fix(gcp_pubsub source): recover from stalled streams and reduce error noise

changelog.d/19418_gcp_pubsub_idle_reconnect.fix.md

@@ -0,0 +1,9 @@
+The `gcp_pubsub` source no longer floods the logs with misleading errors on idle, low-volume, or bursty subscriptions. GCP routinely ends a streaming pull with a transient error (such as an `Unavailable` status when there are simply no messages to deliver), which Vector retries automatically. These transient failures are now logged at `debug`, and a `failed_fetching_events` component error is only emitted once several failures occur in a row without any successful fetch in between. The threshold is configurable via the new `max_retry_errors` option (default `5`); a successful fetch resets the counter.
+
+The source is also more resilient to connections that silently break and stop delivering messages until Vector is restarted:
+
+- HTTP/2 keepalive pings are sent on the gRPC connection so a dead-but-open connection is detected and torn down instead of hanging. The ping interval and timeout are configurable via the new `keepalive` option.
+- The stream keepalive uses a non-blocking send, so a backed-up request channel can no longer deadlock the task.
+- A new `idle_timeout_secs` option (defaulting to `900`) restarts the stream if no activity is seen on an active connection within that window.
+
+authors: SamyDjemai

lib/vector-config/src/external/serde_with.rs

@@ -143,3 +143,12 @@ impl Configurable for Option<serde_with::DurationMilliSeconds<u64, serde_with::f
         generate_optional_schema(&u64::as_configurable_ref(), generator)
     }
 }
+
+impl Configurable for Option<serde_with::DurationSecondsWithFrac<f64, serde_with::formats::Strict>> {
+    fn generate_schema(generator: &RefCell<SchemaGenerator>) -> Result<SchemaObject, GenerateError>
+    where
+        Self: Sized,
+    {
+        generate_optional_schema(&f64::as_configurable_ref(), generator)
+    }
+}