After moving the babylovable Supabase project and DB into the same region as the Vercel workflow
speed jumped about 10x
What kind of hard days were we living before
I thought Supabase itself was slow, or my network was
I even optimized a whole k8s-like reconciliation engine
and still couldn't get preview-server startup any faster
These past two days I remembered we needed a local Supabase dev env and an isolated prod DB
and only then saw how fast it could be
Then I fixed the real issue
At project kickoff, on a whim, I picked the Supabase Tokyo region while everything else was in the US
Round-trip latency was huge
Result: sandbox create → preview ready went from ~1 minute to ~10 seconds
Along the way I also found some idle spinning in the k8s-like reconciler
Hard to believe AI wrote such junior code, for example
1. A hot DB query loading the entire session history when only one field was used
2. That hot query was unnecessary anyway — a reused constant would do; instead it hit the DB every time, and the field is userid, which never changes in that session!
ps. This wasn't a SOTA model. It was Cursor Auto, which I use a lot
Skip to content
After moving the babylovable Supabase project an…
From X · @therealfreddyx
Thoughts
mew～
After moving the babylovable Supabase project and DB into the same region as the Vercel workflow
speed jumped about 10x
What kind of hard days were we living before
I thought Supabase itself was slow, or my network was
I even optimized a whole k8s-like reconciliation engine
and still couldn't get preview-server startup any faster
These past two days I remembered we needed a local Supabase dev env and an isolated prod DB
and only then saw how fast it could be
Then I fixed the real issue
At project kickoff, on a whim, I picked the Supabase Tokyo region while everything else was in the US
Round-trip latency was huge
Result: sandbox create → preview ready went from ~1 minute to ~10 seconds
Along the way I also found some idle spinning in the k8s-like reconciler
Hard to believe AI wrote such junior code, for example
1. A hot DB query loading the entire session history when only one field was used
2. That hot query was unnecessary anyway — a reused constant would do; instead it hit the DB every time, and the field is userid, which never changes in that session!
ps. This wasn't a SOTA model. It was Cursor Auto, which I use a lot
speed jumped about 10x
What kind of hard days were we living before
I thought Supabase itself was slow, or my network was
I even optimized a whole k8s-like reconciliation engine
and still couldn't get preview-server startup any faster
These past two days I remembered we needed a local Supabase dev env and an isolated prod DB
and only then saw how fast it could be
Then I fixed the real issue
At project kickoff, on a whim, I picked the Supabase Tokyo region while everything else was in the US
Round-trip latency was huge
Result: sandbox create → preview ready went from ~1 minute to ~10 seconds
Along the way I also found some idle spinning in the k8s-like reconciler
Hard to believe AI wrote such junior code, for example
1. A hot DB query loading the entire session history when only one field was used
2. That hot query was unnecessary anyway — a reused constant would do; instead it hit the DB every time, and the field is userid, which never changes in that session!
ps. This wasn't a SOTA model. It was Cursor Auto, which I use a lot