Head of Developer Advocacy & Docs at Tiger Data
Chicago, IL
Joined December 2006
- Tweets2.1K
- Following1.4K
- Followers15.5K
- Likes79.1K
Marino Wijay (@virtualized6ix) on the show: "Tech has been the largest accounting hack ever." His case: VCs hedging across five companies on the same non-problem, and a tax structure that rewards hiring tech people either way. arresteddevops.com/wtf-is-go…
"You're comparing your own friction points with everyone else's highlights." @DevOpsJacquie on why every launch makes it look like everyone else learned the new thing last night. From the episode: arresteddevops.com/meaningfu…
New on Arrested DevOps: @DevOpsJacquie and I talk about keeping meaningful control over the tools you depend on. Know what they're tied to, verify they do the job, and have a realistic way out.
arresteddevops.com/meaningfu…
pgvector 0.8.7 fixes CVE-2026-103484. Any database user who can create an IVFFlat index can write out of bounds in the backend, which can get you to arbitrary code execution. Every version through 0.8.6 is affected.
thebuild.com/blog/the-vector…
Upgrade. That part’s easy. Christophe Pettus spends the rest of the post on the harder questions, mostly: who can actually reach this in a real deployment, and what does “a user who can create an index” mean when your app has one database role that does everything.
I’ll argue for consolidating onto Postgres in most rooms I walk into. This is the invoice that comes with that position.
Your attack surface is the union of every extension you installed. Each one has its own release cadence, its own maintainer bandwidth, and its own relationship with the word “soon.” Core Postgres has thirty years of security process behind it. The extension you added in a sprint last spring might have one person with a day job.
That’s a tradeoff I’ll still take. It’s not a free one.
Do you know the installed version of every extension in your production cluster right now, or are you about to go find out?
Your nightly backup is evicting your working set and you probably haven’t measured it.
clickhouse.com/blog/direct-i…
A backup reads every byte of a live database off the same drives answering your queries. By default Linux caches all of it, in case someone asks for those bytes again. Nobody will. To make room, the kernel throws out pages Postgres was about to use, and the next query goes to disk.
Their test box: 48 vCPU, four NVMe drives, 467 GB database under live load. The buffered backup evicted all 40 GiB of a warm table that was sitting idle and not involved in the backup at all. With direct I/O: evicted nothing, finished in 71 seconds, roughly two thirds less latency impact, about 14% less CPU.
The catch is what makes this cool. Direct I/O also turns off kernel readahead, so a small read keeps one drive busy while three sit idle. You have to size each read to span the whole RAID0 stripe and scale reader count to the hardware. The flag alone makes things worse.
Nice writeup. Go look at what your backup window does to your cache hit rate.
Matty Stratton retweeted
I got to help with the pg_vectorsearch launch, our new OSS Postgres vector index.
The days were long but the weeks were short, very excited to have this out there.
100M vectors on a 128 GiB machine: built in 1h40m, 99% recall at 21 ms. pgvector’s HNSW couldn’t even build there.
Lots more impressive stats in the post, along with a deep dive of the architecture for how we got the results:
tigerdata.com/blog/announcin…
one thing i like about our team here at Tiger Data: when we find a better way, we're willing to start over.
today we're releasing pg_vectorsearch, an open source Postgres extension for vector search. alpha, but real. 🧵
pg_vectorsearch's PRISM index partitions and compresses vectors instead of building a graph. that means way less RAM and way faster builds.
at 100M vectors: 6 ms queries on a 128 GB machine.
HNSW couldn't finish building on that box. on a machine with 4x the memory, it took 42 hours.
it's alpha, PG 18+, works with pgvector types, and it's PostgreSQL licensed.
point it at your own data and tell us where it breaks:
tsdb.co/raa9g5ye
Someone built a Postgres index advisor, then built every index it recommended and timed the queries. Including the ones that embarrassed the tool.
prateek-arora.github.io/2026…
179 measured recommendations. 128 at least 15% faster. 32 at least 5% slower. Seven more than twice as slow. The worst estimated 91% cheaper and measured 0.93 seconds to 7.3 seconds.
The reason is structural. HypoPG validates a suggestion by asking the planner if the hypothetical index lowers the cost. But the planner’s row estimates are often the reason the query was slow in the first place. You’re asking the confused party to grade its own confusion.
This is the part of the optimization treadmill that’s hardest to see from inside it. You ship a change, the estimate improves, the dashboard looks better for a while, and you never find out whether the fix was real. At Tiger Data we talk a lot about the difference between operational friction and architectural friction. What to test it cheap? If you can’t measure whether your last five optimizations helped, you don’t yet know which kind you have.
Estimates are free. Measurements are not. Spend accordingly.
Matty Stratton retweeted
The "software is dead" people are missing sufficient empathy.
It's a natural conclusion as a software builder, someone fascinated by computers. But only 0.5% of people are "software engineers."
The remaining 99.5% very well may prefer non-generative ux, even in the long-run.
Matty Stratton retweeted
This next decade is seeing the explosion of data and intelligence around the physical world.
And we are seeing an explosion of @TimescaleDB alongside it. Powering historians, the edge, infrastructure, and more.
Lots of hard, exciting engineering problems to solve. Join us. DMs open. 🐯
Matty Stratton retweeted
23 out of 23 system integrators in @WalkerDReynolds' Industry 4.0 Growth Program are using @TimescaleDB.
A pretty remarkable observation from Walker on a recent 4.0 Solutions podcast about how the industrial historian market is changing:
"If you look at our industry ... it really is staggering to me. The industry 4.0 community is over 100,000 people. All of them are using time-series in some way. All of them used to be buying historians. And now this is crazy: none of them are buying historians.
If you look at the 23 integrators we have in the industry 4.0 growth program: all of them are using TimescaleDB. Every single one, their defacto is TimescaleDB.
10 years ago everybody bought a historian. Then in the last 5-7 years, everybody defaulted to, 'I'm going to build my historian using TimescaleDB as the backend.'"
It's exciting to see how much the modern industrial data stack is changing.
Marino (@virtualized6ix) on why every AI-built site looks the same: "I'm looking at the same set of swatches." arresteddevops.com/wtf-is-go…
New Arrested DevOps: no topic, just me and Marino Wijay (@virtualized6ix) asking what is actually going on in tech right now. Zero interest rate money, and infrastructure work collapsing from three layers to two. arresteddevops.com/wtf-is-go…
New Arrested DevOps episode with @virtualized6ix is live!
arresteddevops.com/wtf-is-go…
