notes of a backend developer — postgres, python, and small tools
Most tables have a soft-delete flag or a status column, and most queries
filter on it. A partial index over the rows you actually select is
smaller, hotter in cache, and cheaper to maintain than an index over
everything. We shaved a p95 report query from 900 ms to 40 ms
by indexing WHERE deleted_at IS NULL and nothing else.
Every team I join debates replacing make with something fancier, and
three years later the Makefile is still there. Keep targets tiny, echo
what they do, and let newcomers run make help. The boring
tool that everyone already understands wins.
py-spy made profiling a service you can do during lunch: record a flame graph of the live process, no restarts, no instrumentation. The trick is to stop reading individual frames and look for wide plateaus instead — that is where the time actually goes.
Not because the documents are valuable — because writing forces the argument to exist. Half of our ADRs end with “revisited, decided against”, and that is the good outcome: a page of writing is cheaper than a week of the wrong architecture.