Andrei Matei

Excited to give a presentation at HYTRADBOI 2025 about what we've been working on!

Ever wanna just know if a line of code is being hit? Or how often? Or from where? In prod!? We've been hard at work making that just a click away with side-eye.io!

@felixge Rando question - is there any code out there for parsing trace files that allows some degree of random access to different time spans, rather than the reader API in x/exp/trace that reads all events? FWIW, for Side-Eye we turn traces into DuckDB, which works out nice for analysis

For the past few days, I've been hacking on a new algorithm for analyzing go execution traces. It seemed great, but fell apart on real data because it was O(N²) and I struggled to figure out how to optimize it. This morning the solution came to me. Thank you Santa 🎅🏻🎁🎉!!

@justinjaffray I figured something out.

@halvarflake FWIW, CockroachDB is written in Go and cares about what's keeping allocations alive - a question for which there's almost no tooling.

Strange thing about memory profiling in GCed languages vs C++ is: In C++ you care what's eating all your memory, in GCed langs you care more "who is causing all the allocations". Different problems...

@MarkCallaghanDB @FranckPachot Postgres lets you query the outcome of a previous txn with txid_status().

This is an important topic. Your network may fail when you COMMIT. Then you retry. Be sure that your transactions are idempotent

@halvarflake @0x0F10 Attributing both IOPS and disk bandwidth to code paths is important for databases.

Perf nerds: What's everybody's view on continuously profiling operations other than CPU cycles in production? Thinking page faults or IOPS, what else is useful? (Forget user space mem profiling for a moment, diff kettle of fish).

@convex Welcome to NY!

Convex is now in NYC! Hit us up if you're excited to work in-person on hard, impactful engineering problems. blog.convex.dev/hello-new-york…

@justinjaffray Elon is on the 10th floor for the midnight meeting.

@largedatabank @justinjaffray Please contribute here for the millennials knowyourmeme.com/memes/lowkey-g…

@justinjaffray please

@asubiotto I don't really know what I'm talking about. The rationalization for LSMs that I've heard was that Google popularized them because GFS is append-only. Otherwise, in CRDB, the read-amp and write-amp are killing us, and the fact that writers externalize the cost to later compaction.

@andreimatei1 Interesting, I want to hear more. Could you elaborate? Twitter might not be the best forum; you have my email.

Are fully in-memory LSMs a thing or is there a better option? We'd still like to compact levels to take advantage of com{paction,pression} & reduce memory size. Thinking of the storage engine as an LSM when doing lvled compaction seems intuitive regardless of the backing storage

@halvarflake Any limitations of the verifier you had to get over.

... about Prodfiler that you are curious about and that you want us to talk about, let me know in replies!

I am happy to announce that @seanhn and me will be giving a talk at re:Invent on the internals of our profiler. We plan on talking about how we do unwinding without frame pointers and symbols, how we deal with the JVM from eBPF and some other stuff. If there is anything ...

Good commentary on the Inflation Reduction Act

@justinjaffray @brandur After Redpanda, I think SingleStore will be the next big WASM user (and they seem to want to go beyond stored procedures github.com/singlestore-la…). Also, Scylla, Postgres, TiDB and CockroachDB seem to all be experimenting with it. It's definitely coming to a DB near you.

Put down some thoughts trying to articulate where I've found it appropriate to push logic down to the database in the form of stored procs and triggers. The opacity of the side effects of such isn't great, but it's a useful technique given some restraint. brandur.org/fragments/code…

@largedatabank @CockroachDB I think you're playing word switcheroo. In an open source context, "freeloading" generally means not contributing back code, not $ (to the extent that it means anything at all). But what we all really want, as you admit, is Amazon's $, not their patches.

The troubling trend I see is megacorps like Amazon freeloading off open source work. I'm proud of @CockroachDB's simple amendment to Apache. It prevents cloud providers from offering CRDB as a service without paying for it. cockroachlabs.com/blog/oss-relic…

@bradfitz I think you can intercept packets/connections at different layers with BPF and keep track of sequence numbers in a map?

As root on Linux, can you get latest TCP sequence numbers of open connections? Without pcap. (so no tcpkill) It's not in /proc/net/tcp, which can be read by regular users. Netlink INET_DIAG? Not seeing there either. I missing something?

@jamesacowling youtu.be/pAD5dz9-Qg8

I've been eating gummy bears my whole life and only just realized that the green ones are strawberry and not "green flavored".

@justinjaffray You could've started reciting some group theory and explain that the cube is just an application.

@natevanben @ajwerner @largedatabank @justinjaffray TransactionRestartWithProtoRefreshError

@ajwerner @largedatabank @justinjaffray They say that about our programs now.