Full Width 
Shortcuts 
Sign Up 
Log In 
Articles 
Reading List More from Krzysztof Kowalczyk blog
I’m porting gpui-component (a Rust UI component library built on GPUI) to C++ as gpui-cpp. By which I mean: my friend Claude does the porting, I’m just directing. It uses markdown-rs (a CommonMark + GFM parser) markdown parser so I ported it too. Then I optimized it. This post describes what I did with the intention of teaching other how to optimize C++ code. The starting point There are 2 kinds of markdown parser: those that stream nodes as they parse those that build an AST in memory markdown-rs builds an AST. The game is about minimizing the size of AST node. In Rust there are various kinds of nodes, the largest being 152 bytes. Claude generated a single Node struct of 232 bytes. I got it down to 16 bytes. Here’s the initial Node struct, before optimizations: Node, 232 bytes k children 24 position 24 8 string fields — 128 bytes align 24 nums 16 grey = padding and small fields · blue = growable vector · yellow = pointer+length strings Node, 16 bytes (same scale) lastKid · sibling · firstStr · kind+flags Where the 232 went: 8 string fields at 16 bytes each (a char* plus a length), two growable vectors at 24 bytes each (children and table alignments), a 24-byte unist Position (line, column and offset at each end), six bools one to a byte, and the padding all of that dragged in. Every node in the tree pays for every field, whichever kind it is. A Text node uses one string field and nothing else. Arena allocator It’s important that all allocations are done in an arena. Nodes in a parse tree all have the same lifetime which makes it a perfect use for an arena: a bump allocator that can only grow. The only way to free memory is to reset the arena. This is different than calling malloc() to allocate each node individually and then having to call free(). It makes it easy to measure memory usage: check the arena size after parsing. It also allows optimization tricks like compressing pointers. How I measured bun cmd/bench.ts markdown parses 64 KB of markdown in four shapes and reports the arena bytes the parse allocated: prose — paragraphs, emphasis, links nested lists — deep blockquotes and lists gfm tables — tables all the way down entities — text that is mostly &-style character references The number is the whole arena: nodes, the tokenizer’s event list, and the strings. Not just sizeof(Node) × node count. We also measure parsing time to make sure we don’t trade size for speed. Baseline, 64 KB of source: prose 1646.1 KB (25.7× the source) nested lists 1067.9 KB gfm tables 2926.0 KB entities 660.2 KB 1. Pointer compression for strings (bed71ee) On 64-bit platforms, pointers are 8 bytes. Pointer compression reduces this to 4 bytes by calculating a 32-bit offset against a base pointer. Google used compressed pointers in v8 with great result. Reduced memory usage and increased speed. Our string type is the simplest possible string: struct Str { char* data; size_t len; }; That’s at least 12 bytes per string, if len is 4 bytes. Due to alignment, the size is 16 bytes. Strings are allocated in Arena so we can use the beginning of an arena as a base pointer and optimize the pointer from 8 bytes to 4 bytes. We typedef ArenaStr as uint64_t. The lower 4 bytes is uint32_t compressed pointer and upper uint32_t is size. We reduced the overhead of strings from 16 bytes to 8 bytes. Times 8 strings that’s 64 bytes saved per node. Added helper functions for allocating ArenaStr in arena and converting ArenaStr to Str. Savings: 8 strings * 8 bytes, 64 bytes per node: 232 → 168 bytes. shape start before after vs before vs start prose 1646.1 KB 1646.1 KB 1285.9 KB -21.9% -21.9% nested lists 1067.9 KB 1067.9 KB 867.5 KB -18.8% -18.8% gfm tables 2926.0 KB 2926.0 KB 2269.7 KB -22.4% -22.4% entities 660.2 KB 660.2 KB 626.2 KB -5.1% -5.1% 2. Growing arena strings in place (a9d4f3a) Some strings had to grow. Arena allocator doesn’t provide freeing or reallocation. You can only allocate new strings, which wastes memory by leaving dead copies of the string we were appending to. We can grow the last allocated string and that’s what this change does. Luckily, most appends were done to the last string. ArenaStrAppend checks whether the string ends exactly where the arena’s next allocation would begin. If it does, the new bytes are pushed straight onto it and nothing is copied. Decoding HTML entities (e.g. &) broke that optimization by doing an allocation before appending to the string. We switched to decoding entities into a 4-byte stack buffer which enabled optimized append. shape start before after vs before vs start prose 1646.1 KB 1285.9 KB 1285.9 KB +0.0% -21.9% nested lists 1067.9 KB 867.5 KB 729.2 KB -15.9% -31.7% gfm tables 2926.0 KB 2269.7 KB 2269.7 KB +0.0% -22.4% entities 660.2 KB 626.2 KB 163.8 KB -73.8% -75.2% 3. Re-order struct fields, pack the bools (5c0ce6e) Unless told to pack the layout of the struct, C++ compilers align struct fields to the size of the largest primitive type. If you sandwich a bool between 2 uint64_t values, the bool will occupy 8 bytes (sizeof(uint64_t)) instead of 1 byte as it should. Our Node had such wasted space due to padding. My friend Claude was careless. A simple fix is to re-arrange fields, putting the largest first. We also had six bool field which we packed into a uint8_t flags field. Result: 168 → 144 bytes, with no padding at all. We’re beating Rust version now. declaration order: bool after vector = 7 bytes of padding, six times over vector b padding strings b padding largest first, bools in one byte: no padding vectors strings nums f shape start before after vs before vs start prose 1646.1 KB 1285.9 KB 1150.9 KB -10.5% -30.1% nested lists 1067.9 KB 729.2 KB 654.0 KB -10.3% -38.8% gfm tables 2926.0 KB 2269.7 KB 2023.6 KB -10.8% -30.8% entities 660.2 KB 163.8 KB 151.0 KB -7.8% -77.1% Free bytes: same fields, same code, different order. 