scale-sweeplisted
Install: claude install-skill samiraklf/claudrunner
# Scale sweep
Assume production is larger than your test data by three orders of magnitude. Most scale
bugs are invisible at development size and fatal at real size.
## Establish the scale profile first
Before judging any query, find out how big the tables actually are. Read the project's own
notes in `.claudrunner/notes.md`, then verify against the database where you have read
access. Record what you learn back into that file. A review without a scale profile is
guessing.
## What to hunt
**Queries in a loop.** The signature is a lookup whose input is one element of a collection
you are already iterating. A filter with a single-element list inside a loop is the same
bug wearing a disguise.
**Unbounded reads.** Any fetch with no limit, on a table that grows. Ask what happens at a
million rows, then at a hundred million.
**Missing indexes.** Every column newly used for filtering, joining, or ordering. Check the
schema, do not assume. Also check that the index order matches the query's leading columns.
**Whole-object buffering.** Files, exports and batches read fully into memory before being
written out. Streaming costs one line and removes the ceiling. Check the memory budget of
the process that runs it.
**Background work.** Three specific traps:
1. A job whose timeout exceeds the queue's retry window — it gets processed twice.
2. A job that is not safe to run twice, when retries are enabled.
3. A job assigned to a queue that no worker actually serves.
**Full-table op