I wonder if Kafka represents an existential angst in these Kubernetized Microservice times. Or is it more simply I am just too dumb to learn and use this shit correctly.
But there are times when you have a problem, and amongst the possible solutions is Kafka.
I've come across Kafkaesque problems only three times in the last seven years: a hosting platform that had to parse logs of over 700 WordPress sites for security and other businesslogic. Putting all events of a financial app backend into datalakes and filtering and parsing all openstreetmap changesets live.
Maybe a few use cases could be switched out for direct API calls, but I think Kafka hits the sweet spot in many situations.
What alternatives would you be looking at?
Kafka is one of those systems that needs to be justified by out-scaling other solutions that don't come wedded with all its baggage.
Want to store data, query it in arbitrary and hard to foresee ways and also want to easily tune performance for these queries? Relational datastore it is.
Want to have ACID? Well, relational datastore it is.
Kafka is not the right solution for these problems.
Related HN discussion of this [1]
I mean, whoever in their right mind would want to:
- have a snapshot of data
- query data, including ad-hoc querying
- query related data
- have trasactional updates to data
When all you need is an unbounded stream of data that you need to traverse in order to do all these things.
Being able to see a snapshot is good, and I would hope to see a higher-level abstraction that can offer that on top of something Kafka-like. But making the current state the primary thing is a huge step backwards, especially when you don't get a history at all by default.
> - query data, including ad-hoc querying
OK, fair, ad-hoc queries are one thing that relational databases are legitimately good at. Something that can maintain secondary indicies and do query planning based on them is definitely useful. But you're asking for trouble if you use them in your live dataflow or allow ad-hoc queries to write to your datastore.
> - have trasactional updates to data
I do think this one is genuinely a mistake. What do you do when a transaction fails? All of the answers I've heard imply that you didn't actually need transactions in the first place.
No more daily SQL dumps from offshore to onshore and big batch procedures to genereate outdated events.
Bank -> debit card purchase -> perform all required database work in a transaction -> transaction fails -> decline debit card purchase
Without transactions, in this scenario, maybe the debit card transaction fails but money is still taken out of your account? Doesn’t sound very pleasant.
But it's interesting to compare his factorio style to mine. Or my factorio style to my regular programming style. They're very different.
Why?
When is "I need to query all of my log to get the current view of data" is a step forward? All businesses operate on the current view of data.
> OK, fair, ad-hoc queries are one thing that relational databases are legitimately good at.
Not just ad-hoc queries. Any queries.
> But you're asking for trouble if you use them in your live dataflow or allow ad-hoc queries to write to your datastore.
In our "live datafows" etc. we use a pre-determined set of queries that are guaranteed to run multiple orders of magnitude faster in a relational database on the current view of data than having to reconstruct all the data from an unbounded stream of raw events.
> What do you do when a transaction fails?
I roll back the transaction. As simple as that.
All businesses operate in response to events. Most of the things you do are because x happened rather than because the current state of the world is y.
> In our "live datafows" etc. we use a pre-determined set of queries that are guaranteed to run multiple orders of magnitude faster in a relational database on the current view of data than having to reconstruct all the data from an unbounded stream of raw events.
If you have a pre-determined set of queries, you can put together a corresponding set of stream transformations that will compute the results you need much faster than querying a relational database.
> I roll back the transaction. As simple as that.
And then what, completely discard the attempt without even a record that it happened?
Yes, but once an event happens, business needs access to current state of data.
> If you have a pre-determined set of queries, you can put together a corresponding set of stream transformations that will compute the results you need much faster than querying a relational database.
No, it won't. Because you won't be able to run "a corresponding set of transformations" on, say, a million clients.
You can, however, easily query this measly set on a laptop with an "overengineered" relational database.
> completely discard the attempt without even a record that it happened?
Somehow in your world audit logging doesn't exist.
[1] https://github.com/edenhill/librdkafka [2] https://github.com/Shopify/sarama
Of course you can. It's a subset of the same computation, you're just doing it in a different place.
> Somehow in your world audit logging doesn't exist.
If you have to use a separate "audit logging" datastore to augment your relational database then I think you've proven my point.
