Hacker News Reader

Comment by craigkerstiens | Hacker News Reader

The native time partitioning in Postgres 10 was a great foundation, but definitely had a few rough edges. With Postgres 10 you mostly needed some additional extension like pg_partman (https://www.citusdata.com/blog/2018/01/24/citus-and-pg-partm...) or timescale to smooth out the rough edges and make things as usable as you'd expect.

This set of changes in 11 is pretty great and makes it much more usable out of the box. For those not reading the article the short list of improvements:

  - Updates move records across partitions
  - A default/catch all partition
  - Automatic index creation
  - Foreign keys for partitions
  - Unique indexes
  - Hash partitioning (where as 10 was just time/range based)

View on HN

Sounds great. Hopefully someone will do up a guide on how to migrate a partitioning scheme created using pg_partman over to Postgres 11. It would be great to remove that dependency.

Still missing:

* Dynamic repartitioning to balance query load/space/create new partitions when tables get over a certain size. Without this, when you get more than a few hundred TB of data, you'll end up with terrible hotspots where single partitions are much more loaded than others and theres nothing you can do about it.

* Query replanning: Query plans can sometimes have a lot of error. By replanning, one can start on another query plan if the existing plan is grossly wrong in its estimates.

* Lock Eliding instead of *Exclusive locks. Some major operations, such as schema changes, take locks which effectively stop the server responding to any other requests till the operation is done. Instead, these operations should work on a copy-on-write image of the database, repeatedly re-syncing with recently changed data until the final change necessary is small, and then atomically commit the change.

Which database does this dynamic repartitioning and how without user intervention?

All the partitioning tools for postgres allow it.

You can say that it should aim for ~100GB per partition, partitioned by time, and it'll aromatically add or remove partitions.

You can go even further with the time scale extension for postgres.

The question was which DB's do this out-of-the-box like you requested, not plugins or tools. However, seeing that there already is an extension for postgres to do it, why is it a problem in the first place?

The partitioning features in PG10 and PG11 were also supported by said extensions, but now they're batteries-included features. The GP is implying that, as long as they're eating the lunch of these extensions, they should eat the whole lunch.

(Which I personally agree with, given that you almost never get these extension—or support for arbitrary extensions—on Postgres DBaaS hosts. Upstreaming a feature = making it available on these hosts.)

The development model of postgres is to start with a simply le implementation and improve it with every release. So maybe with pg13 or 14 this will be available in the core if one developer starts working on the remaining features.

There is a very sharp distinction between providing first-class partitioning, and providing automated maintenance and performance tuning.

Providing first-class partitioning is something that extensions can't possibly provide, due to how many places this has to integrate with.

On the other hand, manipulating partitions (which is effectively automatic table creation) seems like the sort of magic I expect to only exist in extensions.

As mentioned in a different comment, it does sound like people are looking for round-robin load distribution rather than partitioning.

I don't think any of the extensions actually provides the type of auto-tuning / balancing you're asking for, actually. For example pg_partman helps with setting up some partitioning schemes (by time, by ID), but does not rebalance it later (e.g. to split a hot or unexpectedly large partition). It probably provides more convenient interface to do that, but it's still up to the DBA to do it.

I'm sure this is one of the things people on pgsql-hackers will be talking about, but I wouldn't hold my breath for it to get into PG12. I do expect this to emerge in an extension first, and then getting the most useful bits into core (either in the form of built-in features, or hooks allowing extensions to be smarter).

Bigtable does it. [1] . Spanner does it [2].

[1]: https://static.googleusercontent.com/media/research.google.c..., Section 5, "and also splits tablets that have grown too large"

[2]: https://storage.googleapis.com/pub-tools-public-publication-..., Section 1, "it automatically migrates data across machines (even across datacenters) to balance load"

Those are distributed databases and have completely different characteristics. That's an interesting nice feature to have but no traditional RDBMS supports anything close I presume.

#1 seems a bit weird. You could theoretically have the DB automatically create new buckets for a hash partition, but hash partitions shouldn't have hot spots unless you hash on a bad column, and creating new buckets will require all existing rows to be moved.

For other types of partitions, partition hotspots would be the fault of the DBA partitioning on a bad column. The DB could theoretically half partition ranges if a range partition is very large, but there are other things that are more important.

#2 is a nice performance feature which has been talked about before, but do any DB's actually have this yet? I suspect that it is quite tricky, and won't be universally beneficial.

Is #3 really a problem? I certainly wouldn't expect schema changes to perform well, and increasing complexity just to help in this area would seem silly.

> Is #3 really a problem? I certainly wouldn't expect schema changes to perform well, and increasing complexity just to help in this area would seem silly.

Not the parent, but obviously they refer to the online schema migration problem. Without that online schema migrations are basically impossible. Or you get handed the burden of implementing or finding a tool that does it for you.

And yes, pt-online-schema-change[1] is such a tool, but for MySQL. MariaDB has this feature built-in with their `ALTER ONLINE TABLE` [2].

So something similar for Postgres would be godsent.

[1] https://www.percona.com/doc/percona-toolkit/LATEST/pt-online...

