It was a fun moment to finally work on a data problem that did not fit on any (practical) machine. I needed about 50TiB of memory to process a multi-PiB set of logs.
It's worth remembering however that even though it's less efficient per-CPU or whatever to split a large task into many smaller tasks, it may be more efficient overall alongside other workloads as you can bin-pack tasks more efficiently on a cluster, not to mention if tasks fail you are retrying less of the overall work.
All this is to say, the article makes a very good point, but doing it all on one machine also has problems. Just don't cargo cult engineering decisions.
The saddest part about this article being from 2014 is that the situation has arguably gotten worse.
We now have even more layers of abstraction (Airflow, dbt, Snowflake) applied to datasets that often fit entirely in RAM.
I've seen startups burning $5k/mo on distributed compute clusters to process <10GB of daily logs, purely because setting up a 'Modern Data Stack' is what gets you promoted, while writing a robust bash script is seen as 'unscalable' or 'hacky'. The incentives are misaligned with efficiency.
I've done a handful of interviews recently where the 'scaling' problem involves something that comfortably fits on one machine. The funniest one was ingesting something like 1gb of json per day. I explained, from first principals, how it fits, and received feedback along the lines of "our engineers agreed with your technical assessment, but that's not the answer we wanted, so we're going to pass". I've had this experience a good handful of times.
I think a lot of people don't realize machines come with TBs of RAM and hundreds of physical cores. One machine is fucking huge these days.
The wildest part is they’ll take those massive machines, shard them into tiny Kubernetes pods, and then engineer something that “scales horizontally” with the number of pods.
Containers are just processes plus some namespacing, nothing really stops you from running very huge tasks on Kubernetes nodes. I think the argument for containers and Kubernetes is pretty good owing to their operational advantages (OCI images for distributing software, distributed cron jobs in Kubernetes, observability tools like Falco, and so forth).
So I totally understand why people preemptively choose Kubernetes before they are scaling to the point where having a distributed scheduler is strictly necessary. Hadoop, on the other hand, you're definitely paying a large upfront cost for scalability you very much might not need.
Time to market and operational costs are much higher on kubernetes and containers from many years of actual experience. This is both in production and in development. It’s usually a bad engineering decision. If you’re doing a lift and shift, it’s definitely bad. If you’re starting greenfield it makes sense to pick technology stacks that don’t incur this crap.
It only makes sense if you’re managing large amounts of large siloed bits of kit. I’ve not seen this other than at unnamed big tech companies.
99.9% of people are just burning money for a fashion show where everyone is wearing clown suits because someone said clown suits are good.
Thanks. You’ve reassured me that I’m not going mad when I look at our project repo and seriously consider binning the Dockerfile and deploying direct to Ubuntu.
The project is a Ruby on Rails app that talks to PostreSQL and a handful of third party services. It just seems unnecessary to include the complexity of containers.
I have a lot of years of actual experience. Maybe not as much as you, but a good 12 years in the industry (including 3 at Google, and Google doesn't use Docker, it probably wouldn't be effective enough) and a lot more as a hobbyist.
I just don't agree. I don't find Docker too complicated to get started with at all. A lot of my projects have very simple Dockerfiles. For example, here is a Dockerfile I have for a project that has a Node.JS frontend and a Go backend:
FROM node:alpine AS npmbuild
WORKDIR /app
COPY package.json package-lock.json .
RUN npm ci
COPY . .
RUN npm run build
FROM golang:1.25-alpine AS gobuilder
WORKDIR /app
COPY go.mod go.sum .
RUN go mod download
COPY . .
COPY --from=npmbuild /app/dist /app/dist
RUN go build -o /server ./cmd/server
FROM scratch
COPY --from=gobuilder /server /server
ENTRYPOINT ["/server"]
It is a glorified shell script that produces an OCI image with just a single binary. There's a bit of boilerplate but it's nothing out of the ordinary in my opinion. It gives you something you can push to an OCI registry and deploy basically anywhere that can run Docker or Podman, whether it's a Kubernetes cluster in GCP, a bare metal machine with systemd and podman, a NAS running Synology DSM or TrueNAS or similar, or even a Raspberry Pi if you build for aarch64. All of the configuration can be passed via environment variables or if you want, additional command line arguments, since starting a container very much is just like starting a process (because it is.)
But of course, for development you want to be able to iterate rapidly, and don't want to be dealing with a bunch of Docker build BS for that. I agree with this. However, the utility of Docker doesn't really stop at building for production either. Thanks to the utility of OCI images, it's also pretty good for setting up dev environment boilerplate. Here's a docker-compose file for the same project:
And if your application is built from the ground up to handle these environments well, which doesn't take a whole lot (basically, just needs to be able to handle configuration from the environment, and to make things a little neater it can have defaults that work well for development), this provides a one-command, auto-reloading development environment whose only dependency is having Docker or Podman installed. `docker compose up` gives you a full local development environment.
I'm omitting a bit of more advanced topics but these are lightly modified real Docker manifests mainly just reformatted to fewer lines for HN.
I adopted Kubernetes pretty early on. I felt like it was a much better abstraction to use for scheduling compute resources than cloud VMs, and it was how I introduced infrastructure-as-code to one of the first places I ever worked.
