7 ms·

This is exciting: > ...This log architecture is optimized for raft consensus-based replication allowing the implementation of active-active multimaster. I'm a

by latch 5y ago

This is exciting: > ...This log architecture is optimized for raft consensus-based replication allowing the implementation of active-active multimaster. I'm a developer but manage 2 PostgreSQL instances each with an async replica. I also manage 1 three-node CockroachDB cluster. It's night and day when it comes to any type of ops (e.g. upgrading a version). There's a lot of reasons to use PostgreSQL (e.g. stored procedures, extensions, ...), but if none of those apply to you, CockroachDB is a better default. I just finished interviewing someone who went into detail about the difficulty they're having doing a large scale migration or a mission critical large PG cluster (from one DC to another)..it's a multi-team, year long initiative that no one is comfortable about. In my experience, PG is always the most out-of-date piece of any system. I still see 9.6 here and there, which is now EOL.

rad_gruchalski 5y ago
I'd go for YugabyteDB instead of CockroachDB but your point of distributed SQL being better is spot on.
mmontagna9 5y ago
> In my experience, PG is always the most out-of-date piece of any system. Yes
otterley 5y ago
Raft doesn’t magically make a database multi master. It’s a consensus algorithm that has a leader election process. There’s still a leader, and therefore a single master to which writes must be directed. The problem it solves is ambiguity about who the active master is at the moment.
- akorotkov 5y ago
  Right. The idea of active-active OrioleDB multimaster is to apply changes locally and in parallel send it to the leader. Then sync on commit and ensure there is no conflicts. The design document will be published later.
  - otterley 5y ago
    Please have it subjected to a Jepsen test suite ASAP!
    - lillecarl 5y ago
      My thinking is since it's postgres wire compatible the existing tests should work?
      - k__ 5y ago
        Does Jepsen just run automated tests?
        lillecarl 5y ago
        I mean they're automated but also hand crafted. https://github.com/jepsen-io/jepsen https://github.com/jepsen-io/jepsen
      - yencabulator 5y ago
        Jepsen does much more than basic end-to-end tests, including intentionally partitioning the cluster. Tests written for a non-distributed system are downright friendly compared to what Jepsen does to distributed systems.
    - collegeburner 5y ago
      Those are very expensive and booked months in advance. Aphyr sometimes comments on HN and is of course a better source on this but that's what I recall reading.
      - otterley 5y ago
        It’s the price you pay to reassure customers that your newfangled distributed database won’t experience data loss or inconsistency.
porker 5y ago
Feature wise I like postgresql. Operationally? Mysql. I pick MySQL 8 whenever I can because over the lifetime of a project it's easier. Pragmatically I'm coming to the uncomfortable (as a Foss advocate) idea that Microsoft SQL Server is a good choice and worth paying for. Features and relative ease of use.
- jamesfinlayson 5y ago
  Very much this - most of my experience is with Postgres but the rough impression I have is that SQL Server is the best of both Postgres and MySQL, and Postgres and MySQL are improving, but not in any great direction towards what SQL Server offers.
  - fuy 5y ago
    From the operations perspective, SQL Server is definitely way better that Postgres - monitoring is such a pain point in Postgres, for example. From application development/feature perspective, it's much closer, and server-side programming in Postgres (if that's what you like/need) is simply better that in SQL Server.
    - fuy 5y ago
      *better than, not better that.
- tpetry 5y ago
  Can you explain more why MySQL is operational more easy? I personally don‘t see any difference.
  - jamesfinlayson 5y ago
    I haven't used the features but I thought replication and failover were much easier (out of the box at least).
    - 5y ago
vlunkr 5y ago
Maybe postgres is out of date because even an old version like 9.6 is performant, featureful, and stable?
darkr 5y ago
> if none of those apply to you, CockroachDB is a better default. Though it scales horizontally, you will need to throw at least 10x or more hardware at it to achieve the same throughput
- vvern 5y ago
  Do you have benchmarks to back that up? The efficiency per core is worse, but I’ve not seen 10x worse. One thing to be careful of is making sure to have parallelism in your benchmark. Single-threaded Postgres may well be 10x faster on workloads of interest, but that’s not very representative of real use.
  - jordanthoms 5y ago
    We switched to CockroachDB for some of our most heavily loaded tables a few years ago - and the scalability and operational simplicity has been great. I don’t have a like for like benchmark, but at least for writes - we’re running with 5 replicas of each range so that’s 5x the disk write i/o just from that. Add in some network overhead and running raft etc, and you are going to be needing a lot more resources than for PG to maintain a similar throughput - but adding resources when you can scale horizontally is so much easier that the tradeoff is worth it in my experience.

