verdie-g / datadogstatsd Goto Github PK

When using unix domain socket, if the connection to the agent is lost, an exception is thrown and no more metrics will be sent. This is only for unix socket as udp is connectionless.

https://docs.datadoghq.com/getting_started/tagging/unified_service_tagging

If the queue is full a metric will be sent when metrics are discarded but this metric can be discarded too if the queue is still full. A second channel, unbounded, could be used for priority metrics.

Since the socket connection is always on the same machine, writes only block if the buffer is full so async might not be useful here.

• The cost of the async state machine by RavenDB
• Async performance by our lord and savior Stephen Toub

https://docs.datadoghq.com/developers/dogstatsd/?tab=kubernetes

In this example we send a metric every ~50 ms which create a constant rate of 16 increments by seconds. However, the client and agent flush windows almost overlaps and sometimes the client is flushed twice during an agent window, doubling the number of increments, and then flushed 0 times during the next windows.

Max buffering time and metric tick interval need to be tweaked.

Once DataDog/datadog-agent#6506 is merged and available.

Allow to aggregate distributions and histograms.

NaN should throw ArgumentException("Function does not accept floating point Not-a-Number values").

https://github.com/DataDog/datadogpy/blob/058114cc3d65483466684c96a5c23e36c3aa052e/datadog/dogstatsd/base.py#L461-L462