I'm less than thrilled about how complex Kubernetes can be, once you start digging into stuff like Helm and ArgoCD and even more, but in general it's an incredible asset that can take a lot of grunt work out of deployment while providing quite a bit of utility on top.
I can maybe make a case for running in containers if you need some specific security properties but .. mostly I think the proliferation of 'fucked up piles of shit' is the problem.
Are you saying that running your application in a pile of containers somehow helps that problem ..? It's the same problem as CPU scheduling, we just don't have good schedulers yet.. Lots of people are working on it though
To be fair each of those pods can have dedicated, separate external storage volumes which may actually help and it’s def easier than maintaining 200 iscsi or more whatever targets yourself
I recently had to parse 500MB to 2GB daily log files into analytical information for sales. Quick and dirty, the application would of needed 64GB RAM and work laptop only has 48GB RAM. After taking time cleaning it up, it was using under 1GB of RAM and worked faster by only retaining records in RAM if need be between each day.
It is not about what you are doing, it is always about how you do it.
This was the same with doing OCR analysis of assembly and production manuals. Quick and dirty, it would of took over 24 hours of processing time, after moving to semaphores with parallelization it took less than two hours to process all the information.
In interviews just give them what they are looking for. Don't overthink it. Interviews have gotten so stupidly standardized as the industry at large copied the same Big Tech DSA/System Design/Behavioral process. And therefore interview processes have long been decoupled from the business reality most companies face. Just shard the database and don't forget the API Gateway
100%. Interviews should be a two-way filter. I’m sympathetic to unemployed-and-just-need-something, but also: boy are there a lot of companies hiring data engineers.
Meh .. I've played that game; it doesn't work out well for anyone involved.
I optimize my answers for the companies I want to work for, and get rejected by the ones I don't. The hardest part of that strategy is coming to terms with the idea that I constantly get rejected by people that I think are mostly <derogatory_words_here>, but I've developed thick skin over the years.
I'd much rather spend a year unemployed (and do a ton of painful interviews) and find a company who's values align with mine, than work for a year on a team I disagree with constantly and quit out of frustration.
The company's values may align to yours, even though they reject you. It's because the interview process doesn't need to have anything to do with their real-world process. Their engineers probe you for the same "best practices" that they themselves were constantly probed for in their own interviews. Interviewing is its very own skill that doesn't necessarily translate into real-life performance.
I agree with your observation. My issue is (from experience) it's really hard to tell from the outside if a teams' values align with mine. Many teams talk the talk, but don't walk the walk, as the saying goes. It's just easier to not participate than it is to guess, and be wrong.
I also believe that running a broken interview process actively selects for qualities you actually don't want, so it's much more likely that teams conducting those interviews aren't teams I want to work on.
Edit: As credence for my claims, the best team I've ever worked on was a team I did 90%+ of the hiring for, and we didn't do any of the 'typical' interview bullshit most companies do.
What we did instead was sit people down and have deep technical conversations about systems they'd worked on in the past. The candidate would explain, in as much detail as they could muster, a system they'd worked on in the past, down to the lowest level details. Usually, they would talk to us for at least 20-30 minutes, then, we (the interviewers) would pose questions, usually starting with the form 'if we changed X, what effect would it have'. Doing interviews in this style make a few things immediately obvious:
1. Did the candidate have a deep, systemic understanding of the system they worked on?
2. Does the candidate have a good mental model for evaluating change in the system?
That's how I conduct interviews, and unsurprisingly, when I get interviewed like that, my success rate is 100%. I don't think I've ever done an interview like that which did not result in an offer.
Anyways, there's some rambling and unsolicited opinions for you :)
The interview process determines who gets hired, which determines their real-world process. Even if most of their people were hired under a better system, future hires will come in under this one.
This. Most interviewers don't want to do interviews, they have more important job to do (at least, that's what they claim). So they learn questions and approaches from the same materials and guides that are used by candidates. Well, I'm guilty of doing exactly this a few times.
You could have learned this if you were better about collecting requirements. You can tell the interviewer "I'd do it like this for this size data, but I'd do it like this for 100x data. Which size should I design this for?" If they're looking for one direction and you ask which one, interviewers will tell you.
I've done that too and, in my experience, people that ask a scaling question that fits on a single machine don't have the capacity to have that nuanced conversation. I usually try to help the interviewer adjust the scale to something that actually requires many machines, but they usually don't get it.
Said another way, how do you have a meaningful conversation about scaling with a person who thinks their application is huge, but in reality only requires a tiny fraction of a single machine? Sometimes, there's such a massive gulf between perception and reality that the only thing to do is chuckle and move on.
Yes, but then how are these people going to justify the money they're spending on cloud systems?... They need to find only reasons to maintain their "investment", otherwise they could be held as incompetent when their solution is proven to be ineffective. So, they have to show that it was a unanimous technical decision to do whatever they wanted in the first place.
I have a funny story I need to tell some day about how I could get a 4GB JSON loaded purely in the browser at some insane speed, by reading the bytes, identifying the "\n" then making a lookup table. It started low stakes but ended up becoming a multi-million internal project (in man-hours) that virtually everyone on the company used. It's the kind of project that if started "big" from the beginning, I'd bet anything it wouldn't have gotten so far.
