Forcing people to justify, out loud, why they want to use a specific technology or trendy design pattern is usually sufficient to scuttle complex plans.

Frankly, many engineers want to use the latest trends like microservices or NoSQL because they believe that’s what’s best for their resume, even if it’s not necessarily best for the company. It doesn’t help that some companies screen out resumes that don’t have the right signals (Microservices, ReactJS, NoSQL, ...). There’s a certain amount of FOMO that makes early-career engineers feel like they won’t be able to move up unless they can find a way to use the most advanced and complex architectures, even if their problems don’t warrant those solutions.

Does that really work ?

Usually these guys read the sales pitch from some credible source. Then you need to show them that the argument is X works really well for scenario Y but your scenario Z is not really similar to Y so reasons why X is good for Y don't really apply. To do this you usually rely on experience so you need to expand even further.

And the other side is usually attached to their proposal and starts pushing back and because you're the guy arguing against something and need a deep discussion to prove your point chances are people give up and you end up looking hostile. Even if you win you don't really look good - you just shut someone down and spent a lot of time arguing, unless the rest of the team was already against the idea you'll just look bad.

I just don't bother - if I'm in a situation where someone gives these kind of people decision power they deserve what they get - I get paid either way. And if I have the decision making power I just shut it down without much discussion - I just invoke some version of 'I know this approach works and that's good enough for me'.

The sad thing is, they might well be right.

People used to not get hired for a job involving MySQL because their DB experience was with Postgres, but usually more enlightened employers knew better. Today, every major cloud provider offers the basic stuff like VMs and managed databases and scalable storage, and the differences between them are mostly superficial. However, each provider has its own terminology and probably its own dashboard and CLI and config files. Some of them offer additional services that manage more of the infrastructure for you one way or another, too. There is seemingly endless scope for not having some specific combination of buzzwords on an application even for a candidate and a position that are a good fit.

I don’t envy the generation who are applying for relatively junior positions with most big name employers today, and I can hardly blame them for the kind of job-hopping, résumé driven development that seems to have become the norm in some areas.

Funny you mentioned this. I have the exact opposite problem.

That is, I am an engineer trying to push back against management mandating the use of microservices and microfrontends because they are the new “hot” tech noawadays.

(I, too, am in the position of pushing back against microservices for hotness' sake.)

Problem is they no longer have to implement, so they are even more inclined to sell the most complicated tech stack that have marketing pages claiming they scale to basically infinity.

It took me weeks to convince my managers not to migrate to new hot nosql solution because "it's in cloud, it's scalable and it also supports sql queries".

Probably nobody is using NoSQL for their resume. It's because picking a relational database, while usually the correct choice, is HARD when you're operating in an environment that changes quickly and has poorly defined specifications.

When you start seeing engineers have difficulty reasoning about what the data model should be and nobody willing to commit to one, it's the clearest sign that organizationally things are sour and you need to start having very firm and precise conversations with product.

And using AWS generally means using Aurora. Then the choice is already made. Not hard at all.

Yep, my work involves heavy use of SQL and I find it better than the NoSQL insanity.

Postgres is more featureful, but if you don't intend on using those features, MySQL is consistently faster and historically smoother to update and keep running.

Unless you want to do a join.

Most often, you choose a database because of what you application supports and is tested with, not the other way around. Or what your other applications already use. Complete green fields aren't all that common.

And our DBAs are already familiar with the gotchas of MySQL.

You just have to write queries in a different way (subqueries are slow, so they are to be rewritten as joins).

Wouldn't this apply if you are using a static typed language too? what's harder about changing the schema in the DB?

Yes, and on top of that, code-only changes need to be internally consistent to make sense but DB schema changes almost inevitably require some corresponding code change as well to be useful. Then you have all the fun of trying to deploy both changes while keeping everything in sync and working throughout.

If only I had a nickel for every hour I've spent debugging abstractions and indirection that were just there for the sake of adding flexibility that would never be needed.

One might even convert SQL to binary query data at build time, with a code generator. It could work like PIDL, the code generator Samba uses to convert DCE/RPC IDL files to C source with binary data. Binary data goes over the wire. Another way is that both client code and server code could be generated, with the server code getting linked into the database server.

