From why columns are stored together, through row groups, chunks, pages and the table of contents at the tail, a walk through the inside of a Parquet file. Seven animations show min/max and Bloom filters pruning row groups and HTTP Range fetching just the needed slices.
Grouping 10,000+ same-category notifications into one third-party group meant the chunks a batch server throws as jobs get consumed asynchronously. Where waiting-children fits in keeping one group and one send on top of that, and how five races were closed one by one and verified with invariants.
Delivery-history JSON sitting in S3 is hard to query as it stands. Four options weighed on always-on infrastructure, the shape of the query cost, and freshness, and why a nightly Parquet conversion won.
stdio MCP servers duplicated per session added up to 169 processes and 6.7GB of RSS. I built a proxy to share one process across sessions — and finished with a bigger question. Do these servers still need to be MCP at all?
How a Kubernetes pod connects to a database with no password anywhere (IRSA and rds_iam), and how the account that connects keeps its privileges minimal (owner separation and default privileges).