4. Pointer compression for everything (07ec80f) We compress pointer for all objects allocated in the arena, like we compressed a pointer to the string. ArenaVec<Node*> children held 8-byte addresses; ArenaPtr<T> is a 4-byte offset into the arena’s position space, resolved by ArenaAtOffset. Zero is null, which costs nothing because no allocation ever lands at offset zero. The Node itself doesn’t change size — a vector handle is the same three words whatever it holds — so all of the saving is in the child arrays. shape start before after vs before vs start prose 1646.1 KB 1150.9 KB 1091.9 KB -5.1% -33.7% nested lists 1067.9 KB 654.0 KB 611.6 KB -6.5% -42.7% gfm tables 2926.0 KB 2023.6 KB 1866.9 KB -7.7% -36.2% entities 660.2 KB 151.0 KB 144.9 KB -4.0% -78.1% These shapes rank by children-per-node rather than by node count, which is why tables moved most. 5. Varint encoding string length (f9ebc34) ArenaStr was an offset and a length in 8 bytes. Now it’s the offset alone — 4 bytes — and the length is varint-encoded at the beginning of the string data: [varint len][string bytes][NUL] There are many varint encoding schemes. This one is for unsigned number and codes number < 128 as a single byte. Most strings are below that threshold, so they use a single byte for the varint length, saving roughly 3 bytes per string. Node shrinks from 144 → 112 bytes. Str — pointer + length, 16 bytes per field char* s int64 len ArenaStr — offset + length, 8 bytes u32 off u32 len ArenaStr — offset alone, 4 bytes; the length lives in the arena u32 off len characters 0 Caveat: An offset-and-length string can point at a slice of another string, and a length-prefixed one can’t. We weren’t doing it so it doesn’t apply here. shape start before after vs before vs start prose 1646.1 KB 1091.9 KB 918.0 KB -15.9% -44.2% nested lists 1067.9 KB 611.6 KB 512.3 KB -16.2% -52.0% gfm tables 2926.0 KB 1866.9 KB 1543.6 KB -17.3% -47.2% entities 660.2 KB 144.9 KB 128.5 KB -11.3% -80.5% 6. Fusing exclusive fields (5467a18) A List has a start number. A Heading has a depth. No node is ever both, so they became one uint32_t startOrDepth and kind says which it means. It didn’t shrink the size of Node due to the alignment padding but we did it anyway hoping that future optimization would shrink below padding. shape start before after vs before vs start prose 1646.1 KB 918.0 KB 918.0 KB +0.0% -44.2% nested lists 1067.9 KB 512.3 KB 512.3 KB +0.0% -52.0% gfm tables 2926.0 KB 1543.6 KB 1543.6 KB +0.0% -47.2% entities 660.2 KB 128.5 KB 128.5 KB +0.0% -80.5% 7. Compressing text position (ed5e807) Each Node carried the info about its position in parsed text. It was expensive because it was stored as start and end fields and each of them was: a uint32_t line a uint32_t column a uint32_t offset That’s 4*3*2 = 24 bytes. I assume this info is for debugging so not important for me. I replaced it with 2 uint32_t offsets into a source markdown string, srcStart and srcEnd. We can reconstruct the line/column position from that and the source string. shape start before after vs before vs start prose 1646.1 KB 918.0 KB 828.0 KB -9.8% -49.7% nested lists 1067.9 KB 512.3 KB 462.2 KB -9.8% -56.7% gfm tables 2926.0 KB 1543.6 KB 1379.6 KB -10.6% -52.9% entities 660.2 KB 128.5 KB 120.0 KB -6.6% -81.8% 8. Further compression text position (6a558c4) srcEnd is always after srcStart so we can delta-encode it and shrink to uint16_t. What if it’s bigger than 64 KB? I don’t care, we store it as 65535. This is another case where due to padding we didn’t shrink the struct size. But wait for it. shape start before after vs before vs start prose 1646.1 KB 828.0 KB 828.0 KB +0.0% -49.7% nested lists 1067.9 KB 462.2 KB 462.2 KB +0.0% -56.7% gfm tables 2926.0 KB 1379.6 KB 1379.6 KB +0.0% -52.9% entities 660.2 KB 120.0 KB 120.0 KB +0.0% -81.8% 9. Optimizing storing children (d6c4abc) Some nodes have children that were stored as a growable vector. Empty vector was 24 bytes in the node. We replaced it with a ring of compressed pointers: the parent names its last child, each child names the next one, and the last child wraps back to the first. vector: 24 bytes in the node + a separate array of links ptr · len · cap kid0 kid1 kid2 spare spare ring: 4 bytes in the parent, 4 in each child, nothing else allocated parent kid0 kid1 kid2 lastKid We use a ring and not just a linked list because appending is the only thing the parser does to a child list. A single linked list requires walking the list to find the end, while a ring does not. Saving: 96 → 80 bytes. shape start before after vs before vs start prose 1646.1 KB 828.0 KB 619.0 KB -25.2% -62.4% nested lists 1067.9 KB 462.2 KB 308.8 KB -33.2% -71.1% gfm tables 2926.0 KB 1379.6 KB 898.7 KB -34.9% -69.3% entities 660.2 KB 120.0 KB 98.9 KB -17.6% -85.0% Caveat: accessing a child by index would require a walk through the ring, so indexing in a loop would be quadratic. In our code we only ask for the first or the last. 10. Compressing table alignments info (ca0818c) For tables we were storing column alignments in a separate vector on every node, even though only Table nodes have them. Another 24 bytes per node. We switched to a compressed pointer which points to an optimized representation of the column alignments. There are four alignments (left, right, center, none), so a column needs 2 bits: [varint count][2 bits a column, four to a byte] The whole list is known when the table is entered, so it’s counted, allocated once and filled. For an 8-column table that’s 3 bytes in the arena and a 4-byte offset in the node. Saving: 80 → 60 bytes. We saved more than the 20 bytes because with the last pointer-holding member gone alignof(Node) fell from 8 to 4. shape start before after vs before vs start prose 1646.1 KB 619.0 KB 519.3 KB -16.1% -68.5% nested lists 1067.9 KB 308.8 KB 256.3 KB -17.0% -76.0% gfm tables 2926.0 KB 898.7 KB 710.3 KB -21.0% -75.7% entities 660.2 KB 98.9 KB 89.2 KB -9.8% -86.5% The block is pushed byte-aligned rather than through the general allocator, which rounds to 8 and would have handed back exactly what the varint saved. 11. Fusing exclusive fields (07444d6) Previously we fused exclusive fields start of a List node and depth of a Heading node into a single uint32_t. We fused Table node column alignments info from previous optimization into the same field. We called it uint32_t perKind, and kind says what kind of value it is. Saving: 60 → 56 bytes. shape start before after vs before vs start prose 1646.1 KB 519.3 KB 483.9 KB -6.8% -70.6% nested lists 1067.9 KB 256.3 KB 233.7 KB -8.8% -78.1% gfm tables 2926.0 KB 710.3 KB 646.1 KB -9.0% -77.9% entities 660.2 KB 89.2 KB 86.1 KB -3.5% -87.0% 12. Optimizing eight strings (521e32e) We had 8 strings that were not all used by all nodes. Instead of figuring out how many strings we need at most, I created a linked list of strings in the arena. They are different than regular strings in that they carry a 4 byte compressed pointer to the next string within the arena and the kind of the strings. [u32 next][u8 kind][varint len][len bytes][NUL] We can add as many kinds of strings as we need but we only pay for used strings + 5 byte per-string overhead. Some nodes don’t have any strings. 