Edit: I did try JSON.parse() first, which I expected to fail and it did fail BUT it's important that you try anyway.
> I explained, from first principals, how it fits, and received feedback along the lines of "our engineers agreed with your technical assessment, but that's not the answer we wanted, so we're going to pass". I've had this experience a good handful of times.
Probably a better outcome than being hired onto a team where everyone know you're technically correct but they ignore your suggestions for some mysterious (to you) reason.
Yeah I had this problem at a couple of times in startup interviews where the interviewer asked a question I happened to have expertise in and then disagreed with my answer and clearly they didn't know all that much about it. It's ok, they did me a favor.
It may or may not be related that the places that this happened were always very ethnically monotone with narrow age ranges (nothing against any particular ethnic group, they were all different ethnic monotones)
“there’s no wrong answer, we just want to see how you think” gaslighting in tech needs to be studied by the EEOC, Department of Labor, FTC, SEC, and Delaware Chancery Court to name a few
let’s see how they think and turn this into a paid interview
Because developers are incentivized to have marketable software skills. Not marketable build things that are cheap and profitable skills.
Moore's law was supposed to make it simpler and cheaper to do more computationally expensive tasks. But in the meantime, everyone kept inflating the difficulty of a task faster than Moore could keep up.
I think some of this is because of the incredible amounts of capital that startups seem to be able to acquire. If startups had to demonstrate profitability before they were given any money to scale, the story would be very different I think.
I agree - and it's not just what gets you promoted, but also what gets you hired, and what people look for in general.
You're looking for your first DevOps person, so you want someone who has experience doing DevOps. They'll tell you about all the fancy frameworks and tooling they've used to do Serious Business™, and you'll be impressed and hire them. They'll then proceed to do exactly that for your company, and you'll feel good because you feel it sets you up for the future.
Nobody's against it. So you end up in that situation, which even a basic home desktop would be more than capable of handling.
I have been the first (and only) DevOps person at a couple startups. I'm usually pretty guilty of NIH and wanting to develop in-house tooling to improve productivity. But more and more in my career I try to make boring choices.
Cost is usually not a huge problem beyond seed stage. Series A-B the biggest problem is growing the customer base so the fixed infra costs become a rounding error. We've built the product and we're usually focused on customer enablement and technical wins - proving that the product works 100% of the time to large enterprises so we can close deals. We can't afford weird flakiness in the middle of a POC.
Another factor I rarely see discussed is bus factor. I've been in the industry for over a decade, and I like to be able to go on vacation. It's nice to hand off the pager sometimes. Using established technologies makes it possible to delegate responsibility to the rest of the team, instead of me owning a little rats nest fiefdom of my own design.
The fact is that if 5k/month infra cost for a core part of the service sinks your VC backed startup, you've got bigger problems. Investors gave you a big pile of money to go and get customers _now_. An extra month of runway isn't going to save you.
The issue is when all the spending gets you is more complexity, maintenance, and you don't even get a performance benefit.
I once interviewed with a company that did some machine learning stuff, this was a while back when that typically meant "1 layer of weights from a regression we run overnight every night". The company asked how I had solved the complex problem of getting the weights to inference servers. I said we had a 30 line shell script that ssh'd them over and then mv'd them into place. Meanwhile the application reopened the file every so often. Zero problems with it ever. They thought I was a caveman.
I work for a small company with a handful of devs. We don't have a dedicated devops person, so I do it all. Everything is self-hosted. Been that way for years. But, yeah, if I go on vacation and something foes screwy, the business is hosed. However, even if it were hosted on AWS or elsewhere, it would not be any better. If anything, it may be worse. Instead of a person being well versed in standards based tech, they'd have to be an AWS expert. Why would we want that?
I have recently started using terraform/tofu and ansible to automate nearly all of the devops operations. We are at a point where Claude Code can use these tools and our existing configs to make configuration changes, debug issues by reviewing logs etc. It is much faster at debugging an issue than I am and I know our stuff inside and out.
I am beginning to think that AI will soon force people to rethink their cloud hosting strategy.
Those scripts have logs, right? Log a hostname and path when they run. If no one thinks to look at logs, then there's a bigger problem going on than a one-off script.
You can literally have a 20 line Python script on cron that verifies if everything ran properly and fires off a PagerDuty if it didn't. And it looks like PagerDuty even supports heartbeat so that means even if your Python script failed, you could get alerted.
That becomes a problem if you let the shell script mutate into an "everything" script that's solving tons of business problems. Or if you're reinventing kubernetes with shell scripts. There's still a place for simple solutions to simple problems.
Which is why you take the time to put usage docs in the repo README, make sure the script is packaged and deployed via the same methods that the rest of the company uses, and ensure that it logs success/failure conditions. That's been pretty standard at every organization I've been at my entire professional career. Anyone who can't manage that is going to create worse problems when designing/building/maintaining a more complex system.
Yah. A lot of the complexity in data movement or processing is unneeded. But decent standardized orchestration, documentation, and change management isn't optional even for the 20 line shell script. Thankfully, that stuff is a lot easier for the 20 line standard shell script.
Or python. The python3 standard library is pretty capable, and it's ubiquitous. You can do a lot in 50-100 lines (counting documentation) with no dependencies. In turn it's easy to plug into the other stuff.
