The MACI French podcast honoured me with an invitation to a guest appearance on their weekly schedule. As you can imagine, we talked about many things related to PostgreSQL… and also reacted to some newsworthy articles carefully curated by the MACI team. One of the topics we discussed in the podcast started with looking at PostgreSQL through the angle of it being one of the microservices that your application would be composed of.

The podcast episode is available as a video, and it’s in French.

Votre attention s’il vous plaît, le message à caractère informatique suivant est delivré en français :

I must admit looking at Postgres as a microservice is an exercise that I like. It allows focusing on what Postgres offers that your application is happy to re-use rather than having to implement it for itself.

Postgres is not (just) about storage

At a first glance, the Postgres microservice might look like a storage API to many developers, and that’s how we got started in the podcast.

I believe that’s a mistake though. A storage service would typically host files encoded as XML, JSON, Parquet, or some other format, and rely on a get/set API such as the ones found in all the Cloud services nowadays (think AWS S3, Azure Storage, or GCP Cloud Storage… or if you want to host it yourself, have a look at MinIO).

I believe that Postgres is not a storage service. When RDBMS were invented in the 70s developers knew how to serialize memory to disk and read from disk in memory again when needed, of course. Some early file systems already existed, and even the Unix hierarchichal file system was available. So that’s not what our pioneers set themselves to solve as an architecture component.

Postgres is all about concurrency

The Postgres microservice would be the concurrency microservice, if that would ever make sense. A service that maintains a data set with some guarantees (think ACID, schema, constraints) when it’s being accessed in a read/write capacity by several concurrent actors; and a service that even allows online schema changes.

As an thought exercise, imagine implementing a network service that allows your application to insert/update/delete items in a single collection. Let’s target something small, so that it appears to be an easy enough problem to solve, something like a single collection of about a dozen items. Now, the idea is to have half–a-dozen users concurrently accesing the collection to add, remove, or edit items in there. The job is the write a service that is reliable and predictable in how the changes are implemented when the API is used by several concurrent users.

If you’ve been a developper for a little while already, it might be obvious to you why implementing such a service is already non-trivial. If you think it’s trivial, well, spend a couple days on it and write a concurrent test suite that makes you feel safe about the concurrency aspects of what you’ve done.

Now, expanding on our single collection of items, we may introduce relations and constraints such as data types and the NOT NULL check, etc. And make it scale to hundreds of concurrent users and millions of items, or tuples.

Isolation and Locking

The ability to maintain read/write access to a single data set with concurrent users is what Postgres implements for you, in a way that you don’t have to think too hard about the concurrency aspects of the database properties.

Well, until you need to anyway, as shown in my article PostgreSQL Concurrency: Isolation and Locking where we dive in the user settings that impacts transaction characteristics.

And sometimes you might want to relax your approach to concurrency in Postgres and implement a read-only cache that you would only update on a schedule, rather than at transaction boundaries. That’s what my article series titled PostgreSQL Concurrency is all about, give it a read: the series dives into advanced tooling and concepts such as Materialized Views, Listen/Notify, and Triggers.

Say it again: RDBMS are all about concurrency!

These days the RDBMS innovation landscape is shifted towards distributed SQL and how to handle very large databases. Some aspects require an OLAP approach, others want to still handle OLTP systems of records in the billions scale and more, and more and more we see HTAP use cases.

HTAP stands for Hybrid transactional/analytical processing and is a mixed approach that attempts at solving both analytical and transactional aspects of your data production life cycle within a single system. When dealing with a data set in the Petabyte scale, having to operate more than one copy of the system becomes somewhat impractical, or at least quite costly.

One of the major difficulties when implementing a distributed SQL solution is around transaction properties of the system. We like the ACID aspect of RDBMS because it allows the application code and the developers to focus on a single user workflow at a time. That’s quite wonderful, let’s keep it that way!