Previous logger versions would always compute the mode on the Logger
handler which meant that, if any other Logger handler blocked, either because
it was waiting on IO, the system is overloaded or due to long computations, a
burst of messages could make the Logger message queue grow quite large
until it notices it should change its mode.
This commit changes the computation to use a counter which is offset by
the current message length queue. This makes the following script go from
tens of thousands messages in the Logger inbox to only hundreds (assuming
a `sync_threshold` of 100) when writing a rate limited stdio with
`mix run foo.exs | pv -p -L 1k 1>/dev/null`:
```elixir
require Logger
# monitor memory
# :observer.start()
defmodule Loop do
def run() do
Logger.info(inspect(:crypto.strong_rand_bytes(1024)))
Process.sleep(10)
run()
end
end
for _i <- 1..100 do
spawn(fn ->
Loop.run()
end)
end
defmodule OtherLoop do
def run() do
IO.inspect(:stderr, Process.info(Process.whereis(Logger), :message_queue_len), [])
IO.inspect(:stderr, :ets.tab2list(Logger.Config), [])
Process.sleep(100)
run()
end
end
OtherLoop.run()
```
José Valim
6377b324ee