Every recurring term, in plain English. 36 entries. Click any for a fuller explanation and where it shows up.
Two-Phase Commit: distributed transaction protocol. Strong but blocking: failures hang.
Available under partition (Cassandra default, Dynamo). Sacrifices strong consistency.
Front-door service for auth, rate limiting, routing.
Signal from a slow consumer telling producers to slow down (or buffer).
A probabilistic set that says 'maybe' or 'definitely no' to membership queries.
Read from cache; on miss, read from source and populate the cache.
The most common caching pattern. Other patterns: write-through, write-behind, read-through.
Consistency, Availability, Partition tolerance: pick two during a partition.
A network of edge servers caching static content near the user.
Stops calling a failing dependency for a cool-down window to avoid cascading failure.
Distribute keys across N nodes so adding/removing a node only moves ~1/N of keys.
Consistent under partition (etcd, ZooKeeper). Sacrifices availability when the leader is unreachable.
Splits an object into K data fragments + M parity fragments; any K can reconstruct.
Used by S3 and similar to get high durability at much lower storage cost than full replication.
Extract → Transform → Load. The data-pipeline pattern moving data from sources to a warehouse.
Replicas may temporarily disagree but will converge if writes stop.
One write triggers many downstream writes (e.g. posting to all followers' caches).
Trade-off: cheap reads, expensive writes, and writes can spike with celebrities.
Compute the result at read time by querying each source (pure pull).
Avoids write storms but every read does more work. Twitter uses a hybrid.
Push new items into every consumer's data at write time (Instagram-style).
Makes reads O(1) at the cost of huge write storms for popular producers.
An encoding that turns a lat/lng into a string prefix, so nearby points share prefixes.
Lets you shard or index geographically with a normal string-prefix lookup.
Periodic 'I'm alive' messages so peers detect failure quickly.
A single key receiving wildly disproportionate traffic, saturating its shard.
An operation that can be safely repeated without changing the result.
Critical for retries. 'Set balance to $100' is idempotent; 'Add $10' is not.
In-Sync Replicas: the set of follower brokers fully caught up with the leader.
Only ISR members are eligible to take over on leader failure (in Kafka).
A consensus algorithm (Raft, Paxos) picks one node as the writer/coordinator.
Spreads requests across a pool of identical backends.
Client opens a request and the server holds it open until there's data, then responds.
Queries per second, the standard unit for measuring request rate.
A majority of replicas must agree before a write or read is committed.
Keeping copies of data on multiple nodes for durability or read scaling.
Multi-step transaction split into compensating steps so each can roll back independently.
Splitting data across multiple stores so each holds a slice of the keyspace.
Server-Sent Events: one-way streaming over a long-lived HTTP connection.
Every read sees the most recent write. Costs latency or availability under partition.
Decrypting HTTPS at the edge so backends speak plain HTTP internally.
Append-only log of intended changes, written before the data files are updated.
On crash, the system replays the WAL to recover any committed-but-not-yet-flushed changes. The durability foundation of nearly every database (Postgres, SQLite, etc.) and of Kafka-style logs.
Long-lived bidirectional TCP connection between client and server, multiplexed over HTTP upgrade.
Append-only log of intended changes, written before the data files. Crash-recovery foundation.