I've seen the ramifications of this "CV first" kind of engineering. Let's just say that it's a bad time when you're saddled with tech debt solely from a handful of influential people that really just wanted to work elsewhere.
I have hung on to my job for many years now because of being in a similar situation in regards to trying to do the right thing and the fear of not being hire-able.
There is something wrong with the industry in chasing fads and group think. It has always been this way. Businesses chased Java in the late 90s, early 00s. They chased CORBA, WSDL, ESB, ERP and a host of other acronyms back in the day.
More recently, Data Lake, Big Data, Cloud Compute, AI.
Most of the executives I have met really have no clue. They just go with what is being promoted in the space because it offers a safety net. Look, we are "not behind the curve!". We are innovating along with the rest of the industry.
Interviews do not really test much for ability to think and reason. If you ran an entire ISP, if you figured out, on your own, without any help, how to shard databases, put in multiple layers of redundancy, caching... well, nobody cares now. You had to do it in AWS or Azure or whatever stack they have currently.
Sadly, I do not think it will ever be fixed. It is something intrinsic to human nature.
Try Rust? The system programming world isn't very bullshit-infested and Rust is trendy (which is good for a change), also employers can't realistically expect many years of Rust experience.
Need training and something to show? Contribute to some FOSS project.
This exactly, actual doers are most of the time not rewarded meanwhile the AWS senior sucking Jeffs wiener specialist gets a job doing nothing but generating costs and leave behind more shit after his 3 years moving the ladder up to some even bigger bs pretend consulting job at an even bigger company. It's the same bs mostly for developers. I rewrite their library from TS to Rust and it gains them 50x performance increases and saves them 5k+ a week over all their compute now but nobody gives a shit and I do not have a certification for that to show off on my LinkedIn. Meanwhile my PM did nothing got paid to do some shity certificate and then gets the credit and the certificate and pisses of to the next bigger fish collecting another 100k more meanwhile I get a 1k bonus and a pat on the shoulder. Corporate late stage capitalism is complete fucking bs and I think about becoming a PM as well now. I feel like a fool and betrayed. Meanwhile they constantly threaten our Team to lay it off or outsource it as they say we are to expensive in a first world country and they easily find as good people in India etc. What a time to be alive.
If you're willing and able to promote yourself internally, you can make people give a shit, or at least publicly claim they do. That's 260k+ per year, and even big businesses are going to care about that at some level, especially if it's something that can be replicated. Find 10 systems you can do that with, and it's 2.6m+ per year.
But, if you don't want to play the self-promotion game, yeah someone else is going to benefit from your work.
Yep, and a lot more datasets fit entirely into RAM now. Ignoring the recent price spikes for a moment, 128GB of RAM in a laptop is entirely achievable and not even the limit of what is possible. That was a pipe dream in 2014 when computers with only 4GB were still common. And of course for servers the max RAM is much higher, and in a lot of scenarios streaming data off a fast local SSD may be almost as good.
I have actually worked in a company as a consultant data guy in a non technical team, I had a 128 GB PC 10 years back, and did everything with open source R then, and it worked ! The others thought it was wizardry
I think it’s not so much engineers actually setting up a distributed compute, as it is dropping a credit card into a paid cloud service, which behind the scenes sets up a distributed compute cluster and bills you for the compute in an obfuscated way, then gives a 20% discount + SSO if you sign up for annual enterprise plan.
This kind of practice is insidious because early on, they charge $20/month to get started on the first 100mb of log ingestion, and you can have it up and running in 30 seconds with a credit card. Who would turn that down?
Revisit that set up 2 years later and it’s turned into a 60k/y behemoth that no one can unwind
I’ve seen this pattern play out before. The pushback on simpler alternatives seems from a legitimate need for short time to market from the demand some of the equation and a lack of knowledge on the supply side. Every time I hear an engineer call something hacky, they are at the edge of their abilities.
systemd would be a derail even if you weren’t misrepresenting the situation at several levels. Experienced sysadmins in my experience were the ones pushing adoption because they had to clean up the messes caused by SysV’s design limitations and flaws, whereas in this case it’s a different scenario where the extra functionality is both unneeded and making it worse at the core task.
> Experienced sysadmins in my experience were the ones pushing adoption because they had to clean up the messes caused by SysV’s design limitations and flaws
That's funny. I used to have to clean up the messes caused by systemd's design limitations and flaws, until I built my own distro with a sane init system installed.
Many of the noobs groaning about the indignity of shell scripts don't even realize that they could write init 'scripts' in whatever language they want, including Python (the language these types usually love so much, if they do any programming at all.)
It's entirely possible that both SysV init and systemd suck for different reasons. I'm still partial to systemd since it takes care of daemons and supervision in a way that init does not, but I'll take s6 or process-compose or even supervisord if I have to. Horses for courses.
I want to love s6 but every time I see the existence of s6-rc-compile I get heated. I'm sure there are excellent reasons behind it but I personally don't want services to work that way.
I think you’d have a more fruitful discussion if you stopped trying to call people noobs when they don’t agree with you.
