perf-bar

Overview

My performance/algorithmic bar. Usable standalone on a diff, fed into the SPIKE sub-phase of eskills:task-flow (approach selection), and invoked by its per-task review. It gets its own lens because bottlenecks are a recurring core challenge. (This is the 4th lens — the hard cap; a new lens must displace one.)

Hot-path checklist

• Big-O where it matters. Profile or reason about complexity on hot paths; an O(n²) loop on a small list is fine, on a hot path it isn't. Don't micro-optimize cold code.
• N+1 queries. A query inside a loop over rows → batch it, join it, or prefetch.
• Payload size, not just row count. What crosses the wire per operation? Big blobs, over-fetching, sending a whole row when a version number would do.

Per-stack hot-spots

Same lens, applied to the stack in play — the footgun I hit most in each:

• React / Vite: wasted re-renders → memoize (memo/useMemo/useCallback), stable list keys, split context; virtualize long lists. Profile with the React DevTools profiler, don't guess. Rule-level detail (derive in render, derived-boolean subscriptions, content-visibility, localStorage discipline) in references/react.md; chunking, pre-bundling and what hashed chunks mean at deploy time in references/vite-build.md.
• Django: ORM N+1 → select_related (FK) / prefetch_related (M2M); fetch only needed columns (.only()/.values()); never run a query inside a template or loop.
• Flask / FastAPI: blocking I/O on an async path — a sync DB/HTTP call inside async def stalls the event loop → use an async client or offload to a threadpool; keep CPU-bound work off the loop.
• Kotlin coroutines: blocking work on the wrong dispatcher → Dispatchers.IO for blocking calls, wrap with withContext; no runBlocking on hot paths; don't re-collect cold Flows needlessly.

Supabase egress — my binding limit

The Free-tier wall is egress (bytes transmitted), not DB size. Egress ≈ blob_size × updates × connected_clients. Levers, in order:

1. Cap unbounded growth. Any field that grows per action (history, trash, logs) needs a hard cap — uncapped blobs grow quadratically and blow the limit.
2. Don't ship the whole row on every change. Broadcast only {version} over Realtime; fetch the full payload only when the remote version is newer (version-gated fetch).
3. Write-through the local cache so a client doesn't refetch its own write.

Hard challenges

When the task is flagged hard-tech: in the SPIKE, benchmark candidate approaches on representative input before committing the plan — measure, don't guess. Keep the micro-benchmark; it's the evidence the approach was chosen on merit.

Output

Report findings as location → cost (Big-O / bytes / queries) → the lever. Separate a measured regression from a theoretical one; prefer numbers.