8 fields: 32 bytes on every node, 7 of them empty on almost all of them value url title alt ident label lang meta 1 field: 4 bytes, and a record only for what the node actually carries first next kind len characters 0 a stored string costs 5 bytes more · a node storing none saves 28 New records go on the head, so storing is O(1), and the walk that finds a kind is at most 8 long and is almost always 1 or 0. In-place growth still works, because a record being the newest thing in the arena is the same condition it always was. Saving: 56 → 28 bytes. shape start before after vs before vs start prose 1646.1 KB 483.9 KB 358.3 KB -26.0% -78.2% nested lists 1067.9 KB 233.7 KB 159.1 KB -31.9% -85.1% gfm tables 2926.0 KB 646.1 KB 402.9 KB -37.6% -86.2% entities 660.2 KB 86.1 KB 73.7 KB -14.4% -88.8% 13. Fusing two enums into one (f3c14b9) As it happened we had two enums: one needed 6 bits another needed 2 bits We fused them from 2 bytes to 1 byte. Because this 1 byte saving dropped below padding we saved 4 bytes and went from 28 → 24 bytes. shape start before after vs before vs start prose 1646.1 KB 358.3 KB 321.2 KB -10.4% -80.5% nested lists 1067.9 KB 159.1 KB 136.5 KB -14.2% -87.2% gfm tables 2926.0 KB 402.9 KB 341.9 KB -15.1% -88.3% entities 660.2 KB 73.7 KB 70.4 KB -4.5% -89.3% 14. Remove position, reduce allocator’s alignment (861c803) At this point I decided that I didn’t need the position so I removed it. Other markdown parsers don’t carry it around so it doesn’t seem very useful. I reduced overhead of perKind by converting it to a record in the string list from step 12 — varint-encoded, under its own kind byte. A List, Heading or Table pays ~8 bytes for it; every other node pays nothing, where a field cost 4 bytes on all of them. Savings: 24 → 16 bytes. For safety arena allocator aligns allocations to 8 bytes but a 16 bytes Node can be allocated at 4 bytes, which we did. This reduces wasted space between allocations. shape start before after vs before vs start prose 1646.1 KB 321.2 KB 272.0 KB -15.3% -83.5% nested lists 1067.9 KB 136.5 KB 110.2 KB -19.3% -89.7% gfm tables 2926.0 KB 341.9 KB 250.5 KB -26.7% -91.4% entities 660.2 KB 70.4 KB 65.6 KB -6.8% -90.1% End results The results are pretty dramatic: sizeof(Node) prose nested tables entities start 232 1646.1 KB 1067.9 KB 2926.0 KB 660.2 KB end 16 272.0 KB 110.2 KB 250.5 KB 65.6 KB -93% -83.5% -89.7% -91.4% -90.1% A parse of 64 KB of prose cost 25.7× the source in arena bytes. It costs 4.2× now. The entities shape went from 10.3× to 1.02×. The speed was unchanged. Fastest of 3 runs: prose 8.47 → 8.22 ms nested 9.45 → 9.26 ms tables 12.88 → 12.92 ms entities 5.90 → 5.85 ms Those are within margin of error. The phase of building the tree got a measurable speed up: 0.397 → 0.302 ms, about 24% faster. This is from allocating less and touching fewer cache lines. This is not visible on micro benchmarks, but using less memory will slightly speed up the rest of the application. Lessons learned Arranging struct fields by size is good. It costs literally nothing. Pointer compression is good. 8 bytes become 4 bytes and the cost of converting back and forth is negligible, as Google shown in their v8 blog post and is re-inforced by our benchmarks Varint-encoding is good. Most strings are short so varint encoding can save 3 bytes per string on average. Moving rare fields out of line is good. The way we reduced 8 strings into an out-of-line list. Only pays off if savings is bigger than the cost of additional metadata. sizeof only drops when the saving crosses an alignment boundary. Two of our changes didn’t reduce size of Node struct but it paid off in later optimizations. The allocator’s alignment is part of sizeof. A 28-byte struct from an 8-aligned bump allocator is 32 bytes. We need benchmarks. You can’t improve what you can’t measure. Our benchmarks measured both memory usage and speed, to ensure we didn’t regress speed to save memory.
I maintain SumatraPDF on GitHub. People fork it and make their own changes. I want to know which forks are active and what they’re working on. GitHub has a Network tab for this. I find it lousy. It’s hard to see active forks at a glance. Panning and zooming the UI is slow and fiddly. I just want a sorted list of forks with actual changes. So I wrote a small script: github-active-forks.ts. What it does It uses the GitHub API to find forks that have ahead commits (changes not in upstream) from the last year. For each active branch it prints: a link to the branch on GitHub short sha, author, relative date, first line of commit message Forks are sorted by most recent activity, oldest first. Commits within a branch are oldest first. It compares each fork branch against the matching upstream branch (e.g. rel3.6working vs rel3.6working), not always master. That way you only see commits unique to the fork. How to run it Download the script: curl -O https://gist.githubusercontent.com/kjk/71a7679dd408d52bc612cdf5eecace58/raw/7d7f605dde6b739424e5190a79e3988477312e74/github-active-forks.ts You need Bun and the GitHub CLI logged in (gh auth login). Or set GITHUB_TOKEN. Run it on any repo as owner/repo: bun github-active-forks.ts sumatrapdfreader/sumatrapdf Progress goes to stderr; results go to stdout, so you can redirect to a file: bun github-active-forks.ts sumatrapdfreader/sumatrapdf > forks.txt Example output Excerpt from running on SumatraPDF (74 forks with ahead commits): https://github.com/wackget/sumatrapdf-Feldherren-version/tree/single-page-fit-scrollbar (289d ago) c53a51d wackget 294d ago Added scrollbar usable in Fit a Single Page display mode. f927640 wackget 294d ago scrollbar in single page mode now obeys the hideScrollbars setting. 7f57861 wackget 288d ago improved scrolling speed when zoomed in, in non-continuous single-page view. https://github.com/xyzzyx99/sumatrapdf/tree/rel3.6working (5d ago) c8d1f7b xyzzyx99 13d ago Add GitHub Actions workflow for building project 8a92547 xyzzyx99 13d ago Fix bugs, revert using version 3.6 1f27324 xyzzyx99 13d ago Save and restore CHM scroll position 24e7974 xyzzyx99 6d ago Preserve current CHM URL across tab restore 30b1cb4 xyzzyx99 5d ago Update README with new bug fixes and formatting https://github.com/lsq/sumatrapdf/tree/dev (today) 72d7069 lsq 11d ago feat: support share file via socket a1fc427 lsq 10d ago feat: open home page 938bc07 lsq today add localsend v2.1 api support Much easier to scan than the Network graph.
