The network is reliable#

The abstractions developers learn from various technologies and protocols often enforce this common fallacy. As we will see in a later chapter, networking protocols like TCP can make us believe that the network is reliable and never fails. However, this is just an illusion with significant repercussions. Also, we build network connections on top of hardware that will also fail at some point. Hence, we should design our systems accordingly.

Latency is zero

Libraries that attempt to model remote procedure calls as local calls, such as gRPC or Thrift, enforce this assumption. We should always remember that there’s a large difference (from milliseconds to nanoseconds) in latency between a call to a remote system and that to local memory access. This gets even worse when we consider calls between data centers on different continents. Thus, this is another thing to keep in mind when deciding how to geo-distribute our system.

Bandwidth is infinite

This fallacy is weaker nowadays. This is because the bandwidth we can achieve has significantly improved in the last few decades. For instance, we can now build high-bandwidth connections in our own data centers. However, this does not mean we can use all of it if our traffic needs to cross the Internet. This is important to consider when we make decisions about our distributed system’s topology, and when requests travel through the Internet.

The network is secure

This fallacy shows that the wider network used by two nodes to communicate is not necessarily under their control. Thus, we should consider it insecure.

The course dedicates a portion to security, where it explains the various techniques we can use to securely utilize an insecure network.

Topology doesn’t change

Network also comprises many different parts that different organizations may manage with different hardware. Moreover, failures in some parts of this network may require us to change its topology to keep it functional. This also highlights the other two fallacies i.e there is one administrator and the network is homogeneous

Transport cost is zero

The transportation of data between two points incurs financial costs. We should factor this in when we build a distributed system.

The global clock fallacy