I wrote this glossary the way I wish distributed systems terms were introduced to me: not as academic vocabulary, but as words that explain late messages, duplicate messages, missing messages, and conflicting state.
The goal of this glossary is not academic purity. It is to make the words useful in design reviews, incident calls, and code discussions.
The system answers requests even when part of it is broken. Example: a product page still loads from a healthy replica.
A node gives wrong or conflicting answers instead of simply crashing. Example: one node approves a payment for one peer and rejects it for another.
During a network split, a system must choose the safer behavior for consistency or availability. Example: checkout blocks, or it accepts a possibly stale cart.
When I use these first terms in a design review, I try to make the failure visible: what keeps answering, what lies, and what the system chooses during the split.
The same state and input produce the same result everywhere. Example: replicas replay one log and compute the same balance.
Replicas may differ briefly but converge if new writes stop. Example: a profile name reaches every region after propagation.
A rising number proves which worker currently owns a lease. Example: token 42 is rejected after token 43 is issued.
Nodes spread facts by talking to a few peers at a time. Example: each cache tells two peers which nodes look alive.
A small repeated signal that says a node is still alive. Example: followers start an election when leader heartbeats stop.
I learnt to be careful with these words because they sound precise, but in production they usually describe suspicion, delay, and convergence rather than perfect knowledge.
A retry can run again without changing the result twice. Example: create order with request id 123 does not create a duplicate.
The steps a new node follows to enter a cluster safely. Example: a cache joins, receives shard ownership, then serves traffic.
Data is placed by key, often with hashing or ranges. Example: customer id 921 maps to shard 7.
A logical counter orders events without trusting wall clocks. Example: receive time 7, then move your local clock to 8.
The shared view of which nodes belong to the system. Example: rebalancing waits until the cluster agrees that node D left.
My recommendation is to connect these middle terms to one running system, such as checkout or cache routing, because the vocabulary sticks when it explains a real operational choice.
Healthy nodes cannot talk across a network split. Example: two availability zones keep running but cannot exchange messages.
A rule for deciding which event is applied first. Example: every replica applies debit before credit.
The system behaves deliberately when messages are delayed or dropped. Example: an isolated region serves reads but rejects risky writes.
Enough replicas agree for an operation to count. Example: a write waits for 2 of 3 replicas.
A leader follower method for agreeing on a replicated log. Example: the leader appends a config change and followers acknowledge it.
The way I apply these terms is to ask who is allowed to decide, how many participants must agree, and what happens when the answer arrives late.
Concurrent transactions behave as if they ran one at a time in a clear order. Example: two withdrawals from the same account cannot both spend the same remaining balance.
A coordinator asks participants to prepare, then tells them to commit. Example: payment and ledger both prepare before final commit.
A write spreads from one copy to other copies. Example: a password change reaches regional read models.
Several counters record which node saw which version. Example: edits from phone and laptop conflict when neither version includes the other.
A system records intent before applying a change. Example: a broker writes the message to disk before acknowledging it.
A globally unique transaction id names one distributed transaction across every system that participates in it. Example: order, payment, and ledger records all store XID 891 so operators can prove they belong to the same transaction.
A benchmark used to compare key value and cloud data stores. Example: run read heavy workload B before changing quorum settings.
A coordination service used for leader election, locks, and configuration. Example: services watch a path to learn the current leader.