For example, I’ve been dealing with SysV since the early 90s and while it’s gotten better since we no longer have to support the really bizarre Unix variants, my problem with init scripts wasn’t “indignity” but the lack of consistency across distributions and versions, which affects anyone shipping software professionally (“can’t do this easily until $distro upgrades coreutils”), and from an operator’s perspective using Python doesn’t make that better because instead of supporting one consistent thing you’d end up with the subset of features each application team felt like implementing, consistent only to the extent that they care to follow other projects. One virtue of systemd is that having a single common way to specify dependencies, restarts, customization, etc. avoids the ops people having to learn dozens of different variations of the same ideas and especially how to deal with their gaps. A few years back, a data center power outage at one place I worked really highlighted that: the systemd-based servers recovered quickly because they actually had working retries; all of the older stuff using SysV had to be manually reviewed because there were all kinds of problems like races on dependencies like DNS or NFS, retry logic which failed hard after a short period of time, failures because a stale PID file wasn’t removed, or cases where a vendor had simply never implemented retries in their init scripts. While in theory you can handle all of those in SysV most people never did.
After a couple decades of that, a lot of us don’t want to spend time on problems Microsoft solved in Bill Clinton’s first term.
I hate to blather on about systemd in this decade but how in the world does creating something completely different than sysv init help people shipping software? Now they have to support yet another init scheme.
Prior to all of the important distributions consolidating on systemd, you had to support each distribution’s convention for customization, overrides, dependencies, conventions for things like changing users or locations for PID files, not to mention the differences in various shell tools.
Nothing insurmountable but it meant init files were inevitably much longer than the corresponding Upstart or systemd files despite doing less, and anytime we shipped a new version you had more testing since you had to implement a lot of functionality which is built in to other things.
I just created my own OS, with my own init system that does things how I think it should be done--and it does it every time, without the bizarre bugs that come from Linux Puttering's shitware code.
It's the same thing any corporation should be doing if they were smart, instead of outsourcing everything to RedHat, Microsoft, Google, etc.
The reality is unit files are more portable than init scripts, regardless of what anyone says.
Systemd unified and simplified administration across a lot of distributions. Before, it was a hodge podge, and there was a lot of knowledge lost going from rhel to Debian.
Specifying system processes and their dependencies declaratively, rather than in a tangle of arbitrary executable code, is cleaner, more efficient, easier to use, and more auditable. And that's not even getting into the additional process management duties systemd assumes.
You can write arbitrary scripts into systemd... or like one step removed at most? That's not really a difference unless you have some nuance in mind that I don't.
I honestly do not like systemd, either. It is okay for managing processes but I wish it didn't spread into everything else in the machine.
Or if it must, could it actually work cohesively across their concepts? Would be nice to have an obvious and easy way to run Quadlet as its own user to isolate further, would be nice to have systemd-sysusers present in /etc/subuid so they can run containers.
I like what they are doing with atomic distros. It would be great to have a single file declarative setup for something like running a containerized reverse HTTP proxy with an isolated user. Instead of "atomic" but you manually edit files in /etc after install.
Worse in some ways, better in others. DuckDB is often an excellent tool for this kind of task. Since it can run parallelized reads I imagine it's often faster than command line tool, and with easier to understand syntax
More importantly, you have your data in a structured format that can be easily inspected at any stage of the pipeline using a familiar tool: SQL.
I've been using this pattern (scripts or code that execute commands against DuckDB) to process data more recently, and the ability to do deep investigations on the data as you're designing the pipeline (or when things go wrong) is very useful. Doing it with a code-based solution (read data into objects in memory) is much more challenging to view the data. Debugging tools to inspect the objects on the heap is painful compared to being able to JOIN/WHERE/GROUP BY your data.
Yep. It’s literally what SQL was designed for, your business website can running it… the you write a shell script to also pull some data on a cron. It’s beautiful
Pipes are parallelized when you have unidirectional data flow between stages. They really kind of suck for fan-out and joining though. I do love a good long pipeline of do-one-thing-well utilities, but that design still has major limits. To me, the main advantage of pipelines is not so much the parallelism, but being streams that process "lazily".
On the other hand, unix sockets combined with socat can perform some real wizardry, but I never quite got the hang of that style.
Pipelines are indeed one flow, and that works most of the time, but shell scripts make parallel tasks easy too. The shell provides tools to spawn subshells in the background and wait for their completion. Then there are utilities like xargs -P and make -j.
UNIX provides the Makefile as go-to tool if a simple pipeline is not enough. GNUmake makes this even more powerful by being able to generate rules on-the-fly.
If the tool of interest works with files (like the UNIX tools do) it fits very well.
If the tool doesn't work with single files I have had some success in using Makefiles for generic processing tasks by creating a marker file that a given task was complete as part of the target.
On the contrary, the key message from the blog post is not to load the entire dataset to RAM unless necessary. The trick is to stream when the pattern works. This is how our field routinely works with files over 100GB.
For a dasaset that live in RAM, the best solution are DuckDB or clickhouse-local.
Using SQLish data is easier than a bunch of bash script and really powerful.
I see this at work too. They are ingesting a few GB per day but running the data through multiple systems. So the same functionality we delivered with a python script within a week now takes months to develop and constantly breaks.
Our lot burns a fortune on snowflake every month but no one is using it. Not enough data is being piped into it and the shitty old reports we have which just run some SQL work fine.
It looked good on someone’s resume and that was it. They are long gone.
