Highload Architecture

The minimal queueing math for capacity reasoning — Little's law, the M/M/1 utilization hockey stick, and Kingman's variability factor — plus the sizing rules that fall out and the places the models stop applying.

The five caching patterns (cache-aside, read-through, write-through, write-behind, refresh-ahead), why invalidation is genuinely hard, and the production failure modes — stampedes, stale-set races, hot keys — with the mechanisms that contain them.

The state machines behind work queues and publish-subscribe: routing, publisher confirms, consumer acknowledgements, redelivery, ordering, backpressure, retries, dead letters, and the integration patterns that turn broker primitives into reliable application workflows.

How retained partitioned logs work: offsets, keys, consumer groups, rebalances, retention, compaction, replication, lag, replay, and the narrow boundary of exactly-once processing, with a mechanism-level comparison of Kafka, Redpanda, Pulsar, RabbitMQ Streams, and NATS JetStream.

The MapReduce contract, shuffle mechanics, retries, output commits, skew, stragglers, combiners, and the design boundary that makes large batch jobs both scalable and diagnosable.