A field guide to performance and cost engineering: claim-checking with the scaling and queueing laws, measurement discipline, where GPU and CPU budgets leak, and how to turn a benchmark into a hardware decision.
Second in my benchmark best practices series. Little’s law checks that throughput, latency and concurrency describe the same system, exposes the load generator’s ceiling, and turns a latency budget into a sizing requirement.
HEAD probes on delete markers answer 405 and replication counts that as success. 20M stuck versions turned the loop into 684M useless requests in 48 h.
Part 1 of my benchmark best practices series: how I turn a scaling claim into an implied serial fraction with Amdahl’s law, and what the Universal Scalability Law adds. Includes a measured pipeline where assigning more vCPUs per image cut global throughput.
Switching tile formats on a LiDAR ingestion pipeline cut output object count by roughly 25x at the same data volume, independent of compression. Here is why object count is its own cost and performance axis on object storage, and what to check before you scale a pipeline that writes many small files.
Choose a message broker for real-time IoT control: start from the latency budget, test ack mode and packet loss, and when to skip the broker for UDP.
MinIO site replication has no cross-site quorum, and its replication queue exists only inside the source cluster: lose that site and the un-replicated backlog is gone and un-enumerable. Sync mode costs latency without fixing that. The precise model, the RTO/RPO table, and the DR patterns that match real SLAs.
MinIO has no NameNode. It puts the namespace on XFS as real directories and xl.meta files. On Lots of Small Files that means you can exhaust inodes while df -h still looks fine, and a flat leaf can stall PUT, LIST, scanner, and ILM together. Here is the on-disk model, the inode math, and a prefix recipe that keeps XFS inside a regime you can operate.
MinIO fixed the HDFS NameNode limit, but it has no index: it writes one xl.meta per object on every drive of the erasure set. On 5 servers of 12 NVMe, MinIO picks a 15-wide set by default, so each small object is stored 15 times over. A worked sizing that looks fine for two years and dies in days.
Sail is an open-source Apache Spark replacement written in Rust. It drops the JVM, speaks Spark Connect, and runs 4 to 6 times faster than Spark with native accelerators on ClickBench. A deep dive into its architecture, benchmarks, and production readiness.