from Hacker News

Same-Region Read Replicas to Serverless Postgres

by throw14082020 on 7/17/23, 11:06 PM with 6 comments

  • by necovek on 7/18/23, 1:13 AM

    One generally introduces read replicas to overcome I/O and network bottlenecks: how does relying on a single datastore perform in circumstances where having multiple colocated copies of the data would speed things up?

    Btw, this is an interesting take on "serverless" terminology too: most other providers would call this "managed Postgresql".

    Yet this exact feature might enable something akin to "serverless Postgres": quick spin up and tear down allowing for short-lived instances responding to load (scaling setting should be differently laid out too: scaling configuration is too "serverfull"), but whatever Neon had before could not be "serverless".

  • by throw14082020 on 7/17/23, 11:08 PM

    From reading it, I'm not really sure if it's eventually consistent or not.

    > Replicas then update cache pages in the shared buffers. This ensures eventual consistency for read replicas within the same region as your database.

    > Data Consistency: Reading data from a single source ensures data consistency. This addresses a common challenge in traditional read replicas where there might be a replication lag.

    > Resource Customization: Neon allows you to allocate different CPU and memory resources for each replica.

    So there is a single source of data, but can have different CPU and memory and has eventual consistency in the same region?

  • by nikita on 7/17/23, 11:08 PM

    Thank you for submitting (CEO of Neon here). Happy to answer questions on our design decisions.
  • by nikita on 7/17/23, 11:13 PM

    (CEO of Neon). There is a TL;DR: of this blog post: https://twitter.com/nikitabase/status/1680636526823370752