swe@PingCAP - The company behind TiDB. Oracle/MySQL/InnoDB team lead in a past life
California, USA
- Tweets 4,771
- Following 261
- Followers 5,384
- Likes 4,055
Using flags in programming conversations has more or less disappeared, except when describing legacy system interfaces. Now the same bits usually sit behind an enum, a typestate, or a named option, and the design gets described in terms of those states and transitions.
I think that is progress.
2
12
817
Sunny Bains @TiDB retweeted
Tomorrow: live TiDB lab on MySQL fleet consolidation.
Resource Groups, cost modeling, and Plaid's 234-database story.
Sep 24 | 10am PT → pingcap.com/event/tidb-use-c…
3
6
390
This is what the full insert test run graph looks like. I want it to be stable after peak. the drop at 8k connections isn't too bad but it bothers me :-). it's 1.7M/s.
1
10
752
I think a requirement to use specialized malloc libraries is a sign of poorly written software.
Memory management and its performance implications should not be an afterthought.
7
20
3,567
Weekend hacking.
Replacing Tokio with a custom runtime for a storage engine use case, the db engine can now do 2.1M row insert/s with 10 rows per insert. Stated in a less glamorous way, 210K inserts/s 🙂at 512 concurrent threads.
Very pleased.
7
3
111
12,131
Some people like to take a single node database, build their own distributed transaction layer on top. So far so good. Then compare latencies of their single node underlying database with another database that provides strong consistency through distributed transactions and horizontal scalability out of the box.
2
29
2,231
Sunny Bains @TiDB retweeted
I am giving a keynote next Tuesday Sept 22 at the "Rows & Columns Summit" on the torrid history of hybrid transaction/analytical processing (HTAP) database management systems: Broken Dreams, Broken Promises, Broken Marriages.
rowsandcolumnssummit.com/
1
19
134
9,487
This is a very interesting paper: BTRLoG: Low-Latency Logging for Cloud Database Systems
arxiv.org/pdf/2606.27051
3
10
135
7,495
This is a great article. I think it’s more general than just research teams.
Every research team runs down some long dark alleys of the soul. Here's Serdar Benderli, our Research team manager, talking about some of the failure modes he's/we've seen -- hopefully this blog post saves yours from a few.
antithesis.com/blog/2026/way…
1
12
2,640
Databases and storage are the shovels of the AI gold rush.
They persist context, serve retrieval, coordinate state, and turn inference into a reliable system
1
6
71
4,785
I reread the DSQL paper on a recent flight. Viewed through an Accord lens, something that I'm interested in these days, the interesting distinction is DSQL’s disaggregation of the protocol. Accord targeting Serializability and DSQL targeting (S)SI aside.
Accord combines conflict discovery, transaction ordering, and quorum replication in a leaderless, per-transaction consensus protocol with a one-RTT fast path. DSQL separates these responsibilities: sharded adjudicators validate conflicts and constrain commit timestamps, while a selected Journal atomically persists and replicates the transaction and provides the durable order.
2
49
4,088
You need this attitude. You need to just do it, there will always be people wagging fingers, pontificating or giving sage advice. Ignore it.
It’s basically how anything interesting gets built. Nobody who shipped something interesting got there by first securing consensus from the peanut gallery.
Almost every single time I released something in my career, I see people saying: you are clueless. You don't understand the space. You don't know what I am doing.
In the early days of the Turso Cloud, there was this guy here that would every now and they remark that "Turso keeps building the wrong thing", always followed by an explanation of what SQLite *IS*, what is good at, what is not good at, how people use it, etc. Turso now has almost one signup every 5 minutes, and 700,000 databases were created in our cloud this week.
On the rewrite of SQLite itself, the common accusation was that we didn't understand that people use SQLite because of its reliability, and any rewrite would fail because it had no access to their test suite. We took Turso out of beta last week and power many production users already.
Before that, it was Scylla, a rewrite of Cassandra. We were roasted by many, who said that we were morons who didn't understand that Cassandra was I/O bound, so any improvements in the CPU architecture were pointless. Scylla was 10x faster than Cassandra in real life.
My first experience of that was when at Red Hat we patched QEMU to be the virtualization layer for the device abstraction for the KVM Hypervisor. I was 19 at the time, and being part of the team I submitted a talk to a local conference, who rejected me saying that QEMU was an emulator, and I didn't even seen to understand the different between virtualization and emulation if I was conflating the two.
The common thread in all of those is always the same: people getting dead set in their ways, always looking at the world as it is, completely unable to understand that if you just tweak things, most things can be very adaptable. Things can reach new heights. They can be used in different ways.
2
5
60
3,827
The framing sections are genuinely waffle. "We've found a way to make sand think," "10x the Industrial Revolution at 10x the speed," "foothills of the singularity" - none of that is falsifiable or even operationalized. "AGI is probably only a few short years away" does the classic move of making a huge claim while the word "probably" and the undefined term "AGI" give it infinite escape hatches. If timelines slip, nothing in this essay is wrong, which tells you it wasn't really a prediction. And the closing section is pure incense: golden age, human flourishing, the future is not yet written.
1
16
2,328
Nice article, if you use <tenant id, database name> for example, store the segments on separate disks if required, it should be able to isolate the impact of compaction.
I can't recommend this blog post enough if you're interested in databases. It's top tier technical writing and is an excellent first-principles exposition on improving compaction in SlateDB.
slatedb.io/blog/segment-orie…
3
16
3,966
Preliminary results of the custom runtime are pretty good. Worth the effort. Need to run two more performance tests to be convinced that it works for all loads.
3
12
1,457