this is just optimizing for token windows. flat code = less context. we did the same thing with java when memory was expensive, called it "lightweight frameworks"
does it handle skewed distributions? faker's always been useless for this - like, your test data ends up with everyone having 5 orders when real data is all long tail
we had to restrict ours to views only because it kept trying to run updates. still breaks sometimes when it hallucinates column names but at least it can't do anything destructive
curious about the startup latency in practice. docker containers even with warm pools still feel sluggish for agent loops. e2b does firecracker and it's noticeably snappier
how well does the flatten() translation work in practice? every time i've used localstack or similar the queries work locally then break in subtle ways once deployed
the sql-first thing is interesting. main thing that bugs me about alembic is i never know what order migrations will apply in when there's been a merge. how does jetbase handle branching?
does this actually fix metadata filtering during vector search? that's the thing that kills performance in pgvector. weaviate had the same problem, ended up using qdrant instead
does this do point-in-time recovery? tried like 3 postgres backup tools, they all claim to support it but then you actually need to restore and it's a mess
memory voting sounds interesting but does it work? i tried having agents mark useful chunks once, they just marked everything as helpful. accuracy went to shit
had this happen with a retry loop. hit $80 on anthropic before i caught it. how does this handle retries? seems like an agent could just keep retrying and blow past the limit
the lockfile thing is what sold me honestly. had a nightmare with torch pulling different cuda builds on my linux box vs mac. uv just... works? still half expecting it to break
yeah i have this setup. pain point is syncing between them, still haven't figured out a good pattern. just running a cron overnight which feels dumb but works