If there are buffers anywhere in the network you can get unpredictably long latency - your packet happened to arrive at the back of the queue. Adding longer buffers is a common strategy to solve networking problems - if the buffer is long enough, the message will be sent one day - but it's also a mistaken one as the buffer can fill up and stay full, creating a permanent and large delay. IETF's Dave Taht spent a lot of time diagnosing this so-called bufferbloat in WiFi networks and the wider Internet.
The solution he found was to reduce the effective size of the buffer by dropping messages at random when the buffer length went over a target; if you're using a transport layer like TCP/IP that retransmits dropped messages and adjusts the transmission rate, this ensures that senders know that there is a delay and slow down sending so the buffer clears, and the dropped messages get resent. Unlike just having a shorter buffer, this doesn't create a systematic source of packet loss so it is fair to all flows.