In https://codereview.chromium.org/2538833004, a 5-second delay was added to the tearing down process, to better amortize the cost of launching a sandboxed utility process. However, it only covers 5 more seconds of request, but not extending it further even if there are requests coming in.

For example, at t = 0.5, the util process became idle, and tearing down was scheduled on t = 5.5. Assume the requests keep coming every second, so at t = 1, 2, 3, 4, 5, the requests can reuse the same process. However, at t = 6, that process has been killed, and we need to launch a new sandboxed process for this request. In the steady state, we kill and re-spawn it around every 5 seconds.

This issue was discovered by analyzing the image decoding in NTP feeds, where a suatained period of time of spotty image decoding happens.

To better debounce process launching and killing. It might make more sense to keep it alive as long as there haven't been a 5-second period of idling. My main concern is that any possible memory leaks within data decoder would not be automatically fixed by periodical killing, but I guess over-killing might also hinder the opportunity to discover such memory leaks.

Thoughts?