Yep - saw that at a company recently - something in AWS was running a little slower than usual which cascaded to cause massive failures. Dozens of people were trying to get to the bottom of it, it mysteriously fixed itself and no one could offer any good explanation.

Is that the same as what Elixir/Erlang call supervision trees?

But to the point, if you're going to build a distributed system, you need tools to track problems across the distributed system that also works across teams. A poorly performing service could be caused by up/downstream components and doing that without some kind of tracing is hard even if your stack is linear.

The same is true for a giant monolithic app, but the sophisticated tools are just different.

There's very little reason to distribute application code. It's very, very rare that the limiting factor in an application is compute. Typically, it's the data layer, which you can change independently of your application.

But I think very few problems fit into that archetype. Instead, people build distributed systems for reliability and integrity. But it's overkill, because you bring all the baggage and complexity of distributed computing. This area is incredibly difficult. I view it similar to parallelism. If you can avoid it for your problem, then avoid it. If you really can't, then take a less complex approach. There's no reason to jump to "scale to X threads and every thread is unaware of where it's running" type solutions, because those are complex.

As someone using Kafka, I'd like to know what the (good) alternatives are if you have suggestions.

Where I'm at, most of Kafka usage adds nothing of note and could be replaced with a rest service. It sounds good that Kafka makes everything execute in order, but honestly just making requests block does the same thing.

At least then I could autoscale, which Kafka prevents.

I found it easy to get up and running, even as a RAFT cluster, but I have not tried to use JetStream mode heavily yet.

It works, with a lower cognitive burden than that of horizontally scaling.

For the loading concern (i.e. is this enough to handle the load):

For most businesses, being able to serve 20k concurrent requests is way more than they need anyway: an internal app used by 500k users typically has fewer than 20k concurrent requests in flight at peak.

A cheap VPS running PostgreSQL can easily handle that.[1]

For the "if something breaks" concern:

Each "fault-tolerance" criteria added adds some cost. At some point the cost of being resistant to errors exceeds the cost of downtime. The mechanisms to reduce downtime when the single large SQL server shits the bed (failovers, RO followers, whatever) can reduce that downtime to mere minutes.

What is the benefit to removing 3 minutes of downtime? $100? $1k? $100k? $1m? The business will have to decide what those 3 minutes are worth, and if that worth exceeds the cost of using something other than a single large SQL server.

Until and unless you reach the load and downtime-cost of Google, Amazon, Twitter, FB, Netflix, etc, you're simply prematurely optimising for a scenario that, even in the businesses best-case projections, might never exist.

The best thing to do, TBH, is ask the business for their best-case projections and build to handle 90% of that.

[1] An expensive VPS running PostgreSQL can handle a lot more than you think.

Not to forget: those costs are not just in money and time, but also in complexity. And added complexity comes with its own downtime risks. It's not that uncommon for systems to go down due to problems with mechanisms or components that would not exist in a simpler, "not fault tolerant" system.

That's still a business decision.

Customers don't vote with their feet based on what tech stack the business chose, they vote based on a range of other factors, few, if none, of which are related to 3m of downtime.

There are little to no services I know off that would lose customers over 3m of downtime per week.

IOW, 3m of downtime is mostly an imaginary problem.

Services that people might leave, because of downtime are for example a git hoster, or a password manager. When people cannot push their commits and this happens multiple times, they may leave for another git hoster. I have seen this very example when gitlab was less stable and often unreachable for a few minutes. When people need some credentials, but cannot reach their online password manager, they cannot work. They cannot trust that service to be available in critical moments. Not being able to access your credentials leaves a very bad impression. Some will look for more reliable ways of storing their credentials.

> The business can try to decide what those 3min are worth, but ultimately the customers vote by either staying or leaving that service.

What do you think the business is doing when it evaluates what 3 minutes are worth?

I don't understand, what people are arguing about here. Are we really arguing about customers making their own choice? Since that is all I stated. The business can jump up and down all it wants, if the customers decide to leave. Is that not very clear?

I think the point is that, for a few minutes of downtime, businesses lose so little customers that it's not worth avoiding that downtime.

Just now, we had a 5m period where disney+ stopped responding. We aren't going to cut off our toddler from peppa big and bluey for 5m of downtime per day, nevermind per week.

