Simon Eskildsen
13.8K posts
Simon Eskildsen
@Sirupsen
co-founder & ceo @turbopuffer, former infra @Shopify 1k→1m RPS
🇩🇰 → 🍁🇨🇦🌲 Katılım Nisan 2009
808 Takip Edilen14.6K Takipçiler
@Sirupsen This convinced me to book some time next week! Maybe Layers on Turbopuffer soon
English
one of my favorite examples of how agents are fundamentally changing trade-offs is that they write longer text queries, which e.g. turbopuffer takes advantage of in our FTS engine
turbopuffer.com/blog/fts-v2-ma…
English
Simon Eskildsen retweetledi
“make things easy to react to”
this is a core principle of work at @turbopuffer. when you present information to the team for review, load all necessary context into the slack message or notion doc. time is our most precious resource; don’t make people dig.
this is *really* good generalized advice for gtm.
your landing pages and cold DMs should contain all the context your prospect needs to make a decision.
don’t offer to “put some ideas” in front of them or ask to “explain in a follow up call”.
just do it right there.
English
was such a treat to get an hour with @sualehasif996 ❤️ honoured to be part of their journey!
Sualeh Asif@sualehasif996
A conversation with @sirupsen on scaling Shopify, building turbopuffer, and the future of databases. 0:00 - Scaling Shopify through flash sales and outages 8:13 - How top infrastructure teams collaborated in the 2010s 10:35 - Engineering principles from Logrus and on-call 17:38 - The story behind Simon’s famous-ish blog, Napkin Math 23:05 - Why new database companies keep winning 32:21 - How Simon became a fan of databases 35:45 - AI coding, and where agents still fail 42:10 - Hiring P99 engineers in the AI era 48:45 - What’s next for databases
English
@Sirupsen @turbopuffer don't have a use for it yet, but may soon
English
turbopuffer tagline should be 'puff around and find out'
@turbopuffer
English
English
the most egregious case of cultural appropriation is the North American WASP take on charcuterie boards
English
one of my favorite things about onboarding is working with the team to come up with a bio for the newcomers :) @_sebgl @mashbourne
turbopuffer.com/about
English
@tomhaerter @Sirupsen s3 backed lsm kv-store and potentially and separately a vm scheduler
English
a h4x day in Banff project by @jpountz, on the same underlying inverted index and ranking as bm25, regex, etc., so you can combine it all 🥺
turbopuffer@turbopuffer
new: sparse vectors a first-class retrieval primitive that composes with BM25 + attribute ranking in the same query plan, no client-side fusion needed for SPLADE / learned-sparse retrievers (or roll your own weights for custom feature scoring) docs: #sparse-vector-search" target="_blank" rel="nofollow noopener">turbopuffer.com/docs/query#spa…
English
@zoltanarvai @davidcrawshaw No it’s not apples to oranges on every dimension :) if you’re reading off S3 or east-west DRAM caches, this matters!
English
@Sirupsen @davidcrawshaw Faster than what? Hitting the nvme controller? Because network does not persist anything so it’s an apples to oranges comparison.
English
re: NVMe vs S3, it's quite simple:
NVMe random read (8KB): 100 μs
S3 random read (128KB, p99): ~80ms
On throughput, you can find instance types with higher networking throughput (favours S3), or higher NVMe throughput.
English
@davidcrawshaw That’s it!! I think it’s counterintuitive to most that networking can be way faster. It’s also cheaper, if zonal.
English
@Sirupsen For throughput my favorite mental model is counting PCIe lanes. You can dedicate as many or as few as you like to networking or NVMe. (I wish I could convince vendors to not bury this info so deep.)
English
with this, you can index all of commoncrawl into tpuf tomorrow and query it at 1,000 QPS with great economics. some of our customers do.
English
the default "puff into {nvme, dram}" works well for most use-cases, but in cases of high, sustained QPS—coupling compute and storage is best
this is what we mean by "compute-storage flexible"
turbopuffer@turbopuffer
new: namespace pinning pin namespaces to reserved compute for faster, cheaper, more predictable p99 on high QPS workloads ~50x cheaper than shared compute for a 128GB namespace at 500 QPS docs: turbopuffer.com/docs/pinning
English
@thegeeknarrator transparent pricing and discoverable cost calculator
English
@Sirupsen Nice. Naming is hard. Serverless is always confusing, like cloud-native, ai native etc.
Curious what alternative way did you find to communicate the serverless billing model?
English
removed all mentions of serverless on the turbopuffer website. what does it mean anyway
English