Yaroslav Tkachenko

The Advanced Apache Flink Bootcamp is now open for registration! The first cohort is scheduled for January 21st - 22nd, 2026. #ApacheFlink

@systematicls Just a few more tokens…

Want to get agents to be more reliable? Just throw more tokens at the problem with a fresh context. This is exactly what "code review" is doing. Here's the intuition: agents can catch mistakes they missed in the first pass in the second pass. QED.

@trevoro I can hear the whips.

Our shift from engineering to agents is less "weavers to looms" and more "horses to cars". We're the horse.

In the latest edition of the Data Streaming Journal, you can read everything about my benchmarks: which technology is the fastest, by how much, and what you can do to bridge the gap. streamingdata.tech/p/benchmarking…

Can a Rust project be 3x faster than a popular Java project right out of the box? I’m a big fan of RBIR: Rewrite Bigdata in Rust. So when Supermetal recently launched Kafka sink support, I decided to take it for a spin and compare its performance with Debezium and Flink CDC.

@richardartoul Code is a liability, so having friction is great

Wrote a coding agent for WarpStream that runs in our self-hosted github actions runners, and the bottleneck now is validating the result in our test/staging environments. Validation takes 10-100x longer than code generation.

@danshipper Interview starts, Anthropic goes down… what then?

the new "technical" interview: - build: a feature with your AI agent of choice on your own time - submit: a PR, a loom of your feature, and the record of your agents sessions - talk: get on the phone and talk through the decisions you made WITH ai, show us how you chat with it / how you answer questions etc, as if we werent looking. "cheating" encouraged

Engineers don't watch this talk enough and it shows (it's not about Python btw) youtube.com/watch?v=RZ4Sn-…

Current status: async-profiler: 1 - Claude Code: 0.

Since my feed is full of people who didn't code in months because Opus 4.5 is so good, let me chime in with 5 things that I had to manually fix this morning because Claude Code didn't: - Separation of concerns: I asked for new functionality that belonged in a specific class. I told Claude to review that class and make the changes. It made the changes in an unrelated class and then moved some functionality from the class I wanted into a second unrelated class. - Encapsulation. A bunch of private and package-private methods were made public with no good reason. - Concurrency: It implemented a rather blatant concurrency bug and kept insisting that the logic is correct. - Reinventing the wheel: The concurrency bug was in a queue library that didn't need to get written because Guava exists. - Use the wrong build command, then misinterpret the error and make a bunch of totally unnecessary changes. In a loop, since the wrong build command kept failing after each round of useless changes. And yes, I have plan documents, rules document, and I prompted very carefully. Claude Code is impressively good, but human review is very much needed. Especially for concurrency.

@gwenshap I loved most of the papers from 15721.courses.cs.cmu.edu/spring2023/sch…

I want to do some deep reading this holiday season. What were the best technical blogs and papers you’ve read this year? Send me all the recs! Feel free to plug your own blog if you are super proud of it! I will publish a summary for everything I read.

🎄Interesting Links in the data world for December 2025 is here! There's some awesomely excellent content that you'll not want to miss. Grab a mince pie and dive in :) 🔗 rmoff.net/2025/12/16/int… Thanks to all the folk out there writing these great blogs posts, including @SandonJacobs, @themoah, @AdamSouquieres, @mehulbatraa, @sap1ens, @fnthawar, @gunnarmorling, @sspaeti, @markrittman, @J_, @richardartoul and many more!

@iambriccardo Of course I meant without the 2/3 cons 🤦

@iambriccardo Valencia is almost as good, and without the 2/3 pros you mentioned (it's still smelly). But another con is that Spanish becomes a must-have.

My 2 cents on Barcelona after a week there. Pros: - The vibe of the city is great. - Tech scene is growing. - Food is good. - Weather is really good. - Gaudí works are insane. - Great public transport. Cons: - Expensive for what you get. - Smelly as hell. - Too many people.

How it started -> How it is going

wake up babe, new stream-processing framework just dropped

The most destructive belief in the world is that sleep deprivation produces better results.

Do you like databases? Do you want to hear two database professors rant about them? Do you need one of those professors to have a Turing Award for databases? If yes, then join Mike Stonebraker and I next Wed Dec 10 @ 1:00pm EST for database hot takes: dbos.dev/webcast-2025-i…

btw, there is nothing better than Neue Deutsche Härte for locking in

I guess it's that time of the year

@criccomini @kozlovski This is where we disagree 🤷

@sap1ens @kozlovski I think the entire stream processing ecosystem is niche. Selling it is a sliver of that niche.

"Q3 2025: out of a ~$1.1B annual recurring revenue, Flink makes just ~$14M ARR6 (1.25%)" @kozlovski buried the lede in this one. Stream processing is just not going to work as a business. It took me a long time to come to terms with this, but I believe it pretty strongly now. There is too much fundamental, irreducible complexity. Windowed aggregation, state, streaming SQL, and more. It's just too hard for mere mortals trying to get their job done. The simplicity of batch is hard to beat. Then you overlay market fragmentation and it gets even worse. Production stream processing looks very different from business analytics stream processing. It's hard to build a system that can do both, yet you need to address both to have even a chance at a reasonable TAM. The numbers just don't add up. It's a big honey pot for engineers, too. So many fun things to work on. I spent years on Apache Samza, learned a ton, and had a blast. bigdata.2minutestreaming.com/p/event-stream…