I'll also add, that the point is that a devtools unicorn is able to run with a single postgres cluster, so maybe it isn't the rdbms that's going to be the limiting factor on your startup with a hundred users.
Sometimes you'll run into situations where an instance of any software has degraded performance for reasons beyond either your control or understanding. That's one of the situations where horizontal scaling let's you deal with such circumstances better, instead of outright impacting everything and everyone connected to it.
A front end web server container crumbles under the load? No worries, just redirect the traffic to others. A back end API container has problems with the JVM reserved memory and/or the thread pool for processing connections has been filled and all new incoming requests just get queued up? Just have more instances to cope with the load and don't let one instance under load affect all others.
But what are you supposed to do if long running SQL exhausts the resources of your DBMS? What if a background process needs to run some really complicated writes, but your application needs write capability without slowing down?
That's not to say that bad configurations or bad code shouldn't be addressed, but any means of fighting it and minimizing the impact is worthwhile in my eyes.
You can do basically the same thing with read replicas, promote one of them to primary and shut down the machine causing issues.
You're far better off starting with a system that does seamless active-active by default, and then figuring out what kind of transactional guarantees you need. An RDBMS might make sense for a few specialised use cases, but most of the time it doesn't.
It's possibly the sanest default for a data persistence layer. I'd argue otherwise: use a RDBMS unless you know what you are doing.
> the number of web applications that make effective use of database-level transactions is approximately zero
Almost every driver and layer (ORMs, etc) that connect to databases use transactions. Transparently or not, but it uses them. So actually they are using transactions all the time.
> You're far better off starting with a system that does seamless active-active
Think about what means active-active. Either it means that you serialize everything in every active node (pointless); or you need to deal with concurrency and data synchronization. Which in turn means either supporting distributed transactions (e.g. via Paxos or RAFT protocols) or doing conflict resolution (which essentially is a nice way of saying "data loss"). Both problems are much, much harder than the (usually harmless) consequences of using transactions on a "classical" RDBMS.
The only drawbacks of transactions are that are limited to scale in write on a single node. But if this single node can scale to the numbers I mentioned ---and they do, and more-- most use cases will likely fit within this envelope.
Therefore being RDBMSs like Postgres the best and safest, the default, approach that should be taken for data persistence layers.
They're using them but they're not getting any useful effect out of them, since the database-level transactions don't actually align with business-level transactions. Fundamentally you can't actually use database-level transactions in a useful way in a web application because you can't share a transaction between multiple HTTP requests (without introducing much bigger problems).
> Either it means that you serialize everything in every active node (pointless)
Not pointless - you gain redundancy which is the whole point. Performance will be bad but for a low-traffic system it's often actually fine.
> Which in turn means either supporting distributed transactions (e.g. via Paxos or RAFT protocols) or doing conflict resolution (which essentially is a nice way of saying "data loss"). Both problems are much, much harder than the (usually harmless) consequences of using transactions on a "classical" RDBMS.
The consequence of using transactions on a classical RDBMS at scale without thinking carefully about it is deadlocks, which are just as bad as distributed transaction problems, and you still have to do conflict resolution in the case where a transaction gets aborted, which will happen sooner or later (most RDBMS fans just ignore it and lose data). If you want decent performance then eventually you'll have to model your data properly using CRDTs (this is true even in a classical RDBMS) and at that point you gain nothing from traditional RDBMS transactions.
> Therefore being RDBMSs like Postgres the best and safest, the default, approach that should be taken for data persistence layers.
RDBMSes are not safe. The way they achieve their integrity guarantees is by rejecting, and dropping, unexpected writes. This almost always loses data in practice.
Maybe this for consumer web applications, or for CRUD apps.
But my experience with B2B and SaaS applications is that sooner or later, I always need transactions (or locks) for something. Maybe it's for the billing code. Maybe it's for batch jobs or workflow logic. And that's when I'm really happy to have a transactional database.
The alternative to having transactions is setting up a Raft or Paxos server to handle distributed locks, and those require a lot more ops effort than a copy of PostgreSQL.
PostgreSQL works surprisingly well for something like this. It has a rich set of coordination and locking primitives, and it can usually be recovered and restarted in under 10 minutes if the server fails. Yes, connection overhead from 300 workers is a real problem, so it doesn't hurt to put a thin layer in front of it.
One thing that an RDBMS buys me is flexible feature set and a solid ecosystem. Want write-ahead logs and automatic point-in-time recovery? It's out there. Need a specific locking primitive? PostgreSQL probably supports it. Etc.
If I were looking at 2,000 workers, yeah, it would be time to set up a Raft server. But a good RDBMS will do the job for a surprisingly long time.
I do see where you're coming from, but I find they're more trouble than they're worth in the long term. YMMV I guess.
Validation is vital but the datastore is not the place to do it, because handling invalid data by dropping it on the floor is almost never the right behaviour.
There is no substitute for actually understanding your data model, but once you do 99% of the time you'll find using an RDBMS comes with minimal benefits and significant costs.
That being said, if your code doesn't catch and handle errors, then there's a couple of deeper problems already.
Well if you can't get input that violates that integrity then what are you gaining by enforcing that integrity?
> That being said, if your code doesn't catch and handle errors, then there's a couple of deeper problems already.
Sure, but how can you do error handling without data storage, given that your application itself is stateless? Maybe you retry the failure a few times, but ultimately your only option is to drop the data on the floor, which is almost always not what you want.
Fixing up data that was written wrongly takes work, there's no getting away from that. But if you have a record of the original write attempt then at least you can do that work and recover the data (if you decide the cost/benefit is worth it). Whereas if you have the datastore reject it at write time then you're SOL before you've even started.
Have you ever actually encountered any RDBMS that just silently "drops" your write without informing you, or is this just how you think they work?
Otherwise, isn't this just like any other data write erroring out -- MongoDB driver saying "Service down!", simple flat-file write saying "Disk full!", etc -- in that if your app doesn't react to this, it's not the data store but your code that sucks?
Perhaps I'm wording this badly?
Your application should _not_ be generating data which can't go into the database correctly. If it is, that's a different set of problems. ;)
* Read-only queries can easily be scaled out to read replicas.
* Write transactions.
As an example of numbers publicly available, GitLab.com runs more than 250K read-only txs and more than 60K write txs on a single Postgres cluster, with room for further vertical scalability [1].
Do you need millions of transactions per second? Then go ahead. Don't? Then ou can scale very very far with a properly tuned and administered relational database like Postgres.
[1]: https://about.gitlab.com/blog/2020/09/11/gitlab-pg-upgrade/