Spaces:
Sleeping
fix(single_brain): KI-242 — 2-attempt retry on transient Gemini errors
Browse filesZ1 investigation of intermittent single_brain failures (live HF Space logs
2026-05-15T08:15:07Z..45Z) found 100% of failures were "Gemini HTTP 503
UNAVAILABLE — high demand". Zero timeouts, zero 429s, zero JSON-parse,
zero MAX_ITERATIONS. Pure upstream provider capacity blips.
Previously a 1-second blip caused fall-through to legacy orchestrator. Now
the call retries once with 1.5s backoff on:
- HTTP 429 (rate limit)
- HTTP 500/502/503/504 (server/upstream)
- httpx.TimeoutException
- httpx.HTTPError
Permanent 4xx (400/401/403/404) still raises immediately — bad API key or
malformed body is not retryable on the same payload.
Worst-case added latency: 1.5s, well inside the 45s outer budget.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- backend/single_brain.py +88 -26
|
@@ -56,6 +56,16 @@ PER_CALL_TIMEOUT_SEC = 25.0
|
|
| 56 |
# where the LLM keeps calling save_profile_field on the same value.
|
| 57 |
MAX_ITERATIONS = 5
|
| 58 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 59 |
|
| 60 |
SYSTEM_PROMPT = """You are an Indian health-insurance advisor speaking with a customer.
|
| 61 |
|
|
@@ -377,6 +387,13 @@ async def _gemini_call(
|
|
| 377 |
) -> dict:
|
| 378 |
"""Single non-streaming Gemini generateContent call. Returns the raw
|
| 379 |
JSON payload. Raises SingleBrainError on any 4xx/5xx/transport error.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 380 |
"""
|
| 381 |
url = f"{GEMINI_BASE_URL}/{model}:generateContent?key={api_key}"
|
| 382 |
body: dict = {
|
|
@@ -397,34 +414,79 @@ async def _gemini_call(
|
|
| 397 |
pool=2.0,
|
| 398 |
)
|
| 399 |
|
| 400 |
-
|
| 401 |
-
|
| 402 |
-
|
| 403 |
-
|
| 404 |
-
|
| 405 |
-
|
| 406 |
-
|
| 407 |
-
|
| 408 |
-
|
| 409 |
-
|
| 410 |
-
|
| 411 |
-
|
| 412 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 413 |
|
| 414 |
-
if resp.status_code >= 400:
|
| 415 |
-
detail = ""
|
| 416 |
try:
|
| 417 |
-
|
| 418 |
-
except Exception:
|
| 419 |
-
|
| 420 |
-
|
| 421 |
-
|
| 422 |
-
|
| 423 |
-
|
| 424 |
-
|
| 425 |
-
|
| 426 |
-
|
| 427 |
-
raise SingleBrainError(f"Gemini malformed JSON: {e}") from e
|
| 428 |
|
| 429 |
|
| 430 |
def _extract_parts(payload: dict) -> list[dict]:
|
|
|
|
| 56 |
# where the LLM keeps calling save_profile_field on the same value.
|
| 57 |
MAX_ITERATIONS = 5
|
| 58 |
|
| 59 |
+
# Transient-error retry policy (2026-05-15 / KI-singlebrain-503).
|
| 60 |
+
# Live HF Space logs (rohitsar567/InsuranceBot, 2026-05-15 08:15Z) show
|
| 61 |
+
# Gemini intermittently returns HTTP 503 "model is currently experiencing
|
| 62 |
+
# high demand" — sometimes 3 in a row on the same session — which immediately
|
| 63 |
+
# tripped the orchestrator fallback. We retry ONCE on these transient codes
|
| 64 |
+
# with a short backoff before raising SingleBrainError so the legacy
|
| 65 |
+
# orchestrator only takes over on a genuinely sustained outage.
|
| 66 |
+
_TRANSIENT_HTTP_CODES = {429, 500, 502, 503, 504}
|
| 67 |
+
_TRANSIENT_RETRY_BACKOFF_SEC = 1.5
|
| 68 |
+
|
| 69 |
|
| 70 |
SYSTEM_PROMPT = """You are an Indian health-insurance advisor speaking with a customer.
|
| 71 |
|
|
|
|
| 387 |
) -> dict:
|
| 388 |
"""Single non-streaming Gemini generateContent call. Returns the raw
|
| 389 |
JSON payload. Raises SingleBrainError on any 4xx/5xx/transport error.
|
| 390 |
+
|
| 391 |
+
Internal retry: on transient failures (HTTP 429/5xx, httpx
|
| 392 |
+
TimeoutException, httpx.HTTPError) we retry ONCE after a short
|
| 393 |
+
backoff before raising. This soaks up the brief Gemini "high demand"
|
| 394 |
+
503 bursts observed live (2026-05-15) so we don't fall through to
|
| 395 |
+
the legacy orchestrator mid-session for what is usually a sub-second
|
| 396 |
+
blip on the provider side.
