Comment by mgh95 | Hacker News Reader

Comment by mgh95 | Hacker News Reader

Parent Full thread

From the article:

> 4. Performance

> How can we ensure there are no performance regressions, and even performance improvements? When we launched Shepherd we promised to respond to every submission with an indication in 24 hours or less, which directly relates to the performance of our platform.

I actually run a (small, bootstrapped) startup in a very related space (we operate as an agency and MSB, and have direct tie-ins to a much broader range of insurance and financial products) and this article just feels like going the wrong direction. We host on prem (colo) because we knew the DB would be a latency bottleneck and clone operations need to be fast for compliance purposes.

This just feels like the wrong solution for the problem.

I had a similar reaction; one of their constraints was “Reduce server memory”, which tells me their Serverless vendor charges more, which is a constraint they don’t need to have. Using a colo server, memory cost is rarely an important factor.

Pretty much. You can run your business basically saying things like "we will take PITR not less than every seven (7) days" and then just run it on off hours with `pg_dumpall` on each database and the memory/storage costs are basically nil.

This entire blog post reflects a lack of awareness of how much these problems can flat out be ignored with different approaches.

IMO, think of it like switching from a bicycle to a car. The bicycle might be faster in certain conditions, but the car offers more features and flexibility for different terrains. Similarly, Neon’s database branching feature is like adding all-terrain capability.

I very much disagree with this. It is possible (and we do this in production) to branch a pg database without requiring these tools using `pg_dump` or `pg_dumpall` which uses the ability to branch a database. These can be integrated with either standard ORM tools or SQL scripts to achieve the same result with far better performance and lower cost. This very much appears to be a knowledge issue.

Reply on news.ycombinator.com