Skip to content- Sources: primary, discussion
- Summary: After two months of patching and a 1.5 TiB sparse core dump, the author found the hung renderer held 942 shared-memory regions, 938 of them containing the same 27-byte reply to a keepalive presence POST that the app never reads. Blink skips draining a response body into memory when it carries Cache-Control: no-store, so each unread body keeps its Mojo data pipe and fd open until the renderer hits its 1,024 soft descriptor limit. A repro showed the descriptors stay open even after the Response objects are garbage collected, which ruled out the bundled Datadog SDK, and Chromium told the author this works as intended.
- Why it matters: Fire-and-forget fetch calls in Electron or Chromium apps should read or cancel the body, because unread no-store responses leak descriptors until the renderer freezes.
- Background: Chromium hands every response body to the renderer through a Mojo data pipe backed by a shared-memory region, and each one costs a file descriptor in the renderer process. The post also explains why the dumps are so large: V8 reserves a 1 TiB sandbox plus guard regions at startup, and the renderer's seccomp policy blocks the madvise call that would exclude it from core dumps.