Re: Tech Entrepreneurship in Canada and the increase in capital gains tax

problem is the IMMENSE pressure to incorporate tech startups in Delaware

What tips the scales to Canada is SRED (AFAICT). Be careful 🇨🇦

good content for the fellow blob storage fans out there 🫡

ice out in a week or two means spring 🧊 ☀️

soon all of turbopuffer will be zero-copy

shadowing production queries, including all their filtering conditions!

jli was dissatisfied that the previous cache did not saturate NVME. here you are.

Vector search might not seem that hard, but it's hard to make it fast with filters, high recall, and good cold start on object storage...

Bojan pulled off a monumental effort on the foundational query planner for this. 🚀

I think we all just like to assume that if a cloud vendor is in my AWS region, and I'm in that AWS region, network is just free, right? I just connect to their public internet IP by the same DNS as everyone else and won't have to pay a dime in networking?

I wish :'(

I built the first, quick napkin math for turbopuffer's financial planning on Causal, and our fractional CFO has taken it further and integrated Quickbooks. He's a convert

Excel FP&A ends up being terribly idiosyncratic, and impossible for me to self-serve... This is great

watching Bojan Serafimov pulling 2 IOIs a day [0] on the query planner with Morgan Gallant playing evaluator, godly pace

[0]: en.wikipedia.org/wiki/Internati…

took me an hour to spin up a new region of turbopuffer the other day, despite having done nothing special for multi-region provisioning

beauty of object storage being your _only_ stateful dependency.

Very interesting. I think you want your OLTP provider to be profitable. I sympathize with the difficulty of running a free database tier at scale. We haven't opened up turbopuffer fully for that reason (yet), despite it scaling incredibly already with 4 nines since launch.

my prayers have been heard. wow.

welcome Morgan Gallant 🫶🏻

Somehow I didn't know of fallocate(2) to move the head of a file to an offset

E.g. imagine you have a WAL file you persisted up to an offset; you can free space by calling `fallocate(persisted_wal_offset)`

man7.org/linux/man-page…

100m 1536 dim vectors fundamentally are ~600 GiB:

3xNVME (replicate): $190
SSD network disk: $120
Object+spot NVME: $45
HDD network disk: $30

cheapest, fastest, and most scalable to have an architecture of object storage w/ spot NVME caching like turbopuffer does