brotli rob

@annikaze_ pretty good commit to LoC ratio

excited and humbled to announce that I started a new role as junior software engineer - posthog code mobile app

(1/2) Engineers often face a trade-off between scaling fast and maintaining low latency. @P99CONF is designed specifically for those navigating the complexities of high-throughput, low-latency systems. #ScyllaDB #P99CONF

@IanVanagas @posthog and if you want to see what we've done post-rewrite to tune the system as the scale hit the projected numbers, check out this post from one of my teammates! posthog.com/blog/untanglin…

Feature flags are high stakes. An outage affects not only our customers, but our customer's customers. When flags are slow, our customer's entire application slows down. Unreliability doesn't just make us look bad, but our customers too. Our team at @posthog has done a lot to make flags more resilient through local evaluation, caching strategies, and architectural improvements, but as we’ve grown we hit new scaling challenges that required the team to do a more fundamental rework of our flag service. This meant: - Rewriting in Rust (with axum). The team chose this over FastAPI and Node.js. Rust's axum framework was achieving ~32k requests/second, a 21x throughput improvement on Django. The type system made it harder to write buggy code. Most importantly, it gave them proper code-level timeout primitives, eliminating our problematic dependency on PgBouncer-level settings. - Moving evaluation logic out of the database. PostHog now fetches all relevant person properties, evaluates flag conditions in memory, and returns results without database round-trips. - Implementing app-level cohort caching. Cohort-based flags were some of our slowest due to complex joins. PostHog doesn’t compute these on demand anymore. - Simplifying the architecture. The team removed PgBounder entirely. The new service connects to the database directly. All this led to serious improvements: - p99 latency when from 904ms to 85.4ms (10.6x faster, 90.5% reduction) - We reduced costs 68%, dropping the pods required to handle ~500k requests/minute from ~300 to ~90. - Feature flags didn’t have an outage for 3 months after the migration to the new service Curious to learn more, check out @dmarticus's post on how we made feature flags even faster and more reliable → posthog.com/blog/even-fast…

@IanVanagas six seven

It costs PostHog $6.67 to successfully onboard a new user with our AI installation wizard. For this cost, it understands your business logic, captures high quality data, and correctly configures PostHog products in a single command. In one way, this is incredible value. It’s saved thousands of engineering hours, improved onboarding conversion, and is ultimately a net positive. In another way, it’s still A LOT of money. $6.67, that’s like 2 cheeseburgers or a full game on Steam. And more than 90% of users use PostHog or free. The wizard’s success was hiding a lot of waste, so Vincent investigated. He followed the tokens, investigating context explosion, caching, subagent termination, and more before discovering the agent’s real problems: 1. Repeatedly re-reading the codebase even though it already had the context it needed. 2. Forgetting that it had already compacted the context and re-reading files again. 3. Agent spawning multiple subagents that fail to return or timeout. Fixing these meant fewer token reads and subagents, lowering costs while keeping efficiency. It was also a lesson not to take your AI agents at face value. To get the full details on the investigation and everything Vincent learned, check out his blog on “How we caught our AI agent embezzling tokens” → posthog.com/blog/optimizin…

@blisstweeting check out posthog.com/careers

Hey folks! I'm looking for an ambitious project to work on. I've got 15 years of experience across several tech domains, use AI extensively, and have a good eye for product thinking. The kind of generalist everyone needs these days Resume in the next tweet

@gogainda @posthog nah, we hadn't implemented cohort operations in our MCP yet – just did though! github.com/PostHog/postho…

PostHog mcp is promising but missing some write functions : closing errors, creating cohorts @posthog am I missing smth?

@bengold @james406 it's effortless for him

@james406 Why do you have so much time for shitposting when you run PostHog!?

ai is crazy. friend's startup got acquired recently got a surprising windfall of $500k, wanted to put it into a down payment for a house i said that was really stupid and they should instead buy 10-15 nvidia b200 tensor core gpus "think of how many things you could do with that. you might even be able to train a frontier model" unfortunately by this point i realized they left the conversation about 30 seconds earlier and this is why ai is amazing, because claude will always hear out my great ideas. real-life people won't there's a lesson there. i'm not sure what it is, but it's there

@ItsPatrickU thank you!

@dmarticus Sent ya an email ✌️

Update: planned out a solution last night. Sent codex and opus off to implement, spent 2 hours this morning reviewing, refining, and cleaning up and we merged it.

@ItsPatrickU I appreciate it! I'm dylan at posthog dot com if want to email me, feel free to reach out on LinkedIn too (if you'd like!)

@dmarticus Let me put together some notes and I’d be happy to connect! We really like Posthog and still considering to roll it out further, it’s been just FF that sadly gave us too much trouble.

@ItsPatrickU either way, I'd love to understand more about why you wanted to switch, and while I don't expect to win you back given the effort you've put into building out your own solution, I'd love to understand how to improve for future users

@ItsPatrickU and we've stood up a whole team dedicated to scaling out our solution so that we don't run into issues like we've had in the past (e.g. this github.com/PostHog/post-m…) was there also something about our billing model that didn't work?

@adamleithp can't beat blog.jim-nielsen.com/2024/sanding-u… IMO (i think this is a good thing)

@adamleithp UI polish has been really hard for me to nail re: AI-assisted stuff; e.g. for backend things you can use property tests, strong types, etc to encode invariants, but UI is something that oftentimes requires a lot of elbow grease because you don't know it before you see it

I had a coworker reach out to me and ask “do you use AI?” Yes, “ok so how do you make everything pixel perfect with it?” I don’t. These tools, despite all the skills and MCPs, still don’t have taste. Please tell me what I’m doing wrong?

@adityamaru27 @posthog fixing here github.com/PostHog/postho…

.@posthog does so much right but choosing an "enabled" feature flag as the icon for all their feature flags is chaotic evil

@adityamaru27 @posthog ooof, that is good feedback. Let me see if we can come up with a better asset here

wrote about work #fn-1" target="_blank" rel="nofollow noopener">posthog.com/blog/even-fast…

@nwilliams030 Seterra is goated for this

gotta memorize all the flags, countries & capitals i am a 27 year old woman. no reason not to simply buckle down and memorize all the flags, countries, and capitals

@datyayu We’re workin on it!

posthog are you ok?

@datyayu No :( status.posthog.com/incidents/q3dq…