fix(tickerr): fix TypeError, error-type fallback, and thread cap

- _latency_ms(): handle both datetime objects and floats — LiteLLM
  passes datetime, not float, causing TypeError on every invocation
- _ERROR_TYPE_MAP fallback: removed default "overloaded" for unmapped
  codes (400, 404, 502, etc.) — now returns None for unrecognised codes
- _fire_and_forget(): added semaphore cap (_MAX_INFLIGHT=5) to prevent
  unbounded thread spawning during burst failures
- Removed dead is_resolution parameter from _report() — only called
  from failure hooks so it was always False

litellm/integrations/tickerr.py

@@ -18,40 +18,46 @@ from __future__ import annotations
 import os
 import re
 import threading
-from typing import Any, Dict, Optional
+from datetime import datetime
+from typing import Any, Dict, Optional, Union
 
 from litellm.integrations.custom_logger import CustomLogger
 
 _REPORT_URL = "https://tickerr.ai/api/v1/report"
 _UA = "litellm-tickerr/1.0"
 
+# Cap concurrent in-flight reports to avoid thread exhaustion on burst failures
+_MAX_INFLIGHT = 5
+_inflight = threading.Semaphore(_MAX_INFLIGHT)
+
 # Map litellm custom_llm_provider → Tickerr provider slug
 _PROVIDER_MAP: Dict[str, str] = {
-    "openai":           "openai",
-    "anthropic":        "anthropic",
-    "google":           "google",
-    "vertex_ai":        "google",
-    "gemini":           "google",
-    "cohere":           "cohere",
-    "mistral":          "mistral",
-    "groq":             "groq",
-    "together_ai":      "together",
-    "huggingface":      "huggingface",
-    "replicate":        "replicate",
-    "deepinfra":        "deepinfra",
-    "perplexity":       "perplexity",
-    "fireworks_ai":     "fireworks",
-    "openrouter":       "openrouter",
-    "azure":            "azure",
-    "bedrock":          "aws",
-    "ai21":             "ai21",
-    "cerebras":         "cerebras",
-    "xai":              "xai",
-    "deepseek":         "deepseek",
-    "ollama":           "ollama",
-    "nlp_cloud":        "nlp_cloud",
+    "openai":       "openai",
+    "anthropic":    "anthropic",
+    "google":       "google",
+    "vertex_ai":    "google",
+    "gemini":       "google",
+    "cohere":       "cohere",
+    "mistral":      "mistral",
+    "groq":         "groq",
+    "together_ai":  "together",
+    "huggingface":  "huggingface",
+    "replicate":    "replicate",
+    "deepinfra":    "deepinfra",
+    "perplexity":   "perplexity",
+    "fireworks_ai": "fireworks",
+    "openrouter":   "openrouter",
+    "azure":        "azure",
+    "bedrock":      "aws",
+    "ai21":         "ai21",
+    "cerebras":     "cerebras",
+    "xai":          "xai",
+    "deepseek":     "deepseek",
+    "ollama":       "ollama",
+    "nlp_cloud":    "nlp_cloud",
 }
 
+# Only map codes that unambiguously indicate the specific error type
 _ERROR_TYPE_MAP: Dict[int, str] = {
     429: "rate_limit",
     529: "overloaded",
@@ -105,8 +111,23 @@ def _extract_status_code(exception: Optional[BaseException]) -> Optional[int]:
     return None
 
 
+def _latency_ms(
+    start_time: Union[datetime, float], end_time: Union[datetime, float]
+) -> int:
+    """Compute elapsed ms whether LiteLLM passes datetime objects or floats."""
+    if isinstance(start_time, datetime) and isinstance(end_time, datetime):
+        return round((end_time - start_time).total_seconds() * 1000)
+    return round((float(end_time) - float(start_time)) * 1000)  # type: ignore[arg-type]
+
+
 def _fire_and_forget(payload: Dict[str, Any]) -> None:
-    """POST to Tickerr in a daemon thread — never blocks the caller."""
+    """POST to Tickerr in a daemon thread — never blocks the caller.
+
+    A semaphore caps concurrent in-flight threads so a burst of failures
+    (e.g. 100 errors/s) cannot exhaust the thread pool.
+    """
+    if not _inflight.acquire(blocking=False):
+        return  # already at max concurrent reports — drop silently
 
     def _send() -> None:
         try:
@@ -124,6 +145,8 @@ def _fire_and_forget(payload: Dict[str, Any]) -> None:
                 pass
         except Exception:
             pass  # never crash the caller
+        finally:
+            _inflight.release()
 
     t = threading.Thread(target=_send, daemon=True)
     t.start()
@@ -152,10 +175,10 @@ class TickerrLogger(CustomLogger):
         self,
         kwargs: Dict[str, Any],
         response_obj: Any,
-        start_time: float,
-        end_time: float,
+        start_time: Union[datetime, float],
+        end_time: Union[datetime, float],
     ) -> None:
-        self._report(kwargs, start_time, end_time, is_resolution=False)
+        self._report(kwargs, start_time, end_time)
 
     # ── async ─────────────────────────────────────────────────────────────────
 
@@ -163,27 +186,26 @@ class TickerrLogger(CustomLogger):
         self,
         kwargs: Dict[str, Any],
         response_obj: Any,
-        start_time: float,
-        end_time: float,
+        start_time: Union[datetime, float],
+        end_time: Union[datetime, float],
     ) -> None:
-        self._report(kwargs, start_time, end_time, is_resolution=False)
+        self._report(kwargs, start_time, end_time)
 
     # ── internal ──────────────────────────────────────────────────────────────
 
     def _report(
         self,
         kwargs: Dict[str, Any],
-        start_time: float,
-        end_time: float,
-        is_resolution: bool,
+        start_time: Union[datetime, float],
+        end_time: Union[datetime, float],
     ) -> None:
         model: str = kwargs.get("model", "") or ""
         exception: Optional[BaseException] = kwargs.get("exception")
-        latency_ms = round((end_time - start_time) * 1000)
 
         provider = _normalize_provider(model, kwargs)
         status_code = _extract_status_code(exception)
-        error_type = _ERROR_TYPE_MAP.get(status_code, "overloaded") if status_code else None
+        # Only set error_type for codes we can classify with confidence
+        error_type: Optional[str] = _ERROR_TYPE_MAP.get(status_code) if status_code is not None else None
 
         # Strip provider prefix: "anthropic/claude-3-5-haiku" → "claude-3-5-haiku"
         model_clean = model.split("/", 1)[-1] if "/" in model else model
@@ -191,8 +213,7 @@ class TickerrLogger(CustomLogger):
         payload: Dict[str, Any] = {
             "provider": provider,
             "model": model_clean or None,
-            "is_resolution": is_resolution,
-            "latency_ms": latency_ms,
+            "latency_ms": _latency_ms(start_time, end_time),
         }
         if status_code is not None:
             payload["error_code"] = status_code