead_time_ms")
+        
+        # Extract overhead breakdown from metadata or hidden_params
+        overhead_breakdown = (
+            standard_logging_payload.get("metadata", {}).get("overhead_breakdown")
+            or hidden_params.get("overhead_breakdown")
+        )
+        
+        # If overhead_breakdown doesn't exist, try to construct it from available data
+        if overhead_breakdown is None:
+            overhead_breakdown = {}
+            
+            # Extract cache read time from caching_details
+            caching_details = standard_logging_payload.get("metadata", {}).get("caching_details")
+            if caching_details and isinstance(caching_details, dict):
+                cache_duration_ms = caching_details.get("cache_duration_ms")
+                if cache_duration_ms is not None:
+                    overhead_breakdown["cache_read_time_ms"] = cache_duration_ms
+            
+            # Extract retry count from previous_models
+            previous_models = metadata.get("previous_models", [])
+            if previous_models and isinstance(previous_models, list):
+                overhead_breakdown["retry_count"] = len(previous_models)
+            
+            # Extract auth time from metadata if available
+            auth_time_ms = metadata.get("auth_time_ms")
+            if auth_time_ms is not None:
+                overhead_breakdown["auth_time_ms"] = auth_time_ms
+            
+            # Extract translation times from metadata if available
+            request_translation_time_ms = metadata.get("request_translation_time_ms")
+            if request_translation_time_ms is not None:
+                overhead_breakdown["request_translation_time_ms"] = request_translation_time_ms
+            
+            response_translation_time_ms = metadata.get("response_translation_time_ms")
+            if response_translation_time_ms is not None:
+                overhead_breakdown["response_translation_time_ms"] = response_translation_time_ms
+            
+            # Only set overhead_breakdown if we have at least one value
+            if not overhead_breakdown:
+                overhead_breakdown = None
 
     # clean up litellm metadata
     clean_metadata = _get_spend_logs_metadata(
@@ -366,6 +410,7 @@ def get_logging_payload(  # noqa: PLR0915
             if standard_logging_payload is not None
             else None
         ),
+        overhead_breakdown=overhead_breakdown,
     )
 
     special_usage_fields = ["completion_tokens", "prompt_tokens", "total_tokens"]

ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/LogDetailsDrawer.tsx

@@ -190,6 +190,9 @@ export function LogDetailsDrawer({
         {/* Metrics Section */}
         <MetricsSection logEntry={logEntry} metadata={metadata} />
 
+        {/* Overhead Breakdown - Show if overhead breakdown data is available */}
+        <OverheadBreakdownSection metadata={metadata} />
+
         {/* Cost Breakdown - Show if cost breakdown data is available */}
         <CostBreakdownViewer costBreakdown={metadata?.cost_breakdown} totalSpend={logEntry.spend || 0} />
 
@@ -337,6 +340,64 @@ function MetricsSection({ logEntry, metadata }: { logEntry: LogEntry; metadata:
   );
 }
 
+function OverheadBreakdownSection({ metadata }: { metadata: Record<string, any> }) {
+  const overheadBreakdown = metadata?.overhead_breakdown;
+  
+  if (!overheadBreakdown || typeof overheadBreakdown !== 'object') {
+    return null;
+  }
+
+  const hasAnyData = 
+    overheadBreakdown.auth_time_ms !== undefined ||
+    overheadBreakdown.cache_read_time_ms !== undefined ||
+    overheadBreakdown.request_translation_time_ms !== undefined ||
+    overheadBreakdown.response_translation_time_ms !== undefined ||
+    overheadBreakdown.retry_count !== undefined;
+
+  if (!hasAnyData) {
+    return null;
+  }
+
+  const formatTime = (ms: number | undefined) => {
+    if (ms === undefined || ms === null) return "-";
+    return `${ms.toFixed(2)} ms`;
+  };
+
+  return (
+    <div className="bg-white rounded-lg shadow w-full max-w-full overflow-hidden mb-6">
+      <Card title="Overhead Breakdown" size="small" bordered={false} style={{ marginBottom: 0 }}>
+        <Descriptions column={2} size="small">
+          {overheadBreakdown.auth_time_ms !== undefined && (
+            <Descriptions.Item label="Time in Auth">
+              {formatTime(overheadBreakdown.auth_time_ms)}
+            </Descriptions.Item>
+          )}
+          {overheadBreakdown.cache_read_time_ms !== undefined && (
+            <Descriptions.Item label="Time Reading Cache">
+              {formatTime(overheadBreakdown.cache_read_time_ms)}
+            </Descriptions.Item>
+          )}
+          {overheadBreakdown.request_translation_time_ms !== undefined && (
+            <Descriptions.Item label="Time in Request Translation">
+              {formatTime(overheadBreakdown.request_translation_time_ms)}
+            </Descriptions.Item>
+          )}
+          {overheadBreakdown.response_translation_time_ms !== undefined && (
+            <Descriptions.Item label="Time in Response Translation">
+              {formatTime(overheadBreakdown.response_translation_time_ms)}
+            </Descriptions.Item>
+          )}
+          {overheadBreakdown.retry_count !== undefined && overheadBreakdown.retry_count !== null && (
+            <Descriptions.Item label="Number of Retries">
+              {overheadBreakdown.retry_count}
+            </Descriptions.Item>
+          )}
+        </Descriptions>
+      </Card>
+    </div>
+  );
+}
+
 interface RequestResponseSectionProps {
   hasResponse: boolean;
   getRawRequest: () => any;