|
| 397 |
"""
|
| 398 |
url = f"{GEMINI_BASE_URL}/{model}:generateContent?key={api_key}"
|
| 399 |
body: dict = {
|
|
|
|
| 414 |
pool=2.0,
|
| 415 |
)
|
| 416 |
|
| 417 |
+
last_err: Optional[str] = None
|
| 418 |
+
last_status: Optional[int] = None
|
| 419 |
+
# 2 attempts total: initial + 1 retry on transient failure.
|
| 420 |
+
for attempt in range(2):
|
| 421 |
+
async with httpx.AsyncClient(timeout=client_timeout) as client:
|
| 422 |
+
try:
|
| 423 |
+
resp = await client.post(url, headers=headers, json=body)
|
| 424 |
+
except (asyncio.CancelledError, KeyboardInterrupt, SystemExit):
|
| 425 |
+
raise
|
| 426 |
+
except httpx.TimeoutException as e:
|
| 427 |
+
last_err = (
|
| 428 |
+
f"Gemini timeout after {timeout_sec:.1f}s (model={model})"
|
| 429 |
+
)
|
| 430 |
+
last_status = None
|
| 431 |
+
if attempt == 0:
|
| 432 |
+
_log.warning(
|
| 433 |
+
"single_brain transient timeout (attempt=1); "
|
| 434 |
+
"retrying once after %.1fs backoff",
|
| 435 |
+
_TRANSIENT_RETRY_BACKOFF_SEC,
|
| 436 |
+
)
|
| 437 |
+
await asyncio.sleep(_TRANSIENT_RETRY_BACKOFF_SEC)
|
| 438 |
+
continue
|
| 439 |
+
raise SingleBrainError(last_err) from e
|
| 440 |
+
except httpx.HTTPError as e:
|
| 441 |
+
last_err = (
|
| 442 |
+
f"Gemini transport error "
|
| 443 |
+
f"({type(e).__name__}): {str(e)[:200]}"
|
| 444 |
+
)
|
| 445 |
+
last_status = None
|
| 446 |
+
if attempt == 0:
|
| 447 |
+
_log.warning(
|
| 448 |
+
"single_brain transient transport error "
|
| 449 |
+
"(attempt=1, %s); retrying once after %.1fs backoff",
|
| 450 |
+
type(e).__name__, _TRANSIENT_RETRY_BACKOFF_SEC,
|
| 451 |
+
)
|
| 452 |
+
await asyncio.sleep(_TRANSIENT_RETRY_BACKOFF_SEC)
|
| 453 |
+
continue
|
| 454 |
+
raise SingleBrainError(last_err) from e
|
| 455 |
+
|
| 456 |
+
if resp.status_code >= 400:
|
| 457 |
+
detail = ""
|
| 458 |
+
try:
|
| 459 |
+
detail = resp.text[:500]
|
| 460 |
+
except Exception:
|
| 461 |
+
pass
|
| 462 |
+
last_status = resp.status_code
|
| 463 |
+
last_err = f"Gemini HTTP {resp.status_code}: {detail}"
|
| 464 |
+
# Transient → retry once. Permanent (4xx like 400/401/403/404) →
|
| 465 |
+
# raise immediately; retrying won't help.
|
| 466 |
+
if (
|
| 467 |
+
attempt == 0
|
| 468 |
+
and resp.status_code in _TRANSIENT_HTTP_CODES
|
| 469 |
+
):
|
| 470 |
+
_log.warning(
|
| 471 |
+
"single_brain transient HTTP %d (attempt=1); "
|
| 472 |
+
"retrying once after %.1fs backoff",
|
| 473 |
+
resp.status_code, _TRANSIENT_RETRY_BACKOFF_SEC,
|
| 474 |
+
)
|
| 475 |
+
await asyncio.sleep(_TRANSIENT_RETRY_BACKOFF_SEC)
|
| 476 |
+
continue
|
| 477 |
+
raise SingleBrainError(last_err)
|
| 478 |
|
|
|
|
|
|
|
| 479 |
try:
|
| 480 |
+
return resp.json()
|
| 481 |
+
except Exception as e: # noqa: BLE001
|
| 482 |
+
raise SingleBrainError(f"Gemini malformed JSON: {e}") from e
|
| 483 |
+
|
| 484 |
+
# Defensive — loop fell through without returning or raising. Should
|
| 485 |
+
# be unreachable, but raise so we never silently return None.
|
| 486 |
+
raise SingleBrainError(
|
| 487 |
+
last_err
|
| 488 |
+
or f"Gemini exhausted retries (last_status={last_status})"
|
| 489 |
+
)
|
|
|
|
| 490 |
|
| 491 |
|
| 492 |
def _extract_parts(payload: dict) -> list[dict]:
|