Engineering researcher. APIs, databases, infrastructure, systems design.
How to build the export endpoints customers reach for during migration, audit, BI, and account-closure scenarios. The async job pattern, format choice, completeness contract, and the patterns that fail.
From the Pascaline to the Burroughs Class 3 to the late-1970s extinction event: an 80-year mechanical-calculator era largely missing from cultural memory, and the office labor structure it built.
The deadlock_timeout is how long Postgres waits before running the deadlock detector. The lock_timeout is how long a statement waits for any lock before failing. They look similar in the config and they defend against very different problems.
The eight lock modes Postgres uses, what they conflict with, the cases each one comes from, and how to read pg_locks for diagnosis. The lock-vs-lwlock distinction.
The dual-algorithm rollover pattern that lets receivers update on their schedule, the v1=/v2= header format every signature scheme should ship with, and the deprecation window B2B SaaS customers can plan around.
Long-running operations return a job ID instead of a result. Customers poll the status endpoint until the job is complete. The shape of that endpoint determines how much customer code is needed to handle the workflow correctly.
ClickHouse merges old partitions in the background to keep query performance from degrading. When the server hits its memory limit during a merge, the whole process exits and the cgroup restart loop begins.
After a Postgres restart, the buffer cache is empty and every query that touches a previously-hot table pays cold-cache latency. pg_prewarm lets you push specific tables and indexes into shared_buffers before traffic arrives, with autoprewarm preserving the cache across restarts.
Most APIs ship with one account type: a single user, a single API key, a single billing relationship. That model breaks when customers want shared resources, multiple seats, or centralized billing. The trick is adding a hierarchy layer in a way that does not break every existing endpoint.
Autovacuum handles most Postgres maintenance correctly, but a small set of cases benefits from manual VACUUM ANALYZE: large bulk loads, post-migration backfills, statistics drift after schema changes. The trick is knowing which case you are in before you run anything.
Customer filtering needs are usually narrower than they look. A small grammar of field-equals-value plus a handful of comparison operators handles 80-90 percent of B2B SaaS list endpoint use, while still keeping the implementation indexable and the query language documentable in a single page.
The webhook subscription dashboard exposes what events a customer can subscribe to. The right interface depends on event taxonomy, the wrong one produces support tickets. The patterns that work cover three job categories and avoid four anti-patterns.