3.9 KiB
Bolt's Journal ⚡
This journal is for CRITICAL, non-routine performance learnings ONLY.
- Codebase-specific bottlenecks
- Failed optimizations (and why)
- Surprising performance patterns
- Rejected changes with valuable lessons
2024-07-25 - MQL5 Native Functions vs. Scripted Loops
Learning: My assumption that a manual MQL5 loop over a pre-cached array would be faster than built-in functions like iHighest() and iLowest() was incorrect. The code review pointed out that MQL5's native, built-in functions are implemented in highly optimized C++ and are significantly faster than loops executed in the MQL5 scripting layer. The original comment stating this was correct.
Action: Always prefer using MQL5's built-in, native functions for calculations like finding highs/lows over manual loops, even if the data is already in a local array. The performance gain from the native implementation outweighs the overhead of the function call.
2026-01-23 - Python File System Checks
Learning: Checking for file existence (os.path.exists) before getting metadata (os.path.getmtime) introduces a redundant syscall. os.stat() provides both pieces of information in a single syscall and uses the EAFP (Easier to Ask for Forgiveness than Permission) pattern, which is more Pythonic and slightly faster, especially in high-frequency loops or handlers.
Action: Use os.stat() when both existence and metadata are needed, wrapping it in a try...except OSError block.
2026-01-26 - yfinance Bulk Download
Learning: yfinance Ticker.history in a loop is significantly slower than yf.download with a list of tickers due to sequential HTTP requests. yf.download with group_by='ticker' provides a consistent MultiIndex structure even for single tickers, simplifying bulk processing.
Action: Always prefer yf.download(tickers) over iterating yf.Ticker(t) when fetching data for multiple symbols.
2026-02-09 - Git Command Performance
Learning: git for-each-ref is a powerful tool for batch data retrieval, but without filtering, it processes all refs, including thousands of stale merged branches in older repositories. Calculating ahead-behind counts for these stale branches is O(N) where N is total branches, which can be significantly slower than O(M) where M is active branches.
Action: Always filter git for-each-ref with --no-merged (or --merged depending on use case) when only interested in a subset of branches, especially when expensive formatting options like ahead-behind are used.
2026-02-14 - Markdown Parser Reuse
Learning: markdown.markdown() is a convenience function that re-initializes the Markdown parser class on every call. In high-frequency or repetitive rendering scenarios (like a dashboard), this adds significant overhead (~10% in this case). Reusing a markdown.Markdown instance via threading.local() (for thread safety in WSGI apps) eliminates this initialization cost.
Action: When using markdown library in a web server or loop, instantiate markdown.Markdown once per thread/process and reuse it with .reset() instead of calling the top-level markdown.markdown() shortcut.
2026-02-28 - os.walk vs pathlib.Path.rglob for large directory exclusion
Learning: When scanning an entire repository, using Path.rglob("*") and filtering out excluded_dirnames later is extremely slow because rglob will still traverse into massive directories like node_modules and .git before discarding the paths. os.walk(..., topdown=True) allows you to modify the dirs list in-place (e.g., dirs[:] = [d for d in dirs if d not in excluded_dirnames]), completely preventing the traversal of these heavy directories and yielding massive performance improvements.
Action: Always prefer os.walk(..., topdown=True) with in-place dirs pruning over rglob + filtering when large excluded directories (like node_modules, .venv, or .git) are expected.