The issue is you can run sub tib jobs on a few small/standard instances with better tooling. Spark and Hadoop are for when you need multiple machines.
Dbt and airflow let you represent your data as a DAG and operate on that, which is critical if you want to actually maintain and correct data issues and keep your data transforms timely.
edit: a little surprised at multiple downvotes. My point is, you can run airflow and dbt on small instances, and you can do all your data processing on small instances with tools like duckdb or polars.
But it is very useful to use a tool like dbt that allows you to re-build and manage your data in a clear way, or a tool like airflow which lets you specify dependencies for runs.
After say 30 jobs or so, you'll find that being able to re-run all downstreams of a model starts to payoff.
Agreed, airflow and dbt have literally nothing to do with the size of the data and can be useful, or overkill, at any size. Dbt just templates the query strings we use to query the data and airflow just schedules when we query the data and what we do next. The fact that you can fit the whole dataset in duckdb without issue is kind of separate to these tools, we still need to be organised about how and when we query it.
dbt is super useful for building a dag and managing pieces of it that update on different schedules. eg with one dataset that's refreshed monthly and another daily, you can only rebuild the daily one unless the slower-cadence input has a new update.
Well. I try for a middle ground. I am currently ditching both airflow and dbt. In Snowflake, I use scheduled tasks that call stored procedures. The stored procedures do everything I need to do. I even call external APIs like Datadog’s and Okta’s and pull down the logs directly into snowflake. I do try to name my stored procedures with meaningful names. I also add generous comments including urls back to the original story.
I forgot to mention in Snowflake, besides chron scheduled tasks, you can add dependent tasks that only run if the previous task succeeded. I have 40 tasks chained together that way. Each of my task calls a stored procedure. Within each procedure, I have Try Catch and a catch-all clause that Raiseerror.
On the other hand, now we have duckdb for all the “small big data”, and a slew of 10-100x faster than Java equivalent stuff in the data x rust ecosystem, like DataFusion, Feldera, ByteWax, RisingWave, Materialize etc
I call BS on those Rust 10-100x claims. Rust and Java are roughly equal in performance. It is just that there are a lot of old NoSQL frameworks in Java which are trash. I also checked out those companies, some of which are doing interesting stuff. None claim things are 100x faster because of Rust. You just hurt your credibility when you say such clearly false things. That's how you end up with a Hadoop cluster which is 236x slower than a batch script.
PS None of the companies you linked seem to be using a datapath architecture which is the key to the highest level of performance
None of the systems I mentioned existed at the time the article was published. I think the author would love duckdb which is a very speedy CLI SQL thingy that reads and writes data in all sorts of formats. It fits in great with other Unix CLI stuff.
Many of the projects I mentioned you could see as a response to OP and the 2015 “Scalability, but at what COST?” paper which benchmarked distributed systems to see how many cores they need to beat a single thread. (https://news.ycombinator.com/item?id=26925449)
> None of the systems I mentioned existed at the time the article was published
So Hadoop was doing distributed compute wrong but now they have it figured out?
The point is that there is enormous overhead and complexity in going it in any kind of system. And your computer has a lot of power you probably aren’t maxing out.
> which is a very speedy CLI SQL thingy that reads and writes data in all sorts of formats.
Yeah im a big fan of SQLite :). But at analytical workloads like aggregating every row, DuckDB will outperform SQLite by a wide margin. SQLite is great stuff but it’s not a very good data Swiss Army knife because it’s very focused on a single core competency: embeddable OLTP with a simple codebase. DuckDB can read/write many more formats from local disk or via a variety of network protocols. DuckDB also embeds SQLite so you can use it with SQLite DBs as inputs or outputs.
> they were doing distributed compute wrong but now they have it figured out?
Like anything the future is here but it’s unevenly distributed. Frank McSherry, the first author of “Scalability but at what COST” wrote Timely Dataflow as his answer to that question. ByteWax is based on Timely as is Materialize. Stuff is still complex but these more modern systems with performance as their goal are orders of magnitude better than the Hadoop era Java stuff.
"I've seen startups burning $5k/mo on distributed compute clusters to process <10GB of daily logs, purely because setting up a 'Modern Data Stack' is what gets you promoted, while writing a robust bash script is seen as 'unscalable' or 'hacky'."
Also seen strange responses from HN commenters when it's mentioned that bash is large and slow compared to ash and bash is better suited for use as an interactive shell whereas ash is better suited for use as a non-interactive shell, i.e., a scripting shell
I also use ash (with tabcomplete) as an interactive shell for several reasons
> because setting up a 'Modern Data Stack' is what gets you promoted
It’s not just that, it’s that you better know their specific tech stack to even get hired. It’s a lot of dumb engineering leaders pretending that AWS, Azure and Snowflake are such wildly different ecosystems that not having direct experience in theirs is disqualifying (for pure DE roles, not talking broader sysadmin).
The entire data world is rife with people who don’t have the faintest clue what they’re doing, who really like buzzwords, and who have never thought about their problem space critically.
If airflow is a layer of abstraction something is wrong.
Yes it is an additional layer, but if your orchestration starts concerning itself with what it is doing then something is wrong. It is not a layer on top of other logic, it is a single layer where you define how to start your tasks, how to tell when something is wrong, and when to run them.