So I bought all the XP books, dog eared them, left them on my desk. My team nearly mutinied. I asked them to wait and see. Two weeks later, nemesis announced his team was all in for XP, Agile, pair programming, etc.

They never recovered, didn't make another release.

I tossed my copies, unread.

I love that.

I think many people are loathe to "turn the argument around", and pretend they're going the other way.

For example, imagine some legacy app used by 10 people out of 10,000 is incompatible with something like a Microsoft Office upgrade.

In many organisations, the argument goes like this: "We can't upgrade to Office 2023 because StupidApp will break!"

Turning that around: "If Office 2023 was already rolled out, would you roll that back to Office 2021 just for StupidApp?"

Then they are bad engineers. It is true that it’s best for their resume, but I also have my professional integrity to maintain.

Saying “why would you ever do that rather than building the solution the cheapest and with the lowest risk” doesn’t fully appreciate the importance of attractiveness.

I’m selling my hours of work time for a salary, fun, and resume points. My employer pays me in all 3. I’ll always push for $fun_tech or $hot_tech despite it not always being in the short term interest of anyone but myself or my fellow developers. I’ll keep justifying this by “if we do this in $old_tech then me and half the team will leave, and that’s a higher risk than using $new_tech”.

(By tech I here mean things like languages and frameworks not buzzwords like microservices or blockchain, ai... )

Go the simple way and it works, you get paid and off to the next project, if it doesn't work, its your fault.

But on the job experience with scaling tech is way more valued than doing some online course, you get paid to learn by doing on company time, and you don't lose anything by possibly wasting company resources. So you tick the box that often job postings have "proven track record of <insert scale buzz here>" which could possibly lead to a much better salary, its all incentives.

Things that are trivial in monoliths are hard in microservices like error propagation, profilling, line by line debbugging, log aggregation, orquestration, load balancing, health checking and ACID transactions.

It can be done but requires more complex machinery and larger teams.

The real skill of architecture is understanding everything your company has built before and finding the most graceful way to graft your new use case onto that. We get microservices proliferation because people don’t want to do this hard work.

All of this is in the assumption you need to scale horizontally. With modern hardware most systems don’t need that scalability. But it’s one of those “but what if we strike gold” things, where systems will be designed for a fantasy workload instead of a realistic one, because it’s assumed to be hard to go from a monolith to a microservice if that fantasy workload ever presents itself (imho not that hard if you have good abstractions inside the monolith).

Microservice architectures seem to be a recurring example of this phenomenon. Separating medium to long functions into shorter ones based on arbitrary metrics like line count or nesting depth is another.

For me the antipattern poses itself when the cutting up into microservices is done as a general practice, without a clearly defined goal for each service to need to be separate.

(And by the way, i’ve seen a talk before of an application where the entire backend was functions in a function store, exactly as you described. The developer was enthusiastic about that architecture.)

That's a common argument for microservices and one that I always thought was bunk.

What does that even mean? You have a piece of software that provides ten functions, running 100 instances of it in infeasible but running 100 of one, 50 of three and 10 of six is somehow not a problem?

That must be really the perfect margin call of some vsz hungry monstrosity. While not an impossible situation in theory, surely it can't be very common.

There are plenty of reasons to split an application but that seems unlikely at best.

In at least SOME of these cases, the monolith would've been breakable-up into a smaller number of front-end micro-services, with a graph of micro-services behind "the thing you talk to", for a lesser total deployed footprint.

But, I suspect that it requires that "the monolith" has been growing for 10+ years, as a monolith.

And that is the big if! The big advantage of micro services is that it forces developers to think hard about the abstractions, and can’t just reach over the boarder breaking them when they are in a hurry. With good engineers in a well functioning organisation, that is of course superfluous, but those preconditions are unfortunately much rarer than they should be.

I had a fevered dream the other night where it turned out that the bulk of AWS’s electricity consumption was just marshaling and unmarshalling JSON, for no benefit.

Anyway, along this journey I discovered that it's surprisingly difficult to get a HTTPS JSON RPC call below 3ms latency even on localhost! It's mindboggling how inefficient it actually is to encode every call through a bunch of layers, stuff it into a network stream, undo that on the other end, and then repeat on the way back.

Meanwhile, if you tick the right checkboxes on the infrastructure configuration, then a binary protocol between two Azure VMs can easily achieve a latency as low as 50 microseconds.