A new JavaScript library pretext for fast text measuring / layout popped up on social media. Potentially interesting given its focus on speeding up text rendering in web apps and me writing web apps and liking them being fast. I looked at the code and saw a function isCJK(). Given my 3 decades of programming and performance optimization, it looked like it could be sped up. This is a story about ideas on making JavaScript faster and the process of quickly implementing and benchmarking them. The code export function isCJK(s: string): boolean { for (const ch of s) { const c = ch.codePointAt(0)! if ((c >= 0x4E00 && c <= 0x9FFF) || (c >= 0x3400 && c <= 0x4DBF) || (c >= 0x20000 && c <= 0x2A6DF) || (c >= 0x2A700 && c <= 0x2B73F) || (c >= 0x2B740 && c <= 0x2B81F) || (c >= 0x2B820 && c <= 0x2CEAF) || (c >= 0x2CEB0 && c <= 0x2EBEF) || (c >= 0x30000 && c <= 0x3134F) || (c >= 0xF900 && c <= 0xFAFF) || (c >= 0x2F800 && c <= 0x2FA1F) || (c >= 0x3000 && c <= 0x303F) || (c >= 0x3040 && c <= 0x309F) || (c >= 0x30A0 && c <= 0x30FF) || (c >= 0xAC00 && c <= 0xD7AF) || (c >= 0xFF00 && c <= 0xFFEF)) { return true } } return false } My spider sense tingling To make code run fast you have to have mechanical sympathy. You need a good mental model of how CPUs and programming languages work, at the low level. Because I have mechanical sympathy, I know that to evaluate multiple || statements, the program has to check every statement until it finds one that is true. For the case of not matching any range, it has to do all 15 comparisons. My immediate thought was that most characters are ascii (non-cjk) and therefore require 15 comparisons. An early exit should speed things up: if (c < 0x300) continue;. It adds 1 comparison for matching cases but removes 14 (15 - 1) for the most common case of ascii characters. Hunch is good, benchmark is better Micro-benchmarking is subtle. In the past I would google for a JavaScript benchmarking library and trust that it’s sound. Claude wrote a sound, simple benchmark function: function bench(name, fn, iterations = 1_000_000) { for (let i = 0; i < 10_000; i++) fn(); // warmup const start = performance.now(); for (let i = 0; i < iterations; i++) fn(); const elapsed = performance.now() - start; const opsPerSec = ((iterations / elapsed) * 1000).toFixed(0); console.log(`${name}: ${elapsed.toFixed(2)} ms (${opsPerSec} ops/sec)`); } // to call: bench("isCJKRegex ", () => isCJKRegex(input)); The subtle part is warmup phase. JavaScript JIT engines that you’re likely to use (v8 in Chrome, JavaScriptCore in Safari, SpiderMonkey in FireFox) have several optimization tiers and only do most aggressive optimization for frequently called functions. The warmup phase forces JIT to compile fn with most aggressive settings. This is only fair - I assume isCJK is called frequently. If it wasn’t, there’s no point sweating to optimize it. The new AI normal In the past I would have to manually implement each optimization idea. With AI I can delegate the coding and therefore test more ideas. The following was done with Claude 4.6. 100% vibe coding There are levels of delegating things to AI. On the extreme side of giving up control is 100% vibe coding. I had 2 ideas for performance optimization: early exit replace 15 linear checks with binary search But what if I didn’t? What if I just ask AI to come up with ideas? The following is my conversation with Claude 4.6 and what I’ve learned. I created is_cjk.js file and pasted the original function. look at is_cjk.js. come up with 2 ideas for how to speed up isCJK(). Implement those as separate functions and write a javascript micro-benchmark that compares those 2 new implementations against isCJK Two ideas: Pre-compiled regex — push the entire check into the native regex engine, avoiding JS-level for…of iteration and codePointAt() overhead. Typed array bitmap for BMP — for BMP characters (the vast majority of real text), a Uint8Array bitmap gives O(1) lookup with a single array access. Fall back to range checks only for supplementary plane characters. I have to admit: I recognize those as valid optimization ideas but they didn’t immediately occur to me. Regex works because regex implementations in JavaScript engines are insanely optimized. Bitmap is a known optimization for doing true/false lookup for numbers. Let’s say you have total range 0-255 and you want a function that checks if a number is in range 1-10, 20-30, 40-50, 60-70, 80-90. That’s 5 if statements, each doing 2 comparisons. You can instead use 1 bit for each of the 256 numbers (i.e. 256⁄8 = 32 bytes), pre-compute the array of bytes and set the bit for the ranges. Then lookup is just math to get byte offset and bit, load the byte, check the bit. Super fast. Claude not only wrote the benchmark function, 2 new implementations but also ran the benchmarks using node and showed me results in a readable way. The bitmap approach was significantly faster. Human assisted coding Bitmap was the fastest and yet I discarded it using my human judgement. The bitmap needed 65kB of memory and that’s just too much for this function. That’s where human judgement comes in to improve collaboration with AI. I didn’t specify that I also care about memory usage so AI wrote fast code that was using too much memory. You’re benchmarking what? Here’s another tricky part of benchmarking: knowing what you benchmark. There are 3 popular JavaScript engines. Claude ran benchmarks using node which uses Chrome’s v8. That’s good because that’s the most popular browser and therefore most popular JavaScript JIT engine. It’s also good to sanity-check with at least one other JavaScript engine. bun uses Safari’s JavaScriptCore so I asked Claude: run the tests using bun The results were similar, which is good. We don’t want code that is fast in one engine but slow in another. We don’t control which browser the user of our code is running. Continuing collaboration with AI I had some more ideas so here’s the rest of my conversation with Claude. don’t benchmark bitmap, it uses too much memory; write a variant of isCJK that does an early false exit if char is less than smallest Early exit is my first insight I wanted to test. As expected, early exit is good 2x speedup for latin strings, although regex beats it on cjk strings. That is the curse of engineering trade-offs: you have to decide if you want to optimize for cjk strings or non-cjk strings. now implement a variant that does inline bisect / binary search That was my second idea: do a binary search instead of sequential if comparisons. It was faster than original but slower than regex / early exit. Also, Claude didn’t do what I meant. It stored the ranges as array: // Flat array: [lo0, hi0, lo1, hi1, ...] sorted by lo const cjkRanges = new Uint32Array([ 0x3000, 0x303f, 0x3040, 0x309f, ... That’s what I meant: implement a variant isCJBisect that doesn’t use array but unrolls binary search logic into if statements The implementation is gnarly: function inCJKUnrolled(c) { if (c < 0x3000) return false; if (c < 0xff00) { if (c < 0x3400) { if (c < 0x3040) { return c <= 0x303f; } else if (c > 0x309f) { return c >= 0x30a0 && c <= 0x30ff; } return true; // ... more gnarly code But what do I care? Claude wrote it. Arguably it’s hard for me to review for correctness but it’s perfectly testable. We can write a test to compare if it returns the same result as original function for every possible input. As expected, unrolled version was faster than array-based but not always faster than regex or early exit. But Claude wrote something that caught my eye: The for…of + codePointAt iterator overhead remains the bottleneck shared by all non-regex variants on long strings. It’s well known that for ... of loop is slower than for (let i = 0; i < n; i++). It’s one of those things that you need to know. I knew this but I didn’t connect the dots in this particular case. write variant of isCJKEarly which replaces for of with a faster loop Result: isCJKEarlyFast crushes everything across the board on both runtimes The code for isCKJEarlyFast. Looks like we have a winner: Replacing for…of + codePointAt() with charCodeAt() + index loop eliminates the iterator overhead, which was the main bottleneck. The early c < 0x3000 bail then skips all range checks for ASCII/latin characters cheaply. Only regex still wins on Node’s long latin miss case thanks to V8’s native string scanning. There’s a subtlety that makes the rewrite of for .. of into for (let i = 0; i < n; i++) tricky. Logically JavaScript strings are composed from 16-bit numbers (characters). Some Unicode code-points are bigger than 16-bit and are stored as 2 16-bit numbers, called surrogate pairs. So writing a loop over unicode codepoints needs to take this into account. It’s tricky code but again, what do I care? Claude wrote it and it’s perfectly testable. Pressing on The good thing about claude writing most of the code and removing other tedium like running the benchmarks is that we can do more in the same amount of time. So I explored even more. I noticed: export function isCJKUnrolled(s) { for (const ch of s) { if (inCJKUnrolled(ch.codePointAt(0))) return true; } return false; } What happens when inCJKUnrolled is inlined? inline inCJKUnrolled into isCJKUnrolled Result: Done. Inlining didn’t help — it actually got slightly slower in most cases (the engine was already inlining the small function). No change in rankings; EarlyFast remains the winner. Turns out it’s not faster, which was surprising. Inlining is supposed to remove overhead of function calls, which matters in loops. Surprising isn’t bad. We run benchmarks to validate our hunches. For completness, I wanted to compare with original where for .. of was rewritten. also write original with for of replaced by standard for loop and bench all and give results as markdown table Claude was nice enough to show benchmark results in a nicely formatted table, without me even asking. It just knows things. But as you can see above, you can ask it for results in markdown table to e.g. include in a GitHub bug report. Here are final results. The number is operations/second in millions. Higher is better. Bun Scenario Original Regex Early Bisect Unrolled EarlyFast ForLoop Single CJK 11M 22M 12M 12M 16M 79M 57M Single latin 20M 24M 49M 21M 22M 40M 36M CJK string 12M 18M 17M 12M 14M 42M 42M Latin string 1.5M 4.7M 3.6M 3.0M 3.5M 17M 6.5M Mixed string 6.5M 5.9M 9.9M 8.6M 13M 62M 34M Node Scenario Original Regex Early Bisect Unrolled EarlyFast ForLoop Single CJK 85M 54M 51M 44M 56M 112M 110M Single latin 59M 72M 83M 55M 74M 115M 103M CJK string 79M 55M 60M 55M 64M 90M 129M Latin string 4.2M 57M 4.3M 2.8M 4.5M 12M 8.3M Mixed string 14M 15M 16M 10M 15M 41M 38M Conclusions AI is a big unlock. It took me under 30 minutes to test various hypotheses and find out a significant speed up. Without Claude it would take several hours and I would likely not do it at all. It’s just not important enough to spend a working day on it. To get best results we still need to apply human judgement and guide the AI. Programming expert knowledge An expert is simply someone who knows things. We know things because we learn them. If you were paying attention you might have learned the following things: importance of warmup phase when benchmarking JIT compilers for .. of is slower than for (let i = 0; i < n; i++) regex matching in JavaScript engines is fast subtlety of surrogate pairs in JavaScript strings using bitmaps to speed up range lookups Resources All the code is in https://gist.github.com/kjk/bdbea9d90c3bb0454fbe26353c521bfd I like to write fast code. If you want a fast bookmark manager / note taker, try MarkLexis. If you want a fast PDF / ebook / comic book reader for Windows, try SumatraPDF.
