What we do is to periodically turn on profiling for X seconds, collect some
data and turn it off again. We do it at every once a while periodically. We
target a single or a group of instances in a large group.
On Mon, Jul 24, 2017 at 6:11 PM, Dave Cheney wrote:
> Another option
Another option is to profile a % of requests. In the past I've done that by
enabling profiling on a set % of application servers then extrapolating
from there.
On Tuesday, 25 July 2017 10:55:42 UTC+10, Jaana Burcu Dogan wrote:
>
> It would be very speculative to provide reference numbers
Here's an entirely unscientific method to determine the overhead of
profiling. The Go distribution contains a set of basic benchmarks, one of
which is a loopback based http client server benchmark. Running the
benchmark with and without profiling gives a rough ballpark for the
overhead of