You appeared to be under the impression that 3m downtime/week is enough to make people leave. This is simply not true, especially for internet services where the users are conditioned to simply wait.

You saying “the customer can decide to leave” is not a counter to this at all. It’s just a weird way of saying what everyone else is but framing it as a counter argument to what is being said. Which it simply isn’t.

I've never, in my life, seen an error in SQL Server related to SQL Server. It's always been me, the app code developer.

Now, to be fair, the server itself or the hardware CAN fail. But having active/passive database configurations is simple, tried and tested.

I wish we would use something like Debian and take advantage of tech like systemd. But alas, we're still using COM and Windows Services and we still need to remote desktop in and click around on random GUIs to get stuff to work.

Luckily, SQL Server itself is very stable and reliable. But even SQL Server runs on Linux.

This is a very simple distinction and I am not sure why is it not understood

For some reason people design public apps the same as internal apps

The largest companies employ circa 1 million people - that’s Walmart, Amazon, etc. most giants, like Shell, etc. companies have ~ 100k tops. That can be handled by 1 beefy server.

Successful consumer facing apps are hundred millions to billion. It’s 3 orders of magnitude difference

I have seen a company with 5k employees invest into mega-scalable microservice event driven architecture and I was was thinking - I hope they realise what they are doing and it’s just CV-driven development

Agreed, but the cost/benefit should be the same at every level in the stack. I get it when people build a single-server system. I'm just baffled that so many people insist they need a load balancer and multiple instances of their application so that there's no single point of failure (which is not free by any means), but then run them all off of a single SQL server.

There is really nothing wrong with a large vertically scaled up SQL server. You need to be either really really large large scale - or really really UNSKILLED in sql as to keep your relational model and working set in SQL so bad that you reach its limits

Sadly that's the case at my current job. Zero thought put into table design, zero effort into even formatting our stored procedures in a remotely readable way, zero attempts to cache data on the application side even when it's glaringly obvious. We actually brought in a consultant to diagnose our SQL Server performance issues (I'm sure we paid a small fortune for that) and the DB team and all of the other higher ups capable of actually enforcing change rejected every last one of his suggestions.

Large or small, it's always going to be a single point of failure. All hardware is fallible, especially if we're talking about a commodity box rather than a mainframe or something.

It's less true today when redundancy is baked into SaS products (like AWS Aurora, where even if you have a single database instance, it's easy to spin up a replacement one if the hardware on the one running fails).

One of the most stable system archictectures I've built was on Kafka AND it was running with microservices managed by teams across multiple geographies and time zones. It was one of the most reliable systems in the bank. There are situations where it isn't appropriate, which can be said for most tech e.g( K8S vs ECS vs Nomad vs bare metal)

Every system has failure characteristics. Kafka's is defined as Consistent and Available and your system architecture needs to take that into consideration. Also the transactionality of tasks across multiple services and process boundaries is important.

Let's not pretend that kubernetes (or the tech of the day) is at fault while completely ignoring the complex architectural considerations that are being juggled

From my experience most microservices aren’t engineered to handle back pressure. If there is a sudden upsurge in traffic or data the Kafka cluster is expected to absorb all of the throughput. If the cluster starts having IO issues then literally everything in your “distributed” application is now slowly failing until the consumers/brokers can catch up.

The classic dependency loop example that you thought will never encounter again for the rest of your life after OS class

It's very rare to meet a developer who has even the vaguest notion of what an RPC call costs in terms of microseconds.

Fewer still that know about issues such as head-of-line blocking, the effects of load balancer modes such as hash versus round-robin, or the CPU overheads of protocol ser/des.

If you have an architecture that involves about 5 hops combined with sidecars, envoys, reverse proxies, and multiple zones you're almost certainly spending 99% to 99.9% of the wall clock time just waiting. The useful compute time can rapidly start to approach zero.

This is how you end up with apps like Jira taking a solid minute to show an empty form.

I think Web apps have been around so long now people have forgotten how unresponsive things are vs old 2 tier stuff.

I tested the Jira cloud service with a new, blank account. Zero data, zero customisations, zero security rules. Empty.

Almost all basic operations took tens of seconds to run, even when run repeatedly to warm up any internal caches. Opening a new issue ticket form was especially bad, taking nearly a minute.

