Postgres: A better message queue than Kafka? | Hacker News Reader

Postgres: A better message queue than Kafka?

dagster.io

65 pointstylermcginnis45 comments

Not just Postgres.

You can do exactly this with MySQL and SQL server too because they both support SKIP LOCKED.

Interestingly, the plain old file system on Linux also makes the basis of a perfectly acceptable message queue for many use cases - the thing that makes it work is that the file move operation is atomic. Atomic moves are what make queuing systems possible.

You could write a file system based message queue in 100 lines of async python, which I did here:

https://github.com/bootrino/arniesmtpbufferserver

File system based message queues can be written in any language, extremely simple and, most importantly - zero configuration. One of the most frustrating things about queuing systems is configuration - that includes database backed queuing systems. They can also be fast - I wrote one in Rust which maxed out the hard disk's random write capability well before maxing out the CPU - from memory it beat most of the common queuing systems in terms of messages per second.

Not all use cases for queues need to be able to globally distributed messages queues with the sort of guarantees needed for financial transaction processing. I would suggest to you that in fact most queues out there are used as outbound SMTP queues, which are then over engineered to use something like Celery, which is a nightmare to configure and debug.

What makes it atomic is running publishers and consumers on the same box (since you're sharing filesystem between those).

Also listdir is a big bottleneck here:

    while True:
        # get files in outbox
        files_in_outbox = [f'{PREFIX}/outbox/{x}' for x in os.listdir(f'{PREFIX}/outbox')]

>> What makes it atomic is running publishers and consumers on the same box (since you're sharing filesystem between those).

It's the move/rename that is atomic.

https://man7.org/linux/man-pages/man2/rename.2.html

Not really.

       However, there will
       probably be a window in which both oldpath and newpath refer to
       the file being renamed.

But that's not even the main point.

1. Move happens after email is sent, so there is a window where email is already being sent but file still exists. 2. Even if you do it before, there's still a window between os.listdir() and os.remove() 3. Complexity is O(N^2) due to listdir() + getctime() being called on every iteration.

If you just want to ensure order, it probably works fine at a small scale. But it would be unwise to run multiple consumers on a single instance, and impossible to run them on multiple instances.

I worked with a network filesystem that supported atomic renames and we based an entire large-scale production system on the idea that it would work (it did). The system supported Youtube and Google Play model training, regularly processing increments of hundreds of terabytes.

We found difficulty with the purely advisory locking that Linux had. Also the possibility of network filesystems made it a pain.

Do you have any experience with either?

You don't lock, you move the file that it next to be processed. File moves are atomic. You move the file out of the list of files that are being picked up for processing.

Lock free.

Network file systems do not support atomic moves, but you should not run such an application on a network file system.

I also found that advisory locking has a lot of gotchas, especially when used in multithread contexts (apparently you can lose the lock because a different thread closed a different file descriptor on the same file).

love to see this kind of hackery

I love to see effective yet brutally-simple "redneck engineering" solutions to software problems, particularly ones that straddle the line between genius and stupidity in a way that makes architecture astronauts feel uncomfortable.

I used to work at an brokerage that worked with a panel of around 12 providers, all of whom offered 5+ products that were updated multiple times per year to meet changing regulatory requirements. When a sales adviser recommended product X from provider Y to a customer, the adviser would then need to fill in an application form that was precisely tailored to that particular revision of that particular product. Bear in mind that these were complex, multi-sectioned forms. Needless to say, this created a huge workload for the devteam to keep track of all the product changes and update the UI accordingly each time.

At some point, someone on the devteam had the genius idea to simply take the PDF application forms from the provider, extract the individual pages as PNGs and overlay HTML form elements on top of them. The provider would essentially be doing our UI design for us. Add in an editor tool so the sales managers could set up the forms themselves and a tagging system so specific fields could be pre-filled from customer data we already had stored in the DB and the devteam's workload dropped by maybe 90%. Simple, stupid perhaps, but effective.

A controversial, but a very pragmatic take.

Queues are great for semi-infinite scalability, but you rarely need it.

There's numerous subtle benefits to using db compared to regular message queues that are often overlooked.

Being able to delete, reorder, or edit specific messages can be a lifesaver when things go wrong.

Most common issue with queue-based systems is getting overwhelmed with messages. Either your consumers went down. Or your producers had a bug / executed too frequently.

Being able to recover with a simple SQL query is a blessing.

Here's just a few examples from my experience:

1. Huge number of messages from test system were accidentally inserted into production.

Queue solution: disable consumers, move messages to a temporary queue while filtering them, move messages back to the old queue, enable consumers

DB solution: just delete rogue messages

2. We want to store some of the messages, but we're not ready to process them yet

Queue solution: create a separate queue for each message type, insert messages to different queues, manually move them when you're ready to process them (and keep in mind that not every queue can persist messages forever, SQS for example can't hold messages longer than 14 days)

DB solution: just skip those messages while processing

3. Consumers went down and the queue now contains a big number of duplicate messages. While it was fine to just wait a couple hours to let it stabilize, a whale customer started complaining

Queue solution: none (any hacky solution would take longer than it takes for the system to naturally stabilize)

DB solution: move whale customer messages to the front of the queue

1. Kafka isn't a message queue, but advance consumer offsets past bad data. Hit person who inserted test data into prod with a cricket bat and then hit person who designed a system where they could do that with the same cricket bat, but harder.

2. Kafka - do what you did with your DB consumers.

3. Kafka - consumers going down can't cause duplicates, what's even going on with your queue?

> what's even going on with your queue?

A message with the latest state of an object gets pushed into queue. When noone is consuming those, you essentially get duplicates: queue contains all of the updates, but you only care about the most recent one.

Throughput is often where an RDBMS falls over here. Unless you have partitions/isolation around your high throughput workloads (or throw tons of horsepower/$$$ at it) you can’t achieve the same ops/sec as a dedicated queueing mechanism.

But you do have a lot of valid points, particularly about failure recovery. My approach tends to involve a hybrid of the two if high speed is a necessary component somewhere. The system can restore its runtime state from a db, but the active mechanics of it are running on message passing of some kind.

"The write path. We built a daemon that would select log entries that were older than two weeks, copy them into a file in S3, and delete the rows from the database"

Seriously... How can you ever consider saying "a rdbms is just fine as a kafka alternative" under those conditions ?

Why not?

Ideology.

There's alot of people who are ideologically opposed to database backed message queues. They're usually reluctant to give detailed explanations why, because it's an emotional thing.

Because that would be equivalent to saying « rdbms are fine as a kafka alternative provided you build a compensation mechanism for all the things it does worse than kafka». It doesn’t make any sense.

How does Kafka solve this problem any better?

Your messages expire so if you want to archive you need a consumer to write them to disk. That sounds very similar.