[2] https://mariadb.com/kb/en/library/alter-table/

mysql 5.7 has pretty good online DDL support[1]

I have been using it exclusively now instead of pt-online-schema-change.

[1] https://dev.mysql.com/doc/refman/5.7/en/innodb-online-ddl.ht...

> hash partitions shouldn't have hot spots unless you hash on a bad column

In the nice world of all data having an even load, sure...

But in the real world, you can easily get a handful of users in your "Users" table sending millions of requests per second, and in that case, you would really like to re-partition so that they don't happen to all end up on the same partition.

Implementation can be as simple as allowing splittable partitions, so that whenever load/size gets too high, a partition can be split in half, and half the records moved to a new host. The partition map is only a few kilobytes, so is globally shared/updated. There are no concurrency issues, because during the splitting process, either the old or new partition is responsible for each record, and both the old and new partition hosts can respond to a read or write query, either themselves or by forwarding it to the other host.

Whenever two neighbouring partitions both see low load/space usage, join them with the same method in reverse. By joining only neighbouring partitions, you can't suffer terrible fragmentation and blowing up the size of the global partition map.

I'm not sure what you mean by "move to a new host" - all the built-in partitioning features are single-host for now, so this does not make much sense. There are plans to support remote (foreign) partitions, but that is in the future.

A more immediate use case is partition-aware algorithms. For example joining equi-partitioned tables can be done at the partition level, and the smaller the partitions the faster the join (likely). Or partition-aware aggregation, when the partition key is included in the GROUP BY keys. And so on. In that case it makes sense to split the largest partitions to make them smaller.

Partitions provide separation, not load balancing, and does thus not concern themselves with load or size.

While a can imagine a way to make hashed partitioning sub-divide buckets, it won't really be useful, as you can still have a single user blowing their bucket out of proportions with others being almost empty, which partitioning can never solve.

What you want is round robin distribution and parallel query execution, not partitioning. This of course only makes sense if the tables are on different disk systems/hosts or if all tables fit in memory.

No, that's not true. When combined with parallelism, ability to place partitions to other hosts and partition-aware algorithms (partition-wise joins/aggregation etc.) it can be a powerful scalability feature.

It is true. However, while partitioning is not load balancing, partitioning can be used to make some form of load distribution.

Partitioning primarily deals with isolating related data into small pools so that queries do not need to touch full tables. The purpose of the buckets is to group related data so that a single partition may serve a query (instead of having to touch the entire table), thereby lowering the load from each individual query, not distributing all queries.

Load balancing deals with distributing all queries to ensure even node load, not increasing efficiency of queries (which it does not do at all). Data distribution leading to round-robin accesses across nodes/"partitions" is best. This is the exact opposite of partitioning, which tries to group all that data into a single partition.

Hash partitioning on serial or uuid might give you round-robin like behavior, but it's not really what partitioning is meant for.

I'm not really sure what is your point. I was not suggesting partitioning alone is a load-balancing solution, but that partitioning in combination with other features is a powerful scalability feature. Assuming there is a (1) way to keep the partitions withing some size limit automatically, and (2) the partitioning scheme is such that the queries can leverage the partitioning scheme (eliminate most partitions, join tables partitioned in the same way).

We don't have (1) yet - at least not in core PostgreSQL, but this part should be doable in an extension until it gets into the core. We do have (2) to a large extent, although I'm sure there are gaps and room for improvement.

And when combined with ability to place partitions to other hots, that will be another step.

No one is claiming partitioning alone magically solves scalability, but IMHO it's an important building block in doing that (for a large class of applications).

My point is that partitioning is the polar opposite of load balancing, and the wrong tool to implement it.

Oracle has had it for awhile (though I forget what its called.)

SQL Server is getting it in 2017 https://docs.microsoft.com/en-us/sql/relational-databases/pe...

Oracle is also your first born expensive and Postgres is free.

Really? Someone was offended and/ or defending the obscene pricing of Oracle?

Oracle has the right to charge money for their products if they want to. Most people here make a living from exchanging things for money that are better in some way than things you can get for free.

(I didn't downvote you, btw)

They could charge a trillion dollars for their software. I am not poo-pooing that. I am just upset with their practices in extracting those payments.

Of course. And likewise, people have the right to disagree with or be upset about their pricing model, billing approach and business practices.

FKs and unique indexes are a solid win. Very excited for this!

Keep in mind there are limitations - the unique index / constraint has to include the partition keys, and foreign keys are possible in one direction only. This may be improved in the future, of course.

Wait, so you cannot have a FK reference to the partitioned master table? :(

Unfortunately not in PG11, but not because we don't want it - it simply didn't get ready in time.

https://github.com/postgres/postgres/commit/3de241dba86f3dd0...

The relevant limitation is described by CREATE TABLE docs that say

    Partitioned tables do not support <literal>EXCLUDE</literal> constraints
    however, you can define these constraints on individual partitions.
    Also, while it's possible to define <literal>PRIMARY KEY</literal>
    constraints on partitioned tables, creating foreign keys that
    reference a partitioned table is not yet supported.

(the wording is slightly broken in the commit, but got fixed later).

You've really got me on an emotional rollercoaster here. :D

Reply on news.ycombinator.com