Building High-Performance Distributed Systems

Learn the core principles of building resilient, low-latency distributed architectures with event streaming, consensus algorithms, and caching layers.

Most systems do not fail because a single machine went down. They fail because the parts that were meant to cover for each other were all waiting on the same thing at once. This session is about designing for that case rather than the happy path.

Marcus works through the three pieces that carry most of the load in practice: an event log as the source of truth, a consensus layer for the small amount of state that genuinely has to agree, and a cache tier that is allowed to be wrong for a bounded period. Each is introduced with the failure it is there to absorb, then the new failure it introduces in exchange.

The worked example is an order pipeline handling roughly 8,000 writes a second. You will see where idempotency keys go, why retries need jitter, how a backlog on one consumer becomes latency on an unrelated endpoint, and what the dashboards look like an hour before an incident is declared.