So you’ve installed Visual Studio and you want to run the compiler cl.exe from command-line. Microsoft makes it surprisingly hard. They give you a shortcut which opens a terminal window with cmd.exe setup for compilation. But I don’t want a separate window, I want to use the terminal app. You can run cmd.exe /k "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build\vcvarsall.bat" x64 (location for your setup can be different). But I don’t want to run inside cmd.exe. I want to use powershell. What exactly does vcvarsall.bat do? Not much: it just sets some env variables and updates PATH. We can reverse-engineer what it does: cmd.exe set >before.txt cmd.exe /k "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build\vcvarsall.bat" x64 set >after.txt Now compare before.txt and after.txt to see what changed. I asked AI to do it for me and here’s the beginning of what I found: CommandPromptType=Native DevEnvDir=C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\ ExtensionSdkDir=C:\Program Files (x86)\Microsoft SDKs\Windows Kits\10\ExtensionSDKs EXTERNAL_INCLUDE=C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\include;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\ATLMFC\include;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\VS\include;C:\Program Files (x86)\Windows Kits\10\include\10.0.26100.0\ucrt;C:\Program Files (x86)\Windows Kits\10\\include\10.0.26100.0\\um;C:\Program Files (x86)\Windows Kits\10\\include\10.0.26100.0\\shared;C:\Program Files (x86)\Windows Kits\10\\include\10.0.26100.0\\winrt;C:\Program Files (x86)\Windows Kits\10\\include\10.0.26100.0\\cppwinrt;C:\Program Files (x86)\Windows Kits\NETFXSDK\4.8\include\um Framework40Version=v4.0 FrameworkDir=C:\Windows\Microsoft.NET\Framework64\ FrameworkDir64=C:\Windows\Microsoft.NET\Framework64\ ... more stuff I saved that to diff.txt file. Now that we have that we can ask our friendly AI to convert this to a powershell function: write powershell function that sets variables as seen in @diff.txt I get something like this: function set-msvc { <# .SYNOPSIS Sets Visual Studio Developer Command Prompt environment variables. .DESCRIPTION This function sets all the environment variables needed for Visual Studio 2022 native x64 development, equivalent to running vcvars64.bat. .EXAMPLE Set-VsDevEnv #> [CmdletBinding()] param() $env:CommandPromptType = "Native" $env:DevEnvDir = "C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\" ... more stuff We need to change one thing: $env::PATH from absolute to just adding paths to current $env::PATH so that it’ll adapt to future changes in PATH. Now add that to $PROFILE and you can type set-msvc to setup powershell to have access to cl.exe, msbuild and other tools. Below is the whole function but note that this is only for amd64 toolset. For 32-bit or arm setup would be different. You can get by re-running the above logic for vcvarsall.bat with different arguments. function set-msvc { <# .SYNOPSIS Sets Visual Studio Developer Command Prompt environment variables. .DESCRIPTION This function sets all the environment variables needed for Visual Studio 2022 native x64 development, equivalent to running vcvars64.bat. .EXAMPLE Set-VsDevEnv #> [CmdletBinding()] param() $env:CommandPromptType = "Native" $env:DevEnvDir = "C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\" $env:ExtensionSdkDir = "C:\Program Files (x86)\Microsoft SDKs\Windows Kits\10\ExtensionSDKs" $env:EXTERNAL_INCLUDE = "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\include;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\ATLMFC\include;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\VS\include;C:\Program Files (x86)\Windows Kits\10\include\10.0.26100.0\ucrt;C:\Program Files (x86)\Windows Kits\10\\include\10.0.26100.0\\um;C:\Program Files (x86)\Windows Kits\10\\include\10.0.26100.0\\shared;C:\Program Files (x86)\Windows Kits\10\\include\10.0.26100.0\\winrt;C:\Program Files (x86)\Windows Kits\10\\include\10.0.26100.0\\cppwinrt;C:\Program Files (x86)\Windows Kits\NETFXSDK\4.8\include\um" $env:Framework40Version = "v4.0" $env:FrameworkDir = "C:\Windows\Microsoft.NET\Framework64\" $env:FrameworkDir64 = "C:\Windows\Microsoft.NET\Framework64\" $env:FrameworkVersion = "v4.0.30319" $env:FrameworkVersion64 = "v4.0.30319" $env:FSHARPINSTALLDIR = "C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\CommonExtensions\Microsoft\FSharp\Tools" $env:HTMLHelpDir = "C:\Program Files (x86)\HTML Help Workshop" $env:INCLUDE = "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\include;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\ATLMFC\include;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\VS\include;C:\Program Files (x86)\Windows Kits\10\include\10.0.26100.0\ucrt;C:\Program Files (x86)\Windows Kits\10\\include\10.0.26100.0\\um;C:\Program Files (x86)\Windows Kits\10\\include\10.0.26100.0\\shared;C:\Program Files (x86)\Windows Kits\10\\include\10.0.26100.0\\winrt;C:\Program Files (x86)\Windows Kits\10\\include\10.0.26100.0\\cppwinrt;C:\Program Files (x86)\Windows Kits\NETFXSDK\4.8\include\um" $env:is_x64_arch = "true" $env:LIB = "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\ATLMFC\lib\x64;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\lib\x64;C:\Program Files (x86)\Windows Kits\NETFXSDK\4.8\lib\um\x64;C:\Program Files (x86)\Windows Kits\10\lib\10.0.26100.0\ucrt\x64;C:\Program Files (x86)\Windows Kits\10\\lib\10.0.26100.0\\um\x64" $env:LIBPATH = "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\ATLMFC\lib\x64;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\lib\x64;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\lib\x86\store\references;C:\Program Files (x86)\Windows Kits\10\UnionMetadata\10.0.26100.0;C:\Program Files (x86)\Windows Kits\10\References\10.0.26100.0;C:\Windows\Microsoft.NET\Framework64\v4.0.30319" $env:NETFXSDKDir = "C:\Program Files (x86)\Windows Kits\NETFXSDK\4.8\" $env:Path = "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\bin\HostX64\x64;C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\VC\VCPackages;C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\CommonExtensions\Microsoft\TestWindow;C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\CommonExtensions\Microsoft\TeamFoundation\Team Explorer;C:\Program Files\Microsoft Visual Studio\2022\Community\MSBuild\Current\bin\Roslyn;C:\Program Files (x86)\Microsoft SDKs\Windows\v10.0A\bin\NETFX 4.8 Tools\x64\;C:\Program Files (x86)\HTML Help Workshop;C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\CommonExtensions\Microsoft\FSharp\Tools;C:\Program Files\Microsoft Visual Studio\2022\Community\Team Tools\DiagnosticsHub\Collector;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\Llvm\x64\bin;C:\Program Files (x86)\Windows Kits\10\bin\10.0.26100.0\\x64;C:\Program Files (x86)\Windows Kits\10\bin\\x64;C:\Program Files\Microsoft Visual Studio\2022\Community\\MSBuild\Current\Bin\amd64;C:\Windows\Microsoft.NET\Framework64\v4.0.30319;C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\;C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\Tools\;" + $env::Path + ";C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\CMake\bin;C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\Ninja;C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\VC\Linux\bin\ConnectionManagerExe;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\vcpkg" $env:Platform = "x64" $env:UCRTVersion = "10.0.26100.0" $env:UniversalCRTSdkDir = "C:\Program Files (x86)\Windows Kits\10\" $env:VCIDEInstallDir = "C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\VC\" $env:VCINSTALLDIR = "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\" $env:VCPKG_ROOT = "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\vcpkg" $env:VCToolsInstallDir = "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\" $env:VCToolsRedistDir = "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Redist\MSVC\14.44.35112\" $env:VCToolsVersion = "14.44.35207" $env:VisualStudioVersion = "17.0" $env:VS170COMNTOOLS = "C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\Tools\" $env:VSCMD_ARG_app_plat = "Desktop" $env:VSCMD_ARG_HOST_ARCH = "x64" $env:VSCMD_ARG_TGT_ARCH = "x64" $env:VSCMD_VER = "17.14.24" $env:VSINSTALLDIR = "C:\Program Files\Microsoft Visual Studio\2022\Community\" $env:WindowsLibPath = "C:\Program Files (x86)\Windows Kits\10\UnionMetadata\10.0.26100.0;C:\Program Files (x86)\Windows Kits\10\References\10.0.26100.0" $env:WindowsSdkBinPath = "C:\Program Files (x86)\Windows Kits\10\bin\" $env:WindowsSdkDir = "C:\Program Files (x86)\Windows Kits\10\" $env:WindowsSDKLibVersion = "10.0.26100.0\" $env:WindowsSdkVerBinPath = "C:\Program Files (x86)\Windows Kits\10\bin\10.0.26100.0\" $env:WindowsSDKVersion = "10.0.26100.0\" $env:WindowsSDK_ExecutablePath_x64 = "C:\Program Files (x86)\Microsoft SDKs\Windows\v10.0A\bin\NETFX 4.8 Tools\x64\" $env:WindowsSDK_ExecutablePath_x86 = "C:\Program Files (x86)\Microsoft SDKs\Windows\v10.0A\bin\NETFX 4.8 Tools\" $env:__DOTNET_ADD_64BIT = "1" $env:__DOTNET_PREFERRED_BITNESS = "64" $env:__VSCMD_PREINIT_PATH = "C:\Program Files\PowerShell\7;C:\Windows\system32;C:\Windows;C:\Windows\System32\Wbem;C:\Windows\System32\WindowsPowerShell\v1.0\;C:\Windows\System32\OpenSSH\;C:\Program Files\Microsoft VS Code\bin;c:\Users\kjk\AppData\Local\Programs\cursor\resources\app\bin;C:\Program Files\gs\gs10.03.1\bin;C:\WINDOWS\system32;C:\WINDOWS;C:\WINDOWS\System32\Wbem;C:\WINDOWS\System32\WindowsPowerShell\v1.0\;C:\WINDOWS\System32\OpenSSH\;C:\Program Files\GitHub CLI\;C:\Program Files (x86)\Windows Kits\10\Windows Performance Toolkit\;C:\Program Files\Rust stable MSVC 1.88\bin;C:\Program Files\PowerShell\7\;C:\Program Files\nodejs\;C:\Program Files\dotnet\;C:\Program Files\CMake\bin;C:\Program Files\Go\bin;C:\Program Files\RedHat\Podman\;C:\Program Files\Tailscale\;C:\Program Files\Git\cmd;C:\Program Files\Docker\Docker\resources\bin;C:\Users\kjk\AppData\Local\Microsoft\WindowsApps;C:\Users\kjk\AppData\Local\Microsoft\WinGet\Links;C:\Users\kjk\.bun\bin;C:\Users\kjk\.dotnet\tools;C:\Users\kjk\go\bin;C:\Users\kjk\AppData\Local\Programs\superfile\;C:\Users\kjk\AppData\Local\Microsoft\WindowsApps;C:\Users\kjk\AppData\Local\GitHubDesktop\bin;C:\Users\kjk\AppData\Local\Programs\cursor\resources\app\bin;C:\Users\kjk\AppData\Roaming\npm;C:\Users\kjk\.dotnet\tools;C:\Users\kjk\AppData\Local\Programs\Antigravity\bin;C:\Users\kjk\AppData\Local\Programs\Zed\bin;C:\Users\kjk\go\bin;C:\Users\kjk\OneDrive\bin;C:\Users\kjk\.bin\jai\bin;C:\Users\kjk\.bin\mupdf-1.27.0;C:\Users\kjk\.local\bin;C:\Users\kjk\AppData\Local\Programs\WinMerge;C:\Users\kjk\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.12_qbz5n2kfra8p0\LocalCache\local-packages\Python312\Scripts;C:\Users\kjk\OneDrive\bin\sublime_text" Write-Host "Visual Studio 2022 Developer Environment configured for x64." -ForegroundColor Green }
TOON stands for a Token-Oriented Object Notation. It’s a new text format that has the same capability as JSON but uses less space. It was invented to lower costs of sending data (tokens) to LLM AIs but it has 2 advantages over JSON: smaller than JSON more readable than JSON It has implementation in many programming languages, including those I care about: JavaScript and Go. Therefore it’s a good use for non-AI cases e.g.: logging structured data sending data from server to client Here’s an example of JSON and TOON formats: { "table": "A", "currency": "dolaramerykański", "code": "USD", "rates": [ { "no": "001/A/NBP/2024", "effectiveDate": "2024-01-02", "mid": 3.9432 }, { "no": "002/A/NBP/2024", "effectiveDate": "2024-01-03", "mid": 3.9909 }, ] } table: A currency: dolaramerykański code: USD rates[252]{no,effectiveDate,mid}: 001/A/NBP/2024,2024-01-02,3.9432 002/A/NBP/2024,2024-01-03,3.9909 As a result, I’m switching to using TOON whenever possible.
More in programming
A clip of me singing a funny song from Gilbert and Sullivan’s Ruddigore back in 2013
In this video, we look at why fork() needs copy-on-write, how it works inside the kernel, and a memory usage problem that Instagram encountered with Python.
