What it means to be a Postgres extension

What it means to be a Postgres extension(citusdata.com)

144 points by akane 8 years ago | 33 comments

oliwarner 8 years ago |

This seems like a VERY shallow analysis of the benefits.

Forking an industrial-grade tool means the entire lifespan of the entire product becomes your responsibility to your client. Tracking the major upgrade changes might be a pain in the arse but they're nothing compared to tracking every security and data-loss fix that bubbles around the Postgres community.

It's not just developer time that's the cost here. They had to compile the whole Postgres+Citus database, for every platform they support, in a timely manner, test it and distribute packages. Think of all the CPU cycles and bandwidth they're saving by only having to compile as an extension against public headers.

Functioning as an extension means Postgres and its distributors (eg Ubuntu) are the people responsible for keeping Postgres alive and secure. Citus only have to support their thing.

Why aren't they talking about how much this move is saving them in day-to-day? There's no shame in being efficient.

bane 8 years ago | |

Being an extension also makes you entirely dependent on somebody else's platform, and makes it possible that your work will simply be subsumed by the platform if they think it's important enough. It's a very weak business position to be in and you have to have incredible future looking planning and brand buy-in to make sure you succeed like this.

oliwarner 8 years ago | | |

You're already dependent. Upstream can turn around tomorrow and provide everything your fork/extension does for free. They can alter their entire codebase to cause you weeks of work to keep up. It can be a hard slog being downstream, no doubt about it. That's why downstreams tend to "get involved" upstream. Sponsorship, sit on technical advisory boards, etc.

But what you're saying —which wasn't immediately obvious, and correct me if I'm wrong— is your users are using your database product, not Postgres, so you can hold them back as long as you like when they're using a forked product. They won't be carried away by an automatic update and it's much harder for them to jump ship.

And while there is some truth to that, it comes with a karmic cost. People picked you because you were based on their favourite, industry tested database. If you slip behind in features, or (more importantly) can't backport security fixes instantly, you're dead.

simooooo 8 years ago |

They were surprised that maintaining a postgres fork was a pain in the ass?

stubish 8 years ago | |

I think it was more that forks just get ignored. Customers want to run PostgreSQL, not something almost but not quite like PostgreSQL maintained by a company that may or may not employ PostgreSQL core contributors. And we certainly don't want to talk to your sales team about why it is a better fit for our use case.

The stuff Citus has been landing in PostgreSQL is fantastic.

jacques_chester 8 years ago | | |

> I think it was more that forks just get ignored

They call out a number of forks that do quite nicely. I work for Pivotal, which sponsors Greenplum. Companies pay handsomely for the capabilities it brings to the table.

But they are right that rebasing is a nightmare. My understanding (possibly wrong) is that the broad selection of APIs that make an extension-only approach did not appear in PostgreSQL until more recent versions -- anyone who forked earlier (such as Greenplum) have to first catch up and then migrate.

I do know that the Greenplum team have decided to catch up until they are working against mainline. It is, as you might imagine, a slow process: rebasing millions of lines of code a release at a time is not the easiest task on earth. But maintaining a fork will, in the long run, be harder.

vog 8 years ago | |

I also don't get it.

Ever for smaller changes, it is very well known that you better contribute back to the Free Software project - because having your patch included and being maintained there is a lot less trouble than maintaining your fork and updating your patches every few months.

One exception might be one-off changes, but we all know that nothing is more definitive than the temporary.

The only real exception is when your patch is superseded by another patch (or a better solution). Then, you maintian your private patch only until next version is finalized.

stubish 8 years ago | | |

It took a lot of work to not fork, and only recently became practical. Features like logical replication, DDL triggers, foreign data wrappers are all useful for this sort of thing and are all new. Companies like Citus and 2nd Quadrant first needed to get the infrastructure in place, so kudos to them and the PostgreSQL core team.

StreamBright 8 years ago |

Citus is my favorite website when it comes to Postgres content. Their blog posts are usually very informative and useful.

areskib 8 years ago | |

I couldn't agree more, I'm amazed by every single one of their post. They are doing a great job at communicating and giving rich insights on Postgres.

mosselman 8 years ago |

Can anyone shed some light on how the supposed new cluster features of Postgres 10 compare to something like Citus?

In other words. Does Postgres 10 offer the same features in terms of clustering as Citus does?

manigandham 8 years ago | |

No. Postgresql is a single-node database. It supports replication and failover to other nodes and foreign-data-wrapper extensions that let it query other datasources, but it does not have any support for natively working as a distributed database across several nodes.

Citus is an extension that takes several database nodes and makes them appear as a single logical database server (at the table level, by automatically sharding them based on a column).

giancarlostoro 8 years ago | |

All I found was this:

https://wiki.postgresql.org/wiki/Replication,_Clustering,_an...

And Citus is the first link in that list.

riku_iki 8 years ago | |

Citus provides ability to build sharded environment.

Data for different customers can be stored on separate nodes, and your DB is not limited by capacity of one node.

In regular PG, all data need to fit single node.

siscia 8 years ago |

Since we are talking about postgres, please let me go a little OT

Is there any interests in a RDS as a Service? So basically setting up and running a completely fault aware postgres cluster in any infrastructure, either public or private?