>the attempt to charge is recorded in a leger
Hint: how do you think this attempt is recorded and fulfilled? Or, do you think "it's just appended" and bank recalculates your balance from scratch every time you spend 1$ on coke can?
Only bank I've heard of that's not using traditional relational database for ledger is Monzo [1] - but they still use Cassandra's transactions.
[1] https://www.scaleyourapp.com/an-insight-into-the-backend-inf...
That's how the bank I worked with did it. Of course there was caching in place so we didn't actually recompute everything every time, but the implementation of that was a lot closer to "commit a kafka offset" than an RDBMS-style transaction. (E.g. we didn't overwrite the "current balance" in-place, we appended a new "current balance as of time x").
Yeah, who could need to know exactly how many items of a particular product they have in stock currently, or how much money a customer has in her account at the particular moment she wants to do a withdrawal? It's really hard to come up with any useful real world examples when this could be the case.
> What do you do when a transaction fails?
It depends on why the transaction fails and in which way. But sometimes it is really useful to make sure that when one account is debited, another one is credited at the same time.
> Of course you can.
Of course, you can't. Because you can't run a million transformations. Whereas querying specific data for any of the one million clients? It's trivial on a relational database.
Moreover. If you need new queries into data, it's again trivial. Because you have the current view of your data, and you don't need to recalculate everything from the beginning of time just because your requirements ever so slightly changed.
> If you have to use a separate "audit logging" datastore to augment your relational database then I think you've proven my point.
No, I haven't.
It's funny, however, that you think that businesses don't require a current view of data and need to re-calc everything from scratch.
I think you've proved our point
There are just too many scenarios where not having transactions is dog slow or really really unwieldy.
For me, Kafka sits in the same area of solutions as Kubernetes, Hadoop clusters, or anything "webscale": you don't need it. Untill you do, but by then you'll (i) have Serious Problems which such systems solve and (ii) the manpower and budgets to fix them.
With which I don't mean to avoid Kafka at all costs. By all means, play around with it: if anything, the event-driven will teach you things that make your a better Rails/Flask/WordPress developer if that is what you do.
* Just keep your architecture a monolith. You'll do fine the majority of the cases.
* Event-sourcing doesn't require Kafka clusters. Nor do event-driven setups. You don't need complex tooling to pass around strings/json-blurps. An S3 bucket or a Postgresql database storing "Events-as-json" is often fine.
* Postgres can do most of what you need (except for the "webscale" clustering etc)[0] in practice already.
* Redis[1]
My main point is that while Kafka is a fantastic tool, you don't need that tool to achieve what you want in many cases.
> It seems as good a way as any to decouple systems
IMO relying on a tool to achieve a good software design, rather than design-patterns, is a recipe for trouble. If anything, because it locks you in (do you suddenly get a tightly coupled system if you remove Kafka?) or because its details force you into directions that don't naturally fit your domain or problem.
--
[0] https://spin.atomicobject.com/2021/02/04/redis-postgresql/ [1] https://redis.com/redis-best-practices/communication-pattern... etc.
Lessened somewhat with SaaS products like Amazon Kinesis (technically not a Kafka, but close).
Another "baggage" is that an event-driven setup is eventual-consistent -and async- by nature. If your software already is eventual-consistent, this is not a problem. But it is a huge change if you come from a blocking/simple "crud" setup.
MY problems are so special that my use of Kafka was perfect, but YOURS are trivial and you shouldn’t even consider Kafka.
One of those things being “store it in a relational model” or “write a sum to a key value store” or something else.
This ability comes for free with Kafka, but is very not-free when using a relational model.
It's not missing that because it doesn't even address that. I'm answering a specific point.
But yeah, parallel programming is easy to explain with Factorio.
Transactions were kept by humans, literally for a few centuries, before the algorithm was adapted for computers.
It also only reflects one of the many tasks a developer usually does as an employee, that is creating algorithms. For a senior developer, that's probably not an apt task for differentiation, because it's quite bounded.
I consider this a form of gatekeeping of advice on using Kafka.
> Relational database seem to be a crazily overengineered solution in search of a problem
Why would an answer to that need to mention Kafka consumers?
This is the part I was responding to.
Having access to the current state of the world is useful, having a log of what happened / how it got that way is essential. You've got to get the foundations right before you build a monumental edifice on top.