A SIGTERM during graceful shutdown could close the OTLP exporter before the final flush completed, truncating up to two seconds of in-flight spans. Symptom: the last request of a rolling deploy would appear to have no downstream spans, making it look like the new pod was failing silently when in fact the span data had simply been dropped.

The shutdown path now waits on the flush completion (capped at the configured exporter timeout) before closing the transport. We backfilled the missing spans for affected tenants where the buffered data was still within the 7-day retention window; anything older is unrecoverable.

Thanks to the customer who filed the reproducer — the bug had been latent since v0.5 and was invisible without a tightly-timed rolling restart.