Comments require commitment, but they’re worth it.
Basecamp 5 runs on Puma in cluster mode: one master process with preload_app! and 63 single-threaded workers per host, deployed as a Docker container with Kamal. We serve Basecamp from several sites. Each site has its own web hosts and a read replica of the database, and writes go to a single primary database in one of them. On our busiest hosts, each deploy left up to 2,000 requests waiting while the new workers warmed up. We reduced those queues by running signed-in requests through the app in the Puma master, before it forked the workers. Why 63 single-threaded workers? Basecamp has always served web requests from processes rather than threads. It ran on Unicorn, which only does processes, until we moved to Puma in January 2025, and we kept the same setup: workers (Concurrent.physical_processor_count * 1.3).ceil threads 1, 1 preload_app! On a 48-core host that’s 63 workers, each handling one request at a time. We chose 1.3 after benchmarking HEY in 2023, when we moved our apps out of the cloud and onto our own hardware. We tested several combinations of workers and threads with a mix of GET and POST requests on a 32-vCPU VM. Every multithreaded configuration we tested was slower and handled fewer requests than single-threaded workers. Adding workers beyond about 1.2 to 1.3 per vCPU brought little benefit. The threaded workers spent a lot of their time waiting for Ruby’s global VM lock. That made single-threaded workers a good fit for this workload, and we use the same setup for Basecamp. An app that spends more time waiting on its database or other services may benefit from more threads, so benchmark your own app. The other reason is the app itself. Basecamp has class-level state in places and has never needed to be thread-safe. With one request per process, it still doesn’t. Processes do use more memory than threads, and preload_app! reduces the difference. The master loads the app once and the workers share its memory through copy-on-write until they write to it. Shopify’s comparison of Ruby execution models explains the trade-off well. In the HEY benchmark the best setup came to about 260 MB of PSS per core, where PSS counts each shared page once, split between the processes using it, and the gap to a threaded setup was smaller than we’d expected. What Puma does on each host when a container starts: one master, then 63 forked workers that share its memory until they write to it. Two things about this setup matter for the rest of the post. A worker that’s compiling or loading something is fully blocked — there’s no other thread to pick up the next request. And whatever the master has in memory before it forks, all 63 workers share. Whatever they build after the fork, they build 63 times. What happens when we deploy Kamal starts the new container alongside the old one, and kamal-proxy moves the host’s traffic across as soon as the health check passes. At that moment, the new workers have handled health checks but no customer requests. preload_app! means the master loads the app once and the workers inherit it through fork. That covers the code. It doesn’t cover anything Ruby and Rails set up on first use: YJIT compiled code. YJIT compiles a method once it’s been called a certain number of times. The master calls very little during boot, so every worker compiles the same methods again on its own first requests. Compiled templates. Action View turns each ERB template into a Ruby method the first time it’s rendered. The schema cache. Active Record reads each model’s columns from the database the first time that model is used. Inline caches and memoized values throughout Ruby, Rails and the app. All 63 workers did all of this at once, while serving the traffic the old container had been handling a second earlier. In the test environment with YJIT on, the first request to a project page on a cold process took 652 ms, 151 ms of it YJIT compiling. The same request to a warm process took 28 ms. In production, CPU time per request peaked at around 200 ms while kamal-proxy moved traffic to the new container, against about 30 ms once the workers had warmed up. A host with spare CPU absorbs this. Every one of our web hosts has 48 cores and 63 workers, but each Amsterdam host serves around 250 requests per second, against 25 to 60 at our other sites. In Amsterdam the slow first requests turned into a queue. At a peak-hour deploy, the Puma backlog on an Amsterdam host reached anywhere from 250 to 2,238 requests, and kamal-proxy’s p99 response time hit about 10 seconds. Eron, our Director of Operations, had been tracking this since June. Another server in Amsterdam would help, but it would take weeks to arrive, so we also wanted to make deploys cheaper on the hardware we already had. What didn’t work We tried a few things first. In June, Donal tested the first two on a single Amsterdam host, comparing it with its neighbors, and they ruled out two likely causes. Warming each worker’s database connections. Puma’s before_fork hook clears the master’s connections, and each worker opened its own on its first request. Opening them in before_worker_boot instead made no difference. Queries on a freshly booted production host were already under a millisecond, so connections weren’t the problem. A synthetic request in each worker. Next, each worker made a few requests in before_worker_boot to an internal controller that touched every model. That ran the middleware, routing and Active Record paths, but it ran them in 63 workers at once — exactly the CPU spike we were trying to avoid. And a request with no real data renders no real views, so most of the app stayed cold. Spreading YJIT compilation out. Delaying YJIT in each worker by a random interval spread the compiling out over a few minutes, but every worker still ran interpreted until its delay ended. The queue didn’t change. Reforking from a warm worker. This is what Shopify’s Pitchfork does: let one worker serve traffic until it’s warm, then fork the others from it. Puma has an experimental version called fork_worker, and on beta it worked — the reforked workers were warm after three to five requests, where fresh ones took up to 30 seconds. But with fork_worker the template is worker 0, and it keeps serving requests. If it exits, the workers waiting to be forked never start (puma/puma#3596). If it gets no traffic, the refork never happens, which is what we saw on beta. Instacart have a mold_worker patch that promotes a warm worker to a template that stops serving, but it isn’t in a Puma release. We have a branch of it, and we may come back to it. That last experiment did show us where the fix was, though. Everything a warm worker has that a cold one lacks is in its memory, and fork copies memory. The master already has the app loaded. It just never runs it. Run the requests in the master So now, before the master binds its socket and forks, it makes the app’s own requests, in-process, the way a signed-in user would. Rack has a hook for exactly this. Rack::Builder#warmup takes a block that’s called once with the built app, before the server starts. rails server builds the app from config.ru, so the change to boot is one line: require_relative "config/environment" warmup { WarmUp.configured.run } if ENV["WARM_UP"] run Rails.application With preload_app! this runs in the master, and the workers inherit whatever it did. Puma binds its socket after the app is built, so until the warm-up finishes the health check’s connection is refused and kamal-proxy keeps retrying. No request reaches a worker that hasn’t been warmed. The warm-up has three steps. After precompiling the views, it gives the page requests and schema loading a shared 20-second budget, checked before each page or model. 1. Precompile the views actionview_precompiler reads every template for its render calls and compiles each one with the locals it’s passed. For us that’s 1,394 templates in about two seconds. A first request to a project page then compiles 2 templates instead of 44. 