by latch 5y ago

rad_gruchalski 5y ago

I'd go for YugabyteDB instead of CockroachDB but your point of distributed SQL being better is spot on.

mmontagna9 5y ago

> In my experience, PG is always the most out-of-date piece of any system. Yes

otterley 5y ago

Raft doesn’t magically make a database multi master. It’s a consensus algorithm that has a leader election process. There’s still a leader, and therefore a single master to which writes must be directed. The problem it solves is ambiguity about who the active master is at the moment.

akorotkov 5y ago
Right. The idea of active-active OrioleDB multimaster is to apply changes locally and in parallel send it to the leader. Then sync on commit and ensure there is no conflicts. The design document will be published later.
- otterley 5y ago
  Please have it subjected to a Jepsen test suite ASAP!
  - lillecarl 5y ago
    My thinking is since it's postgres wire compatible the existing tests should work?
    - k__ 5y ago
      Does Jepsen just run automated tests?
      - lillecarl 5y ago
        I mean they're automated but also hand crafted. https://github.com/jepsen-io/jepsen https://github.com/jepsen-io/jepsen
    - yencabulator 5y ago
      Jepsen does much more than basic end-to-end tests, including intentionally partitioning the cluster. Tests written for a non-distributed system are downright friendly compared to what Jepsen does to distributed systems.
  - collegeburner 5y ago
    Those are very expensive and booked months in advance. Aphyr sometimes comments on HN and is of course a better source on this but that's what I recall reading.
    - otterley 5y ago
      It’s the price you pay to reassure customers that your newfangled distributed database won’t experience data loss or inconsistency.

porker 5y ago

Feature wise I like postgresql. Operationally? Mysql. I pick MySQL 8 whenever I can because over the lifetime of a project it's easier. Pragmatically I'm coming to the uncomfortable (as a Foss advocate) idea that Microsoft SQL Server is a good choice and worth paying for. Features and relative ease of use.

jamesfinlayson 5y ago
Very much this - most of my experience is with Postgres but the rough impression I have is that SQL Server is the best of both Postgres and MySQL, and Postgres and MySQL are improving, but not in any great direction towards what SQL Server offers.
- fuy 5y ago
  From the operations perspective, SQL Server is definitely way better that Postgres - monitoring is such a pain point in Postgres, for example. From application development/feature perspective, it's much closer, and server-side programming in Postgres (if that's what you like/need) is simply better that in SQL Server.
  - fuy 5y ago
    *better than, not better that.
tpetry 5y ago
Can you explain more why MySQL is operational more easy? I personally don‘t see any difference.
- jamesfinlayson 5y ago
  I haven't used the features but I thought replication and failover were much easier (out of the box at least).
  - 5y ago

vlunkr 5y ago

Maybe postgres is out of date because even an old version like 9.6 is performant, featureful, and stable?

darkr 5y ago

> if none of those apply to you, CockroachDB is a better default. Though it scales horizontally, you will need to throw at least 10x or more hardware at it to achieve the same throughput

vvern 5y ago
Do you have benchmarks to back that up? The efficiency per core is worse, but I’ve not seen 10x worse. One thing to be careful of is making sure to have parallelism in your benchmark. Single-threaded Postgres may well be 10x faster on workloads of interest, but that’s not very representative of real use.
- jordanthoms 5y ago
  We switched to CockroachDB for some of our most heavily loaded tables a few years ago - and the scalability and operational simplicity has been great. I don’t have a like for like benchmark, but at least for writes - we’re running with 5 replicas of each range so that’s 5x the disk write i/o just from that. Add in some network overhead and running raft etc, and you are going to be needing a lot more resources than for PG to maintain a similar throughput - but adding resources when you can scale horizontally is so much easier that the tradeoff is worth it in my experience.