Why would I want to use a graph DB directly then?

Postgres really is a wonderful database.

For a good example of a high volume site using a proper rdbms approach I would look at stackoverflow. It can (and has) run on a single ms sql server instance.

I do know for Wikipedia, english wikipedia is mostly a single master mysql db + slaves, with most of the sharding being on the site language level (article text contents stored elsewhere)

But if you're doing more complicated query-like stuff, especially if you want to allow for queries you haven't thought of yet, then the DB might be useful.

Sometimes a hybrid of query-able metadata in a DB along with plain old data files is good.

That depends very much on your data, how much things key to each other, and what you're doing with it.

That's some kind of fallacy - standard datastructures would totally destroy any-sql-alike thing, if it comes to performance (and memory footprint). I guess it does depend on where the background comes when it comes to convenience - or how people tend to see their data. However like I said - for close to 3 decades I have not seen a single reason to do so. On the contrary I've had cases where optimization of 3 orders of magnitude was possible.

Evidently you don't have dataset far exceeding the amount of RAM you can afford.

For a good example look at LMDB.

All of the same reasons you would store relational data in a dbms...

There is no D from the ACID. For the D to happen, it takes transaction logs + write barrier (on the non-volatile memory). Doing Atomic, consistent and isolated is trivial in memory (esp. in GC setup), and a lot faster: no locks needed.

Validations and constraints are simple if-statements, I'd never think of them as sql.

You also have to do backup and recovery. And for that, you need to write to disk, which becomes a big bottleneck since besides backup and checkpointing there is no other reason to ever write to disk.

Then, you have to know that even in mem database, data needs to be queried, and for that you need special data structures like a cache aware B+tree. Implementing one is non trivial.

Thirdly, doing atomic, consistent and isolated transaction is certainly trivial in a toy example but in an actual database where you have a high number of transactions, it's a lot harder. For example, when you have multiple cores, you certainly will have resource contention, and then you do need locks.

And last thing about gc, again, gc is great, but there has to be a custom gc for a database. You need to make sure the transaction log in memory is flushed before committing. And malloc is also very slow.

I'd suggest reading more into in mem research to understand this better. But in mem db is certainly not the same as a disk db with cache or a simple Hashmat/B+tree structure.

Isn't one of the advantages of a GC environment that malloc is basically free? Afaik the implementation of malloc_in_gc comes down to

    result_address = first_free_address;
    first_free_address += requested_bytes;
    return result_address;

That's also the reason why, depending on the patterns of memory usage, a GC can be faster than malloc+free.

The original talk was explicitly about SqlLite and in-memory databases, no idea where you got the rest of.

https://docs.rocket.chat/installation/docker-containers/mong...

I'm happy to justify every single point I made with research papers.

Lastly I know I came off as a bit condescending. Just having a bad day, nothing personal. But you should read more about in mem dbs.

Universities taught SQL for years, so everyone knows it and its edge cases.

NoSQL databases are all different AND they weren't all taught for decades.

If you put real effort into learning a specific NoSQL database and it is suited for your problem things work out pretty well.

I find dynamodb to be unnecessary but I prefer nosql systems

Over the past decade I've realised using a RDBMS is the right call basically 100% of the time. Now pgsql has jsonb column types that work great, I cannot see why you would ever use a NoSQL DB, unless you are working at such crazy scale postgres wouldn't work. In 99.999% of cases people are not.

In retrospect we should have gone with mongo from the start, but postgres was chosen because in 99% of circumstances it is good enough. It was the wrong decision for the right reasons.

So really the use case for mongo etc is 'very high performance requirements' AND 'does not require relations'.

Many projects may be ok with just one of those. But very few require meet both of those constraints.

FWIW I've seen many cases which are sort of the opposite: great performance with mongodb, but then because of the lack of relations for a reporting feature (for example) performance completely plummets due to horrible hacks being done to query the data model that doesn't work with the schema, eventually requiring a rewrite to RDBMS. I would guess that this is much more common.

If your company needs to shutdown RDS to save a couple of bucks a month, there's a much larger problem at hand than RDS vs Dynamo.

As far as what makes Dynamo a good fit, I almost take the other approach and try to ask myself, what makes Postgres a bad fit? Postgres is so flexible and powerful that IMO you need a really good reason to walk away from that as the default.