Releases

v1.9.0

Improved `IO::Event::Profiler` for detecting stalls.

A new IO::Event::Profiler class has been added to help detect stalls in the event loop. The previous approach was insufficient to detect all possible stalls. This new approach uses the RUBY_EVENT_FIBER_SWITCH event to track context switching by the scheduler, and can detect stalls no matter how they occur.

profiler = IO::Event::Profiler.new

profiler.start
		
Fiber.new do
	sleep 1.0
end.transfer

profiler.stop

A default profiler is exposed using IO::Event::Profiler.default which is controlled by the following environment variables:

IO_EVENT_PROFILER=true - Enable the profiler, otherwise IO::Event::Profiler.default will return nil.
IO_EVENT_PROFILER_LOG_THRESHOLD - Specify the threshold in seconds for logging a stall. Defaults to 0.01.
IO_EVENT_PROFILER_TRACK_CALLS - Track the method call for each event, in order to log specifically which method is causing the stall. Defaults to true.

The previous environment variables IO_EVENT_SELECTOR_STALL_LOG_THRESHOLD and IO_EVENT_SELECTOR_STALL_LOG no longer have any effect.

v1.8.0

Detecting fibers that are stalling the event loop.

A new (experimental) feature for detecting fiber stalls has been added. This feature is disabled by default and can be enabled by setting the IO_EVENT_SELECTOR_STALL_LOG_THRESHOLD to true or a floating point number representing the threshold in seconds.

When enabled, the event loop will measure and profile user code when resuming a fiber. If the fiber takes too long to return back to the event loop, the event loop will log a warning message with a profile of the fiber's execution.

> cat test.rb 
#!/usr/bin/env ruby

require_relative "lib/async"

Async do
	Fiber.blocking do
		sleep 1
	end
end

> IO_EVENT_SELECTOR_STALL_LOG_THRESHOLD=true bundle exec ./test.rb
Fiber stalled for 1.003 seconds
	/home/samuel/Developer/socketry/async/test.rb:6 in '#<Class:Fiber>#blocking' (1s)
		/home/samuel/Developer/socketry/async/test.rb:7 in 'Kernel#sleep' (1s)

There is a performance overhead to this feature, so it is recommended to only enable it when debugging performance issues.

v1.7.5

Fix process_wait race condition on EPoll that could cause a hang.

2.3 KiB Raw Blame History