Other Atlassian excuses included: corporate web proxy servers (I have none), slow Internet (gigabit fibre), slow PCs (gaming laptop on "high performance" settings), browser security plugins (none), etc...

At that point, something mustve been wrong with your instance. I'd never call jira fast, but the new ticked dialog on a unconfigured instance opens within <10s (which is absolutely horrendous performance to be clear. Anything more then 200-500ms is.)

it's open within 1-2 sec (which is still bad performance, objectively speaking. It's an empty instance after all and already >1s )

It's still sluggish compared to a desktop app from the 1990s, but it's much faster than just a couple of years ago.

To be fair, small time units are difficult to internalize. Just look at what happens when someone finds out that it takes tens of nanoseconds to call a C function in Go (gc). They regularly conclude that it's completely unusable, and not just in a tight loop with an unfathomable number of calls, but even for a single call in their program that runs once per day. You can flat out tell another developer exactly how many microseconds the RPC is going to add and they still aren't apt to get it.

It is not rare to find developers who understand that RPC has a higher cost than a local function, though, and with enough understanding of that to know that there could be a problem if overused. Where they often fall down, however, is when the tools and frameworks try to hide the complexity by making RPC look like a local function. It then becomes easy to miss that there is additional overheard to consider. Make the complexity explicit and you won't find many developers oblivious to it.

I'm not too familiar with Go but my default assumption is that it's just used as a convenient excuse to avoid learning how to do FFI.

Nah. Jira was horrible and slow long before the microservice trend.

In fact, I'd argue that one of the biggest annoyances I had in last job was that I actually missed JIRA, or rather it's glorious issue list view with its keyboard shortcuts.

So really eventually we'll all be optimizing around network interfaces as the bottleneck.

I agree on how it's technically a waste of time to pursue fads, but it's also a huge PITA to have a platform that good engineers actively try to avoid, as their careers would stagnate (even as they themselves know that it's half a fad)

1. Has been at org for so long, management condones them flaunting the rules, like pushing straight to prod. Hates the "inefficiency" of open source platforms and purpose-built something "suitable for the company" by themselves, no documentation you have to ask them to fix issues because they don't accept code or suggestions from others. The DSL they developed is inconsistent and has no parser/linter.

absoutely start as simple as you can, but plan to move to a hosted kube something asap instead of writing your own base images, unless that's a differentiator for your company.

2. Fad-driven

I wonder why I don’t often see (1) critiqued on the basis of (2).

Of course, it was because the client app recently went over the 100MB JS budget. Which they decided to make because the last time that happened, customers abroad reported seeing “white screens”. International conversion dropped sharply not long after that.

It’s pretty silly. So ya, good times indeed. Time to learn k8s.

They already has their architecture, largely, and just moved it over to K8s.

They even mention they aren’t a micro service company.

to illustrate performance cost i usually ask people what ping they have say from one component/pod/service to another and to compare that value to what ping time they'd get between 2 linux boxes sitting on that gorgeous 10Gb/40Gb/100Gb or even 1000Gb network that they are running their modern microservices architecture over.

> More broadly, we’re not a microservices company, and we don’t plan to become one

Is this generally correct? How well is this term defined?

If yes, then I'm not surprised this type of design has become a target for frustration. State is smeared across system, which implies a lot of messaging and arbitrary connections between services.

That type of design is useful if you are an application platform (or similar) where you have no say in what the individual entities are, and actually have no idea what they will be.

But if you have the birds-eye view and implement all of it, then why would you do that?

And the "test cases" porn! Goodness! People want "coverage" and that's it. It's a box to tick, irrespective of whether those test cases and they way those are written are actually meaningful in anyway. There are more like "let's have dependency injection everywhere" charade.

    grug wonder why big brain take hardest problem, factoring system correctly, and introduce network call too

    seem very confusing to grug

So they didn't do microservices correctly. Big surprise.

There are certainly ways to mitigate the SPOFiness of each of those cases, but that doesn’t make having them an antipattern.

Personally, i think, it does not have to be hundreds of microservices, basically each function a service. But i see it more as the web in the internet. Things are sometimes not reachable or overloaded. I think that is normal life.