A few comments

Author: ericallam

apps/webapp/app/v3/eventRepository/clickhouseEventRepository.server.ts

@@ -1379,6 +1379,11 @@ export class ClickhouseEventRepository implements IEventRepository {
         }
       }
 
+      // Parse attributes from the first record that has them, then re-parse for the
+      // completed SPAN record. The completed record's attributes are a superset of the
+      // partial's (includes enriched trigger.llm.* cost data added during ingestion).
+      // This means at most 2x JSON.parse per span detail query, but only on this
+      // read path (span detail view), not on ingestion.
       if (typeof record.attributes_text === "string") {
         const shouldUpdate =
           span.properties == null ||

internal-packages/llm-pricing/src/registry.ts

@@ -22,6 +22,9 @@ function compilePattern(pattern: string): RegExp {
 export class ModelPricingRegistry {
   private _prisma: PrismaClient | PrismaReplicaClient;
   private _patterns: CompiledPattern[] = [];
+  // TODO: When we add project-based models (users adding their own), this cache grows unbounded
+  // between reloads. Fine-tuned model IDs (e.g. "ft:gpt-3.5-turbo:org:name:id") create unique
+  // entries per model string. Consider adding an LRU cap or size limit at that point.
   private _exactMatchCache: Map<string, LlmModelWithPricing | null> = new Map();
   private _loaded = false;
   private _readyResolve!: () => void;