Proud contributors
We just crossed 800 Github ⭐️ !! and v0.32.1 of parquet-go is out, which includes changes from 11 new contributors!
Seeing so many users send bug fixes is the best part of maintaining it
github.com/parquet-go/parque…
Perfloop retweeted
In September I made a chat thread my VP of Engineering. It built an org of twenty agent threads.
In October it audited the bureaucracy it had created, handed off its work and fired itself. That took 47 minutes.
Last week the org that replaced it merged 1,119 pull requests.
Fixing the World
Perfloop found a dangling stack pointer in the Go runtime. It made recover miss a panic.
In September, Go tip made recover work inside range-over-func loop bodies. The loop body keeps the address of the panic record to do that. But that record lives on the stack, and Go moves stacks.
So if the loop body escaped to the heap and the stack got copied, the address went stale. recover returned nil and the panic kept going.
Perfloop filed it with the cause and a small program that fails on tip. The Go team marked it NeedsFix. @cuonglm_ found it also happens on a shrink and sent the fix the next day. David Chase and Michael Pratt reviewed it, and it's now merged.
1
101
💯💯💯
Perfloop just hit 100 changes merged into open source!
Started with parquet-go in June. Torbjörn Einarsson merged number 100 into mp4ff on Tuesday. Two more since then, so 102 now in 25 projects.
And number 102 is in Postgres. Perfloop found a memory leak in the checkpointer. Freeing a WAL reader could leak the last record it decoded. The checkpointer doesn't reset that memory in normal operation, so big PREPARE records piled up. Postgres takes patches by email so I was the meat proxy. Fujii Masao reviewed and committed it. Backpatched to 15.
Thanks to all the maintainers who merged these. You're in the credits at the end of the video.
47
Perfloop retweeted
Prometheus merged another Perfloop PR this morning.
PromQL range functions like predict_linear and quantile_over_time can take scalar arguments. For every argument, on every series, at every step, the evaluator built a fresh one-sample vector.
Now each slot is set up once and overwritten before each call. On the range query benchmark Prometheus already had, at 100 series and 1,000 steps, that's 52% fewer bytes and 89% fewer allocations.
Thanks @bboreham, Julien Pivotto and @bwplotka for the review.
🪲🪨
Cpp2Rust merged its first Perfloop PR. This one is a bug fix, not a speedup.
Cpp2Rust translates C++ to safe Rust. A C++ std::map uses its ordering (<) to decide if a key is already there, not ==. Usually the two agree. When a type defines them differently, say it sorts by name but == compares IDs, the translated map[key] went by == and got it wrong. It could hand back another key's value, or replace an entry it should have reused.
Claude Opus wrote that code and reviewed it, and it went through human review too. You can spend a lot of tokens on review and bugs like this still get through.
Perfloop is a safety net behind review. It's always on, hunting your code for bugs and performance wins, and it hands you the change with proof, made the way your team wants it. It learns your environment and your preferences as it goes.
This PR came with a test that runs the C++ and the translated Rust side by side. It fails before the change and passes after.
Thanks Nuno Lopes from Lisbon's IST and Furiosa AI for running Perfloop on Cpp2Rust.
55
Perfloop retweeted
Agentic Performance Engineering is out now..
Timestamps:
00:00 The performance problem behind Perfloop
00:39 Why performance work is a battle
03:59 How AI agents raise the stakes
08:55 Quality control at the scale of agentic development
10:34 How Perfloop maps systems and finds opportunities
14:32 Using telemetry to prioritize hot paths
18:13 Perfloop’s open-source performance wins
22:18 How Perfloop measures and verifies improvements
25:11 Common performance patterns and bottlenecks 28:47 The “apex hunt” for bigger architectural gains 30:58 Initiatives and setting high-level performance goals
35:21 Connecting code optimizations to customer impact
39:59 Why independent verification matters
44:36 Choosing the right AI models for each task
49:45 What a harness does in Perfloop
53:59 The possibility of pre-merge performance reviews 57:24 How AI may change the role of performance engineers
63:54 Why performance requires measurement
featuring @tsenart from @perfloop
Perfloop retweeted
Software Engineering is NOT DEAD, but it has changed forever. But how about Performance Engineering?
Can you use Agents to find performance bottleneck, find the root cause and fix it for you?
But how would the approach look like? another prompt?
How do you do the verification?
most importantly how do you avoid Performance engineering SLOP?
Checkout the latest episode on The GeekNarrator with @tsenart who built @perfloop to enable Agentic Performance Engineering. Link in the comments.
Also like, share and repost to make it visible to folks who care about Performance.
Perfloop retweeted
Pipecat, Daily's voice agent framework, merged two Perfloop contributions.
The first is chatty I/O, the pattern where something sends way more messages than it needs to: the VAD controller was firing a speech event on every audio frame, when once every 200 ms was the intent, and each one fanned out through the LLM aggregator as a user-speaking frame. One line changed, 17% less CPU on the VAD pipeline.
The second is an avoidable copy: the audio volume tracker was rebuilding its 400 ms window from scratch on every chunk. 48.5% fewer bytes copied per update, allocations from 2 to 0.
Perfloop retweeted
Writing code got cheap, and attention did not. A system stays fast only while someone holds its envelope, and where agents write the code, holding it takes superhuman attention.
A freestyle coding agent runs on your attention. What changes the accounting is a harness every change must survive before a person ever looks.
perfloop.ai/blog/superhuman-…
🤝
Replying to @TigrisData
@TigrisData is now a @perfloop customer.
Tigris met Perfloop as a pull request: an optimization our system opened in their open source acceleration gateway, tigrisdata/tag, with the evidence attached.
@ovaistariq read it and merged it, then asked on X: "How do I get access :)". He signed up the next evening.
Within a week he had connected repos himself and brought his cofounders in. In the two weeks since, Tigris has merged sixteen more of Perfloop's pull requests, four of them in tag: github.com/tigrisdata/tag
1
367
Perfloop retweeted
Just got some nice perf work stream done with @perfloop and Claude working together github.com/tigrisdata/tag/pu…
Perfloop retweeted
News flash! First public Perfloop PR not driven by me just landed (link in reply).
I'm giving direct access to Perfloop to open source maintainers who care about performance, in return for their feedback.
@__Achille__ is driving it on his repos including parquet-go, which we had landed many contributions on already via our OSS tenant. Now he has his own.
We’re all about chop chop.
I'm that boss telling my agents to chop chop.
Many people have had the failure mode of over-engineering and defending against an ever larger bucket of corner cases.
And boy oh boy, isn't GPT 5.6 Sol mimicking that behavior.
That's one of my next challenges: how to keep automated review loops but cut them short at the right point of pragmatism.
1
548
Perfloop retweeted
See you in October! May the tail be with you.
We can't wait to learn about a software loop that does the full arc of performance engineering from @tsenart at this year's #P99CONF. Don't miss your chance to learn what this looks like for teams running performance-sensitive workloads: ow.ly/P68h50ZnQHc
#ScyllaDB

















