Unreliable network

A network of interconnected computers is based on some physical layer (Ethernet cables or radio in case of wireless networks)¹. The physical layer itself is subject to different disruptions, noisiness, and other problems. So we always have to assume that the network is unreliable and implement different retry, ordering, and healthcheck policies in our services.

Moreover, network capacity is limited and we should estimate ahead what an amount of data the system under development is going to send through the network.

Network throughput is an amount of data that can be transferred through a network within a certain time period Oxford Dictionary of the Internet (4 ed.). Usually, an average throughput is used: the number of the transferred bits divided by the number of passed seconds.

We can measure a network throughput in bits per second (bit/s or bps) or even in data packets per a unit of time.

Stop&Think: How do you measure a network throughput? How do you estimate a maximum throughput (also called bandwidth or channel capacity) of a network?²

Congestion control algorithms https://cloud.google.com/blog/products/networking/tcp-bbr-congestion-control-comes-to-gcp-your-internet-just-got-faster https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/net/ipv4/tcp_bbr.c

https://cacm.acm.org/magazines/2017/2/212428-bbr-congestion-based-congestion-control/fulltext

https://dl.acm.org/doi/10.1145/52324.52356

https://github.com/google/bbr

Typical interservice communication within a corporate network often causes to think that the communication is almost instanteneous. But in case of a world wide internet connection spanning several networks the time to deliver data significantly increases.

Network latency is the time taken for a given signal or data packet to propagate through a circuit or network from its source point to its destination Oxford Dictionary of Electronics and Electrical Engineering (5 ed.)

Network latency usually measured from source to destination and back constituting the so called round trip time (RTT) which is rougly speaking a double latency.³

Caching, keep-alive connections, geoghraphical proximity to the users are among usual options to decrease latency⁴.

Network partition happens if some nodes of a cluster in a distributed system cannot communicate due to a network failure but are supposed to.

Depending on the network usage profile, targets on throughput, latency, paket loss and rate a standard Linux networking stack can be partially or completely replaced with:

• a user-space networking stack (no kernel at all) like DPDK⁵ and other packet processing frameworks⁶
• a kernel executed user-defined network modules (eBPF)
• raw sockets ()

So for a distributed system it is absolutely necessary to handle different network failures (which are quite common):

• meet bandwith requirements and handle network congestion on scaling;
• deal with increased latencies;
• have policies for network partitions: if a network partition creates two and more subclusters and such subclusters can behave independently (situation known as a split-brain⁷), how should the system evolve and finally merge conflicting states?
• implement backpressure

• L1 - Physical layer (bits)
• L2 - Data link (frames)
• L3 - Network (packets)
• L4 - Transport (datagrams)
• L5 - Session (data)
• L6 - Presentation (data)
• L7 - Application (data)