The major news are:
- SQL/PGQ - A Graph Query Language
- JSON improvements (a JSON type, simplified notations)
Peter Eisentraut gives a nice overview here: https://peter.eisentraut.org/blog/2023/04/04/sql-2023-is-fin...
- SQL/PGQ - A Graph Query Language
- JSON improvements (a JSON type, simplified notations)
Peter Eisentraut gives a nice overview here: https://peter.eisentraut.org/blog/2023/04/04/sql-2023-is-fin...
SQL:2023 is finished: Here is what's new - https://news.ycombinator.com/item?id=35562430 - April 2023 (153 comments)
SELECT t.j.foo.bar[2], ... FROM tbl t ...Using a graph DB with many underlying KV-store nodes, you can have a single graph spread over many machines representing e.g. Facebook's social graph, and run a query which "chases around" edges between vertices that live on different nodes, to solve that query, while ensuring that as little of that has to happen as possible — both by rebalancing vertices so that data is sharded at low-connection-degree points in the graph; and by consolidating the steps of queries that occur on the same node into single batch queries, such that the whole thing becomes (close to) a single map/reduce step.
There's nothing in Postgres that knows how to do that; if you had e.g. a graph stored in a Citus hypertable, and did a recursive CTE over it to do graph search, then you'd get pretty dang bad perf.
A problem everbody would love to have but pretty much nobody actually has.
Except the people who do have it and need to keep their business running off of one postgres instance.
It's usually not that data doesn't fit on one machine but that load on the database exceeds what one machine can serve. A failover configuration might enable you to use the spares for some read operations and take a little load of the primary, but you lose ACID semantics when you do that and it generally doesn't help you for long.
a) At some point you will have more data or more users than one instance can handle. And instead of simply adding another node you need to throttle usage in order to do a rolling upgrade. Which is far easier said than done and involves impact to the business.
b) With distributed databases you are constantly testing that everything works in Dev, Test etc environments. With failover you really only test it every now and again usually before you deploy to Production. And in most companies which are hopeless the testing will be guaranteed to be inadequate.
c) Vendors lie. They promise that failover will just work but in my experience it very often doesn't. Which is another reason why b) is such important to validate their claims.
one of the things I would like to see in my lifetime is somehow it should be easier to "run the whole enterprise" from one box. Sure, it will probably be seriously underpowered and I can't do all things at once but for most small to mid-size companies, it should be possible to run all our "code" from one machine.
I think of this as some kind of development or pre-qa environment. It really shouldn't be that big of an ask...
I am thinking most, if not all, companies will be able to fit their entire enterprise on a Supermicro A1+ server with two 96-core processors. Sure, there is no machine in the world that can fit all of YouTube videos but there is no reason why we can't have YouTube, with a limited set of non-production data, running from just one box. Thoughts?
I think that's the mainframe idea. There's probably some interesting philosophical question in there about whether or not a data center is just a big mainframe. It sort of feels like it verges on semantics.
> Sure, there is no machine in the world that can fit all of YouTube videos but there is no reason why we can't have YouTube, with a limited set of non-production data, running from just one box. Thoughts?
I'm not sure what you mean exactly, but if you're streaming YouTube's traffic through one box (even if that box isn't directly connected to the disks on which the video is stored) you'll run into I/O bottlenecks--such a machine would need to push terrabytes per second which is probably not trivial. Moreover, having a single machine that can handle YouTube's peak traffic probably means you're underutilizing it most of the time.
Basically, somewhere I can run something and feel safe knowing this is production code except this one change I have made.
1. compute and memory bandwidth to serve complicated queries 2. IO
You can't scale memory bandwidth beyond some pretty low limit on one machine. You can't scale IO bandwidth beyond some limit. To give you an example, I've seen database servers with 20GB of data being so overloaded by compute requirements of complex queries that they needed to be scaled horizontally.
I mean I expect things to be slower I guess but to test for correctness and spec?
For example, at one place just over a year ago they were well into this territory. One of weird problems for them was with pgBadger's memory usage (https://github.com/darold/pgbadger). That's written in perl, which doesn't seem to do garbage collection well. So even on a reporting node with a few hundred GB's of ram, it could take more than 24 hours to do a "monthly" reporting run to analyse PG usage for the time period.
There wasn't a solution in place at the time I left, so they're probably still having the issue... ;)
It's not for scalability reasons it's for high-availability.
Which as cloud adoption has increased and server uptime has decreased is even more important.
