fix(webapp): cap AI SDK OTel attribute size so ClickHouse JSON parse doesn't drop spans

Vercel AI SDK spans carry tens of KB of prompt/response content per attribute,
producing an assembled attributes JSON that ClickHouse rejects with
"Cannot parse JSON object" and silently drops the whole batch.

- Add otlpAttributeLimits module with per-key overrides for ai.* / gen_ai.*
  content keys (tighter 1KB cap) plus a 32KB total-attributes backstop that
  drops AI content keys in priority order when exceeded.
- Wire SERVER_OTEL_SPAN_ATTRIBUTE_VALUE_LENGTH_LIMIT /
  SERVER_OTEL_AI_CONTENT_ATTRIBUTE_VALUE_LENGTH_LIMIT /
  SERVER_OTEL_SPAN_TOTAL_ATTRIBUTES_LENGTH_LIMIT env vars through the OTLP
  exporter for spans and logs.
- DynamicFlushScheduler now recognises ClickHouse JSON parse errors and
  recursively splits the failing batch (up to depth 4) to isolate the bad
  row instead of poisoning the whole 5-10k-row batch, with a droppedRows
  metric for rows that can't be isolated further.

Author: 0ski

.server-changes/otel-ai-sdk-attribute-truncation.md

@@ -0,0 +1,20 @@
+---
+area: webapp
+type: fix
+---
+
+Tighten OTel span attribute truncation for Vercel AI SDK content keys
+(`ai.prompt*`, `ai.response.text/object/toolCalls/reasoning*`,
+`gen_ai.prompt`, `gen_ai.completion`, `gen_ai.request.messages`,
+`gen_ai.response.text`) to a 1KB per-attribute cap, plus a 32KB per-span
+backstop that drops these content keys in priority order if the assembled
+attributes JSON still exceeds it. Cost/token metadata (`ai.usage.*`,
+`ai.model.*`, `gen_ai.usage.*`, `gen_ai.response.model`, etc.) keeps the
+default 8KB cap so LLM enrichment continues to work.
+
+Adds a parse-error-aware safety net in `DynamicFlushScheduler`: when
+ClickHouse rejects a batch with `Cannot parse JSON object here`, the
+batch is split in half and retried (up to 4 split levels / 16-way
+isolation) instead of failing all 5–10k rows at once. Singleton rows
+that still fail are logged with a 1KB sample and dropped so the rest of
+the queue keeps flowing.

apps/webapp/app/env.server.ts

@@ -498,6 +498,17 @@ const EnvironmentSchema = z
     TRIGGER_OTEL_ATTRIBUTE_PER_LINK_COUNT_LIMIT: z.string().default("10"),
     TRIGGER_OTEL_ATTRIBUTE_PER_EVENT_COUNT_LIMIT: z.string().default("10"),
 
+    // Server-side OTel ingestion limits applied in otlpExporter.server.ts.
+    // Default per-attribute cap (8KB) is enough for nearly all keys, but
+    // Vercel AI SDK content keys (ai.prompt*, ai.response.text/object/etc.,
+    // gen_ai.prompt, gen_ai.completion) carry tens of KB and have a tighter
+    // dedicated cap. The total cap is a backstop applied to the assembled
+    // attributes JSON; if exceeded, AI content keys are dropped in priority
+    // order. Both prevent oversized JSON from breaking ClickHouse inserts.
+    SERVER_OTEL_SPAN_ATTRIBUTE_VALUE_LENGTH_LIMIT: z.coerce.number().int().default(8192),
+    SERVER_OTEL_AI_CONTENT_ATTRIBUTE_VALUE_LENGTH_LIMIT: z.coerce.number().int().default(1024),
+    SERVER_OTEL_SPAN_TOTAL_ATTRIBUTES_LENGTH_LIMIT: z.coerce.number().int().default(32768),
+
     CHECKPOINT_THRESHOLD_IN_MS: z.coerce.number().int().default(30000),
 
     // Internal OTEL environment variables

apps/webapp/app/v3/dynamicFlushScheduler.server.ts

@@ -18,6 +18,19 @@ export type DynamicFlushSchedulerConfig<T> = {
   isDroppableEvent?: (item: T) => boolean; // Function to determine if an event can be dropped
 };
 
+// Bound on the recursive batch-split safety net. 4 → up to 16-way split,
+// which isolates a single bad row inside a 16k-element batch.
+const MAX_SPLIT_DEPTH = 4;
+
+function isClickHouseJsonParseError(error: unknown): boolean {
+  if (!error) return false;
+  const message =
+    typeof error === "object" && error !== null && "message" in error
+      ? String((error as { message?: unknown }).message ?? "")
+      : String(error);
+  return message.includes("Cannot parse JSON object");
+}
+
 export class DynamicFlushScheduler<T> {
   private batchQueue: T[][];
   private currentBatch: T[];
@@ -43,6 +56,10 @@ export class DynamicFlushScheduler<T> {
     totalItemsFlushed: 0,
     droppedEvents: 0,
     droppedEventsByKind: new Map<string, number>(),
+    // Rows dropped at flush time because ClickHouse rejected them and the
+    // batch-split safety net couldn't isolate them further. Distinct from
+    // `droppedEvents`, which counts pre-batch load-shedding drops.
+    droppedRows: 0,
   };
   private isShuttingDown: boolean = false;
 
