report key and model queue saturation separately

This commit is contained in:
wlx committed 2026-09-27 03:17:44 +08:00
1 parent 16981fd307
commit b54b6a8ac4
2 files changed
+5 -2

No files matched your search

@@ -125,11 +125,14 @@ public class GatewayGuard {
String token = UUID.randomUUID().toString();
long waitMs = Math.max(0, Math.min(queueWaitMs, requestedWaitMs > 0 ? requestedWaitMs : queueWaitMs));
long deadline = System.nanoTime() + java.util.concurrent.TimeUnit.MILLISECONDS.toNanos(waitMs);
long lastBusyCode = 0;
while (true) {
long acquired = executePair(modelBucket, keyBucket, Integer.toString(modelConcurrency(model)), Integer.toString(defaultKeyConcurrency), Long.toString(SLOT_TTL_MS), token);
if (acquired == 1) return new Admission(modelBucket, keyBucket, token);
lastBusyCode = acquired;
if (waitMs == 0 || System.nanoTime() >= deadline) {
throw new GatewayLimitException("MODEL_BUSY", "当前模型或 API Key 并发已满,已排队等待但仍未获得资源", 429, Math.max(1, (int) Math.ceil(queuePollMs / 1000.0)));
boolean keyBusy = lastBusyCode == 2;
throw new GatewayLimitException(keyBusy ? "KEY_BUSY" : "MODEL_BUSY", keyBusy ? "当前 API Key 并发已满,已排队等待但仍未获得资源" : "当前模型分组并发已满,已排队等待但仍未获得资源", 429, Math.max(1, (int) Math.ceil(queuePollMs / 1000.0)));
}
try {
long remaining = java.util.concurrent.TimeUnit.NANOSECONDS.toMillis(deadline - System.nanoTime());