Backend Systems · Cluster
Service Reliability
Timeouts, backpressure, and circuit breaking.
- 01 Deadline Propagation and the Shared Request Time Budget Why independent per-hop timeouts cannot bound end-to-end latency, and how an absolute deadline carried in request metadata turns the budget into arithmetic.
- 02 Retry Budgets and the Load Multiplication They Prevent How per-client retry counts turn a partial dependency failure into a full outage, and why a budget denominated in successful requests bounds the amplification instead.
- 03 Queue Time, Service Time, and What a Timeout Measures A client-side timeout covers every wait in the path, including queues the client never sees, and that decomposition explains why a server can look fast while its callers time out.
- 04 Bounded Queues and Load Shedding at the Admission Boundary An unbounded queue converts an overload into a latency problem and then into a memory problem, while a bounded one forces the choice of which requests to drop and when.
- 05 Circuit Breaker State Machines and Half-Open Probing The three states, the counters that drive transitions, and why the half-open probe count is the parameter that decides whether a recovering dependency survives its own recovery.
- 06 Concurrency Limits Derived from Little's Law How the relation between in-flight requests, arrival rate, and latency produces a concurrency limit, and why an adaptive limit tracks a dependency that changes speed.
- 07 Hedged Requests and the Arithmetic of Tail Latency Why the slowest of several parallel calls sets the latency of a fan-out, and how a delayed second attempt trades a fixed slice of extra load for a large reduction in the tail.
- 08 Cancellation Propagation and the Work That Outlives Its Caller Cancellation is a cooperative signal that reaches only the code that checks it, so a cancelled request can still commit a write; scoping the non-cancellable part is the design work.
- 09 Metastable Failure and the Load That Keeps a System Down Why some outages persist after the trigger is gone, how retries and cold caches hold a system in a low-capacity state, and what it takes to break the loop.
- 10 Bulkhead Isolation and the Design of Partial Degradation Partitioning capacity so one dependency cannot consume all of it, and deciding in advance which features are worth serving when the partition is full.