If you don't insist on doing heavy compitations within the airflow worker it is dirt cheap. If it's something that can easily be done in bash or python you can do it within the worker as long as you're willing to throw a minimal amount of hardware at it.
It's great to see this post I wrote years ago still being useful for people.
I agree with many here that the situation is arguably worse in many ways. However, along similar lines, I've been pleased to see a move away from cargo culting microservices (another topic I addressed in a separate post on that site).
To all those helping companies and teams improve performance, keep it up! There is hope!
This times a zillion! I think there's been a huge industry push to convince managers and more junior engineers that spark and distributed tools are the correct way to do data engineering.
I think its a similar pattern to web dev influencers have convinced everyone to build huge hydrated-spa-framework-craziness where a static site would do.
My advice to get out of this mess:
- Managers, don't ask for specific solutions (spark, react). Ask for clever engineers to solve problems and optimise / track what you vare about (cost, performance etc). You hired them to know best, and they probably do.
- Technical leads, if your manager is saying "what about hyperscale?" You don't have to say "our existing solution will scale forever". It's fine to say, "our pipelines handle dataset up to 20GB, we don't expect to see anything larger soon, and if we do we'll do x/y/z to meet that scale". Your manager probably just wants to know scaling isn't going to crash everything, not that you've optimised the hell out of everything for your excel spreadsheet processing pipeline.
Here’s the thing though, most companies work with small data. The distribution of data set size follows a power law which means that few engineers get to work with petabyte sized datasets. That said, the job market definitely incentivizes people to have that kind of experience on their resume if they want to keep progressing in salary. This incentivizes over engineering.
Absolutely, when I worked at (semi-well-known unicorn) a half-dozen years ago on the data-engineering team the manager told me "Hey we want to use spark next quarter, that's a huge initiative."
And I immediately asked, "in what capacity?" And the answer was don't-know/doesn't-matter, it's just important that we can say we're using it. I really wish I understood where that was coming from (his manager resume-building? somebody getting a kickback?)
The most interesting part is that you can say you're doing/using something entirely independent of if you actually are. Sure, that's a lie, but so is only using something so you can say you're using it (sure, they admitted to you that was the reason, but that won't be the reason they put on LinkedIn).
A little bit of history related to the article for any who might be interested...
mrjob, the tool mentioned in the article, has a local mode that does not use Hadoop, but just runs on the local computer. That mode is primarily for developing jobs you'll later run on a Hadoop cluster over more data. But, for smaller datasets, that local mode can be significantly faster than running on a cluster with Hadoop. That's especially true for transient AWS EMR clusters — for smaller jobs, local mode often finishes before the cluster is up and ready to start working.
Even so, I bet the author's approach is still significantly faster than mrjob's local mode for that dataset. What MapReduce brought was a constrained computation model that made it easy to scale way up. That has trade-offs that typically aren't worth it if you don't need that scale. Scaling up here refers to data that wouldn't easily fit on disks of the day — the ability to seamlessly stream input/output data from/to S3 was powerful.
I used mrjob a lot in the early 2010s — jobs that I worked on cumulatively processed many petabytes of data. What it enabled you to do, and how easy it was to do it, was pretty amazing when it was first released in 2010. But it hasn't been very relevant for a while now.
When I worked as a data engineer, I rewrote some Bash and Python scripts into C# that were previously processing gigabytes of JSON at 10s of MB/s - creating a huge bottleneck.
By applying some trivial optimizations, like streaming the parsing, I essentially managed to get it to run at almost disk speed (1GB/s on an SSD back then).
Just how much data do you need when these sort of clustered approaches really start to make sense?
> I rewrote some Bash and Python scripts into C# that were previously processing gigabytes of JSON
Hah, incredibly funny, I remember doing the complete opposite about 15 years ago, some beginner developer had setup a whole interconnected system with multiple processes and what not in order to process a bunch of JSON and it took forever. Got replaced with a bash script + Python!
> Just how much data do you need when these sort of clustered approaches really start to make sense?
I dunno exactly what thresholds others use, but I usually say if it'd take longer than a day to process (efficiently), then you probably want to figure out a better way than just running a program on a single machine to do it.
I remember a panel once at a PyCon where we were discussing, I think, the anaconda distribution in the context of packaging and a respected data scientist (whose talks have always been hugely popular) made the point that he doesn't like Pandas because it's not excel. The latter was his go to tool for most of his exploratory work. If the data were too big, he'd sample it and things like that but his work finally was in Excel.
Quick Python/bash to cleanup data is fine too I suppose and with LLMs, it's easier than ever to write the quick throwaway script.
I took a bio statistic class. The tools were Excel/ R or Stata.
I think most people used R. Free and great graphing. Though the interactivity of Excel is great for what ifs. I never got R till I took that class. Though RStudio makes R seem like scriptable excel.
R/Python are fast enough for most things though a lot of genomic stuff (Blast alignments etc..) are in compiled languages.
I like the peer comment's answer about a processing time threshold (e.g., a day). Another obvious threshold is data that doesn't conveniently fit on local disks. Large scale processing solutions can often process directly from/to object stores like S3. And if it's running inside the same provider (e.g., AWS in the case of S3), data can often be streamed much faster than with local SSDs. 10GB/s has been available for a decade or more, and I think 100GB/s is available these days.
