Febontheweb

Change is inevitable, especially so in applications that ebb and flow with system changes and user requirements. Generally, changes are codified in the form of an event; an event like a ddos attack for instance triggers a reaction or a response from one or more nodes, which ultimately mutates the overall state of the universe in a system. Undoubtedly, events rarely happen as one offs. At any given time, a sequence of events flows through the system. Whether it be a cascade of related events or multiple isolated events, the system is consistently responding and adapting. ...

In a distributed system, failures are to be expected. Whether it be a fault of the network or the node, failure is disruptive to the entire system. Failed processes struggle to recover and receive updates. Healthy nodes that try to connect to a failed node might fail completely or receive stale information, which can jeopardize the integrity and overall (C)onsistency of the system. To mitigate this, nodes have a failure mode, which it codifies in the form of a membership protocol. ...

In a distributed system, communication is crucial to ensuring no node becomes stale. Updates can happen in 2 main ways, through consensus, or through gossip. In a consensus model, nodes attempt to reach quorum by mutually agreeing on the order of updates or periodically voting on and rotating the leader in charge of distributing updates to fellow nodes—I covered the ins and outs of this in an earlier post, if you’re curious to learn more about it. Gossip is a concept analogous to consensus. It similarly keeps nodes updated except it achieves this via epidemic theory; If one node is infected, the entire population of nodes will eventually become infected. ...

In an earlier post, I glossed over a concept that forms a fundamental tenet of systems design; the CAP theorem. To reiterate, the CAP theorem states that of 3 desirable traits in a distributed system, Consistency, Availability and Partition tolerance, only 2 can be prioritized at any given time. It highlights the tradeoffs we make when we design for a distributed setting. The CAP theorem was birthed from the move away from vertical scaling where we simply sized up machines for greater processing capacity to horizontal scaling where machines work in a distributed and parallel fashion. The CAP theorem specifically focuses on databases, since it highlights the difficulty of maintaining state when reads and writes can originate from anywhere and can easily be disrupted in the event of a network failure. ...

In a distributed setup where nodes share the load of processing requests and storing/retrieving data, reaching consensus is crucial to ensuring consistency in an application. To return to the concert ticket sales example from the last post, we wouldn’t want 2 users buying the same seat since that violates the condition of assigned seating. Consensus can then be described as ensuring all nodes agree on a specific value or a state in spite of a potential failure or delay; there lies the rub. ...

Demystifying Distributed Systems: the what and why If you deploy an app today, dealing with a distributed system is an inescapable problem. Whether it’s understanding how to balance requests, or synchronizing data replication, at least some aspect of your deployment will require orchestrating and coordinating multiple servers to operate in unison. The goal of “distributedness” here is to have things work across regions and systems in a way that is completely imperceptible to a user. ...