Anyone with dogmatic opinions about this stuff need to be taken with a grain of salt. If you scale out PeopleSoft, your accounting system will exceed the value of your company. If you’re worried about webscaling your random app, that’s more wasting time navel gazing than accomplishing anything! :)
Postgres uses foreign data wrappers and table partitions to achieve sharding of relational databases over many nodes. It's certainly possible to make the FDW layer smarter wrt. being able to distribute certain queries to the shards, including some recursive queries that happen to be of practical interest.
But a lot of graph databases have sizes where they are fully replicated on each node and some small internal optimizations for the graph are good enough. Given the design of the graph query language (and what you can query, hint: not everything arbitrarily) there is no reason postgres can't gain some additional optimizations to do that effectively as long as it's not facebook scale.
SQL now having syntax for querying graph means Graph databases can support SQL, at least a subset of it, in a standardized way.
What you really want is to apply graph processing to data as it is. The SQL 2023 additions are a step in the right direction. I need to find a good detailed description of the constraints and semantics to assess how good it is.
I suspect if Postgres had a solid implementation of SQL/PCQ it would be a similar story for Neo4j.
There's also a good article by Martin Fowler about how "NoSQL" was really "NoDBA" for a lot of folks, and I definitely saw that dynamic. JSON fields can also be a good middle ground here, where a DBA can insure good "structural integrity" of your schema, but you don't need to go through the hassle of adding a new column and schema update if you're just adding some "trivial" bit of data.
Another good example is for generalized classified ads, different categories may have additional details, but you don't necessarily want to create the plethora of tables to store said additional details.
Some other controversial thoughts: SQL itself is a really not-ergonomical query language, and also the lack of any decent Rust-like enum typing is really unfortunate. I know lots of people think that databases aren't for typing, but (1) clearly SQL aspires toward that but gives up half way and (2) that's a shame because they have a lot of potential in that capacity. Also while you can sort of hack together something like sum types / Rust enums, it's a lot of work to do it reasonably well and even then there are gaps.
pg_jsonschema is a postgres extension that implements schema validation for JSON columns. I'm not particularly familiar with Rust, so not sure exactly what you mean by "Rust-like enum typing", but postgres has enums, composite types, array types, and custom scalars, so not sure what's missing.
Assuming you've got good integration test coverage of the database schema alterations end up taking a minuscule amount of time and if you lack test coverage than please reconsider and add more tests.
Sibling comment, "is when you want to store/use additional payment processor details for a transaction", is a great example IMO. I've dealt with card processing systems where the card transaction data can be reams of JSON. Now, to be clear, there are a lot of subfields here that are important that I do pull out as columns, but a lot of them are just extra custom metadata specific to the card network. When I'm syncing data from another API, it's awesome that I can just dump the whole JSON blob in a single field, and then pull out the columns that I need. Even more importantly, by sticking the API object blob in a single field, unchanged, it guarantees that I have the full set of data from the API. If I only had individual columns, I'd be losing that audit trail of the API results, and if, for example, the processor added some fields later, I wouldn't be able to store them without updating my DB, too.
Before JSON columns were really standard, saw lots of cases where people would pull down external APIs into something like mongo, then sync that to a relational DB. Tons of overhead for a worse solution where instead I can just keep the source JSON blob right next to my structured data in postgres.
And there are four major reasons still to choose MongoDB over something like PostgreSQL.
a) PostgreSQL has terrible support for horizontal scalability. Nothing is built-in, proven or supported.
b) MongoDB has superior ability to manipulate and query the JSON.
c) MongoDB is significantly faster for document-attribute updates.
d) MongoDB has better tooling for those of us that prefer to manage our schema in the application layer.
Sharding has nothing to do with high-availability.
And primary-secondary failover in my experience is rarely without issues.
There is a reason almost every new database aims to be distributed from the beginning.
That's partly because you can't compete with the existing RDBMSs if you're single node: they are good enough already. Nobody will buy your database if you don't introduce something more novel than PostgreSQL, whether that novelty is worth it or not.
---
And to respond sibling comment about "noticeable" downtime....
Primary-secondary failover in <1m is very feasible. And each minute downtime is a mere 0.002% for the month.
Primary-secondary isn't what is hurting your availability.
And given that the entire industry has moved to a distributed model despite its complexity gives you a hint as to which way the wind has been blowing for the last decade.
Failover is automatic for PG when using e.g. Patroni. Of course you lose active transactions and that might be a showstopper, but monitoring failover? I am curious when you'll have to do that.
In particular it is nice to see that a core dev views JSON dot accessing and PCQ as "sensible" future additions to Postgres.
[1] - https://peter.eisentraut.org/blog/2023/04/18/postgresql-and-...
Now with 23c, they are adopting PGQ and one should be able to interface through the regular JDBC connectors. Will see how it shapes up.
It may be interesting to see if real world usage of PGQ pushes the database engines to do interesting things.