2. Request the pages, signed in A small browser class makes the requests through Rack::MockRequest, with the two cookies a real sign-in sets, then goes back for each page’s lazy Turbo frames: class WarmUp::Browser def initialize(signed_in_as:) @client = Rack::MockRequest.new(Rails.application) @headers = { "HTTP_USER_AGENT" => "Basecamp warm-up", "HTTP_COOKIE" => cookie_for(signed_in_as), "bc3.warm_up" => true } end def visit(path) page = get(path) frames_in(page).each { |id, src| get(src, "HTTP_TURBO_FRAME" => id) } end private def get(path, headers = {}) @client.get("https://#{host}#{path}", @headers.merge(headers)) end def frames_in(page) Nokogiri::HTML5(page.body).css("turbo-frame[src]").map { |frame| [ frame["id"], frame["src"] ] } end end The requests are signed in. The user is a monitoring account we already use for automated checks, and the pages are its own project, Campfire, to-dos, documents and messages. Public pages weren’t enough: after warming up with signed-out pages only, the first signed-in request to the projects page still took 131 ms, because authentication, the signed-in controllers and their views had never run. With signed-in pages it took 40 ms. cookie_for writes the same signed cookie the sign-in controller does, using the app’s own cookie jar, so there’s no API token and no secret to store. The frames are followed. The busiest HTML requests in production aren’t pages at all but Turbo frames — the sidebar badge, the inbox, the navigation menus. The browser parses each page and requests its <turbo-frame src> URLs with the Turbo-Frame header, so those controllers and views get warmed too. Our first four pages turned into 60 requests. The requests are excluded from rate limiting. They are internal, so they do not count against the rate limits that apply to real visitors. 3. Load the rest of the schema The page requests load the schema for the models they touch. The last step loads the rest, from the read replica: ApplicationRecord.reading do models.lazy.take_while { time_left? }.each { |model| model.load_schema if model.table_exists? } end The step checks 261 models and loads any schema information still missing. Those database round trips add up when the primary is far away: outside a request, Active Record uses the writing role, and from a host a long way from the primary each round trip is tens of milliseconds. Reading from the local replica brings the step down from about 20 seconds to 3.5. The pages go first because they load most of the schema anyway. If the time budget runs out, the step stops, logs how many models it got through, and the workers load the rest on first use like they always did. Rails can also load the schema from a dumped cache file at boot (bin/rails db:schema:cache:dump), which would make this step unnecessary. We don’t ship one in our image yet, because the dump needs a database to read from at build time, and we have several databases to cover. It’s on the list. What to close before the fork Running requests in the master opens things the master never opened before, and every worker inherits them. Two processes writing to the same socket will corrupt each other’s traffic, so you need to know what’s open before you fork. The way to find out is to list the master’s open file descriptors — ls -l /proc/<pid>/fd — before and after a warm-up, in an environment set up like production. Development wasn’t enough for us: it stores files on disk, so our S3 connections only showed up in production. Then, for each thing that’s open, check how its library handles a fork. We found three kinds: Already handled. Plenty of libraries detect a fork on their own, either by recording the PID they connected from and reconnecting in the child, by opening per-process files, or by resetting their thread pools. Redis clients, metrics libraries and concurrency libraries tend to be in this group. Check, but you probably don’t need to do anything. Already closed. Database connections are the classic one, and most Puma configs already clear them in before_fork. Anything else that’s opened per process — we have a SQLite cache the workers open on boot — needs closing when the warm-up finishes. Needs a new step. HTTP clients with keep-alive connections are the ones to look for: cloud SDKs with connection pools, tracing exporters, error reporters. They usually have no fork handling at all. We empty the aws-sdk connection pools in before_fork, and we run the warm-up untraced so the OpenTelemetry exporter never opens its connection to Tempo in the first place. Once that’s done, before_fork finishes with Process.warmup, which Ruby 3.3 added for this purpose: a major GC, a heap compaction, and every surviving object promoted to the old generation, so the memory pages the workers share change as little as possible afterwards. Choosing the pages The first list was the four pages that ran the busiest requests on beta. Once the warm-up was live, production showed us which endpoints were still cold. For one deploy, we compared each endpoint’s mean duration in the six minutes after kamal-proxy moved traffic to the new container with the same endpoint an hour later, then multiplied the difference by the number of requests in those six minutes. That gives the extra time each endpoint cost us because it was cold: Endpoint Cold Warm Requests in 6 min Extra seconds Campfire 246 ms 70 ms 6,490 1,140 Projects (JSON API) 84 ms 50 ms 22,077 771 Docs & Files 262 ms 177 ms 4,996 421 To-dos tool 205 ms 113 ms 4,018 371 To-dos (JSON API) 33 ms 16 ms 18,738 320 The pages already in the warm-up showed what to expect: the project page kept a 36 ms gap after a deploy, and the to-do page 10 ms. We’ve proposed adding these five requests, and expect them to add about five to seven seconds to the page step. The two JSON endpoints were a surprise. The warm-up’s page list had no API requests in it, so nothing on the API path had run before the first real request: not the API controllers, and not the Jbuilder templates rendering real records. Precompiling the views covers JSON templates too, but it isn’t a substitute for running the request. Results The warm-up is on for all 68 web hosts. With the first four pages it took 12 to 16 seconds per host: about 2 seconds to precompile the views, 7 to 9 for the 60 requests, and 3.5 for the schema. Deploys take that much longer per host, and we raised the deploy timeout from 30 to 60 seconds to cover it. In Amsterdam, at a peak-hour deploy: During deploy Before After Peak Puma backlog per host 250–2,238 requests 19–223 requests Peak kamal-proxy p99 about 10 s 2.4–4.8 s Peak CPU time per request 201–214 ms 88–132 ms Peak database time per request 56–69 ms 39–47 ms The same eight hosts at three deploys on 1 October, an hour apart, as the warm-up went from one host to four to all eight. The deploy in the middle, with four hosts warmed and four not, shows why every host needed the warm-up. Each warmed host recovered faster on its own: mean request duration peaked at 130 to 173 ms, against 203 to 311 ms on the hosts that weren’t warmed. But the backlog on all eight was about the same, because they were all waiting on the same database. Mean request duration on each host at the 07:21 UTC deploy. Blue hosts warmed up in the master before forking, orange hosts did not. Memory came down too. The workers now share compiled templates, YJIT code and the schema with the master instead of each building their own copy. On beta, the view precompiler alone took a busy worker’s private memory from 174–202 MB to 119–135 MB. Thirty minutes after the deploy, the web containers used about 39 GB less memory than the previous day’s containers at the same age and traffic. Amsterdam served most of our traffic at the times we tested. In Amsterdam, each new container used about 2 GB less just after traffic moved to it, which lowers the peak while the old and new containers overlap. Working with Claude Claude Code helped throughout. It combed through the per-worker backlogs and per-endpoint timings in Prometheus and Loki after each deploy, worked out the cold-versus-warm cost of each endpoint, and prepared the changes and the pull request descriptions with the benchmarks in them. We decided what to try, deployed it and read the results. If you do this Warm the master before it forks. Compile common code and templates and load their schema in the master, so workers inherit that work. With preload_app!, Rack::Builder#warmup runs before the workers start accepting traffic. Use the app’s real requests. Public pages, internal endpoints and synthetic queries warm the paths they run and nothing else. Signed-in requests to real records, frames included, run what production runs. Measure the cold penalty per endpoint. The difference between an endpoint’s cold and warm duration, times its request count after a deploy, ranks the pages worth adding. Ours weren’t the ones we’d have guessed, and two of them were JSON. Check what the warm-up leaves open. List the master’s file descriptors after a warm-up and account for every one before the fork. Two of ours needed changes. Set a time budget. A warm-up that runs long on one slow host fails the deploy on that host. Ours gives the page requests and schema loading a shared 20-second budget, checked before each page or model, puts the most valuable pages first, and logs what it skipped. Reforking from a warm worker, as Pitchfork does, solves the same problem continuously rather than once at boot, and it would warm paths no fixed list of pages covers. We may still get there: our branch brings Instacart’s mold_worker up to date with Puma’s main branch and fixes the bugs we found in it. But warming the master works with the Puma we already run, took a few days to implement, and substantially reduced the queues after deployment.
