4 ms·

That doesn't sound right at all.

by PrimeMcFly 3y ago

ReflectedImage 3y ago
Sounds right to me. Postgres on AWS doesn't work because relational databases don't work on networked storage.
- zbentley 3y ago
  Not really. EBS isn't really network storage in the traditional sense. It's closer to iSCSI-attached NVMe over a dedicated low-congestion storage backbone network.
  - kossae 3y ago
    Lol I get what you’re saying but it’s funny your description is essentially “storage that is networked”.
  - ReflectedImage 3y ago
    This is very simple: EBS has a latency of around 2 ms. SSD has a latency of around of around 0.25 ms. A relational database will have around 10x the performance on an SSD compared to EBS because relational databases need to ensure data has been fully written to disk.
    - zbentley 3y ago
      Sometimes. The more expensive varieties of EBS provide comparable latency numbers to local SSDs (by using fancy on-hypervisor volatile storage or something? No idea), for an added cost. A separate set of behavior exists for RDS Aurora, which isn't using EBS underneath, so it might be worth looking into the Aurora latency characteristics (vs. cost, of course; Aurora ain't cheap) if you're concerned about the performance impact of EBS-vs-local-disk.
      - ReflectedImage 3y ago
        Ultimately, hosting a relational database on AWS will cost 10x as much regardless of which non-working flavour of it you choose. Million dollar database bills do not happen outside of the cloud world.
        zbentley 3y ago
        > Million dollar database bills do not happen outside of the cloud world. Laughs in Oracle per-CPU licensing terms
    - magicalhippo 3y ago
      Reminds me of when a customer complained a DB-heavy processing step took about 10x as long as expected, ie 10 minutes instead of 1. We used on-prem DB server and fat clients. After a long debugging session I hadn't found anything. Gradping at straws I asked what kind of network they ran. Turned out the clients all ran on laptops connected through Wi-Fi... so yeah 10x latency turned into 10x longer job.
- PrimeMcFly 3y ago
  I wasn't saying anything about Postgres not working on AWS but rather the alleged speed difference.
  - ReflectedImage 3y ago
    In theory, Postgres will be 10x faster on bare metal than AWS. From practical experience that is also what I see. It's the difference between using SSD and EBS for your disk storage.

That doesn't sound right at all.

ReflectedImage 3y ago

Sounds right to me. Postgres on AWS doesn't work because relational databases don't work on networked storage.

zbentley 3y ago
Not really. EBS isn't really network storage in the traditional sense. It's closer to iSCSI-attached NVMe over a dedicated low-congestion storage backbone network.
- kossae 3y ago
  Lol I get what you’re saying but it’s funny your description is essentially “storage that is networked”.
- ReflectedImage 3y ago
  This is very simple: EBS has a latency of around 2 ms. SSD has a latency of around of around 0.25 ms. A relational database will have around 10x the performance on an SSD compared to EBS because relational databases need to ensure data has been fully written to disk.
  - zbentley 3y ago
    Sometimes. The more expensive varieties of EBS provide comparable latency numbers to local SSDs (by using fancy on-hypervisor volatile storage or something? No idea), for an added cost. A separate set of behavior exists for RDS Aurora, which isn't using EBS underneath, so it might be worth looking into the Aurora latency characteristics (vs. cost, of course; Aurora ain't cheap) if you're concerned about the performance impact of EBS-vs-local-disk.
    - ReflectedImage 3y ago
      Ultimately, hosting a relational database on AWS will cost 10x as much regardless of which non-working flavour of it you choose. Million dollar database bills do not happen outside of the cloud world.
      - zbentley 3y ago
        > Million dollar database bills do not happen outside of the cloud world. Laughs in Oracle per-CPU licensing terms
  - magicalhippo 3y ago
    Reminds me of when a customer complained a DB-heavy processing step took about 10x as long as expected, ie 10 minutes instead of 1. We used on-prem DB server and fat clients. After a long debugging session I hadn't found anything. Gradping at straws I asked what kind of network they ran. Turned out the clients all ran on laptops connected through Wi-Fi... so yeah 10x latency turned into 10x longer job.
PrimeMcFly 3y ago
I wasn't saying anything about Postgres not working on AWS but rather the alleged speed difference.
- ReflectedImage 3y ago
  In theory, Postgres will be 10x faster on bare metal than AWS. From practical experience that is also what I see. It's the difference between using SSD and EBS for your disk storage.