Enable log buffering

[steve-chavez]

Problem

Currently there's no log buffering, as noted in docs. This means that log levels higher than error cannot be enabled without reduced RPS.

Additionally, this might also cause other issues with the default log-level=error. For example on pool timeouts, the RPS will also suffer.

Solution

Enable log buffering, currently it's all set to line buffering on:

postgrest/main/Main.hs

Lines 15 to 21 in bf79766

	setBuffering :: IO ()
	setBuffering = do
	-- LineBuffering: the entire output buffer is flushed whenever a newline is
	-- output, the buffer overflows, a hFlush is issued or the handle is closed
	hSetBuffering stdout LineBuffering
	hSetBuffering stdin LineBuffering
	hSetBuffering stderr LineBuffering

This concern should be centralized on the Logger.hs module and possibly receive a config option as input.

Notes

The logger middleware obfuscates this logic. Ideally we should reuse our Observation module as mentioned on:

postgrest/src/PostgREST/Logger.hs

Lines 58 to 70 in bf79766

	-- TODO stop using this middleware to reuse the same "observer" pattern for all our logs
	middleware :: LogLevel -> (Wai.Request -> Maybe BS.ByteString) -> Wai.Middleware
	middleware logLevel getAuthRole =
	unsafePerformIO $
	Wai.mkRequestLogger Wai.defaultRequestLoggerSettings
	{ Wai.outputFormat =
	Wai.ApacheWithSettings $
	Wai.defaultApacheSettings &
	Wai.setApacheRequestFilter (\_ res -> shouldLogResponse logLevel $ Wai.responseStatus res) &
	Wai.setApacheUserGetter getAuthRole
	, Wai.autoFlush = True
	, Wai.destination = Wai.Handle stdout
	}

[steve-chavez]

One way to test buffering is working as intended would be to enable log-level=info for the loadtests, in that case throughput should be more or less maintained.