@@ -196,40 +213,112 @@ export class DynamicFlushScheduler<T> {
     // Schedule all batches for concurrent processing
     const flushPromises = batchesToFlush.map((batch) =>
       this.limiter(async () => {
-        const itemCount = batch.length;
-
         const self = this;
 
-        async function tryFlush(flushId: string, batchToFlush: T[], attempt: number = 1) {
+        async function tryFlush(
+          flushId: string,
+          batchToFlush: T[],
+          attempt: number = 1,
+          splitDepth: number = 0
+        ) {
+          const subBatchSize = batchToFlush.length;
+
           try {
             const startTime = Date.now();
             await self.callback(flushId, batchToFlush);
 
             const duration = Date.now() - startTime;
-            self.totalQueuedItems -= itemCount;
+            self.totalQueuedItems -= subBatchSize;
             self.consecutiveFlushFailures = 0;
             self.lastFlushTime = Date.now();
             self.metrics.flushedBatches++;
-            self.metrics.totalItemsFlushed += itemCount;
+            self.metrics.totalItemsFlushed += subBatchSize;
 
             self.logger.debug("Batch flushed successfully", {
               flushId,
-              itemCount,
+              itemCount: subBatchSize,
               duration,
               remainingQueueDepth: self.totalQueuedItems,
               activeConcurrency: self.limiter.activeCount,
               pendingConcurrency: self.limiter.pendingCount,
             });
           } catch (error) {
+            // ClickHouse rejects an entire batch when a single row's
+            // attributes JSON is unparseable. Retrying the same batch will
+            // just fail again, so split-and-retry isolates the offender
+            // instead of poisoning the whole 5–10k-row batch.
+            const isParseError = isClickHouseJsonParseError(error);
+
+            if (isParseError && subBatchSize > 1 && splitDepth < MAX_SPLIT_DEPTH) {
+              const mid = Math.floor(subBatchSize / 2);
+              const left = batchToFlush.slice(0, mid);
+              const right = batchToFlush.slice(mid);
+
+              self.logger.warn(
+                "Splitting OTel batch after ClickHouse JSON parse failure",
+                {
+                  flushId,
+                  itemCount: subBatchSize,
+                  splitDepth,
+                  leftSize: left.length,
+                  rightSize: right.length,
+                }
+              );
+
+              // Run halves concurrently and tolerate independent failures —
+              // a rejection from one half must not prevent the other half
+              // from completing. Each leaf's tryFlush updates totalQueuedItems
+              // and metrics on its own success/drop paths.
+              const results = await Promise.allSettled([
+                tryFlush(flushId + "-L", left, 1, splitDepth + 1),
+                tryFlush(flushId + "-R", right, 1, splitDepth + 1),
+              ]);
+
+              for (const [index, result] of results.entries()) {
+                if (result.status === "rejected") {
+                  self.metrics.failedBatches++;
+                  self.logger.error(
+                    "Split half failed after exhausting retries",
+                    {
+                      flushId: flushId + (index === 0 ? "-L" : "-R"),
+                      error: result.reason,
+                      splitDepth: splitDepth + 1,
+                    }
+                  );
+                }
+              }
+              return;
+            }
+
+            if (isParseError && subBatchSize === 1) {
+              // Singleton row that ClickHouse still rejects. Drop it so
+              // the rest of the queue keeps flowing, and log enough of
+              // the offending event for someone to investigate later
+              // without dumping multi-KB of attributes into the log.
+              self.metrics.droppedRows += 1;
+              self.metrics.failedBatches++;
+              self.totalQueuedItems -= 1;
+              self.logger.error(
+                "Dropping single OTel row rejected by ClickHouse JSON parser",
+                {
+                  flushId,
+                  sample: JSON.stringify(batchToFlush[0]).slice(0, 1024),
+                  splitDepth,
+                }
+              );
+              return;
+            }
+
             self.consecutiveFlushFailures++;
             self.metrics.failedBatches++;
 
             self.logger.error("Error attempting to flush batch", {
               flushId,
-              itemCount,
+              itemCount: subBatchSize,
               error,
               consecutiveFailures: self.consecutiveFlushFailures,
               attempt,
+              splitDepth,
             });
 
             // Back off on failures
@@ -239,7 +328,7 @@ export class DynamicFlushScheduler<T> {
 
             if (attempt <= 3) {
               await new Promise((resolve) => setTimeout(resolve, 500));
-              return await tryFlush(flushId, batchToFlush, attempt + 1);
+              return await tryFlush(flushId, batchToFlush, attempt + 1, splitDepth);
             } else {
               throw error;
             }