> data can often be streamed much faster than with local SSDs. 10GB/s has been available for a decade or more, and I think 100GB/s is available these days.
In practice most AWS instances are 10Gbps capped. I have seen ~5Gbps consistently read from GCS and S3. Nitro based images are in theory 100Gbps capable, in practice I've never seen that.
Also, anything under 16 vCPUs generally has baseline / burst bandwidth, with the burst being best-effort, 5-60 minutes.
This has, at multiple companies for me, been the cause of surprise incidents, where people were unaware of this fact and were then surprised when the bandwidth suddenly plummeted by 50% or more after a sustained load.
How do you stream parse json? I thought you need to ingest it whole to ensure it is syntactically valid, and most parsers don't work with inchoate or invalid json? Or at least it doesn't seem trivial.
I used Newtonsoft.Json which takes in a stream, and while it can give you objects, it can also expose it as a stream of tokens.
The bulk of the data was in big JSON arrays, so you basically consumed the array start token, then used the parser to consume an entire objects which could be turned into a C# object by the deserializer, then you consumed a comma or end array token until you ran out of tokens.
I had to do it like this because DS-es were running into the problem that some of the files didn't fit into memory. The previous approach took 1 hour, involved reading the whole file into memory and parsing it as JSON (when some of the files got over 10GB, even 64GB memory wasnt enough and the system started swapping).
It wasn't fast even before swapping (I learned just how slow Python can be), but then basically it took a day to run a single experiment. Then the data got turned into a dataframe.
I replaced that part of the Python code processing and outputted a CSV which Pandas could read without having to trip through Python code (I guess it has an internal optimized C implementation).
The preprocessor was able to run on the build machines and DSes consumed the CSV directly.
This sounds similar to how in C#/.NET there are (at least) 3 methods to reading XML: XmlDocument, XPathDocument, or XmlReader. The first 2 are in-memory object models that must parse the entire document to build up an object hierarchy, which you then access object-oriented representations of XML constructs like elements and attributes. The XmlReader is stream-based, where you handle tokens in the XML as they are read (forward-only.)
Any large XML document will clobber a program using the in-memory representations, and the solution is to move to XmlReader. System.Text.Json (.NET built-in parsing) has a similar token-based reader in addition to the standard (de)serialization to objects approach.
I'm going to go out on a limb and say no - this library seems to do the parsing in Python, and Python is slow, like many times slower than Java, C# or languages in this class - which you find out if you try to do heavy data processing with it, which is one of the reasons I dislike the language. It's also very hard to parallelize - in C# if you feed stuff into LINQ and entries are independent, you can make the work parallel with PLINQ very quickly, while threads aren't really a thing in Python (or at least they werent back then).
I've seen so many times that data processing quickly became a bottleneck and source of frustration with Python that stuff needed to be rewritten, that I came to not bother writing stuff in Python in the first place.
You can make Python fast by relying on NumPy and pandas with array programming, but doing so can be quite challenging to format and massage the data so that the things that you want can be expressed as array programming ops, that usually it became too much of a burden for me.
I wish Python was at least as fast as Node (which also can have its own share of performance cliffs)
It's possible that nowadays Python has JITs that improve performance to Java levels while keeping compatibility with most existing code - I haven't used Python professionally in quite a few years.
> native code parsing speedups for most common platforms
Which is to say, roughly analogous to "relying on NumPy". (A well-designed system avoids repeatedly calling from Python to C and prefers to let loops live within the C code; that applies at least as much to tree-like data as array-like data.)
> I wish Python was at least as fast as Node (which also can have its own share of performance cliffs) It's possible that nowadays Python has JITs that improve performance to Java levels while keeping compatibility with most existing code - I haven't used Python professionally in quite a few years.
No guarantees, but have you tried PyPy? It's existed since 2007 and definitely improved over time.
I would say that "performance cliffs" are just endemic to programming. Even in C you find people writing bad algorithms because better ones seem (at least superficially) much harder to write — especially if the good algorithm requires, say, a hash table. (C++ standard library containers definitely ameliorate this effect, but you pay in code complexity, especially where templates are needed.) And on the other hand you sometimes see big improvements from dropping to assembly (cf. ffmpeg).
I don't know what the GP was referring too, but often this is about "JSONL" / "JSON Lines" - files containing one JSON object per line. This is common for things like log files. So, process the data as each line is deserialized rather than deserializing the entire file first.
You assume it is valid, until it isn't and you can have different strategies to handle that, like just skipping the broken part and carrying on.
Anyway, you write a state machine that processes the string in chunks – as you would do with a regular parser – but the difference is that the parser is eager to spit out a stream of data that matches the query as soon as you find it.
The objective is to reduce the memory consumption as much as possible, so that your program can handle an unbounded JSON string and only keep track of where in the structure it currently is – like a jQuery selector.
There's a whole heap of approaches, each with their own tradeoffs. But most of them aren't trivial, no. And most end up behaving erratically with invalid json.
You can buffer data, or yield as it becomes available before discarding, or use the visitor pattern, and others.
> Just how much data do you need when these sort of clustered approaches really start to make sense?
You really need an enormous amount of data (or data processing) to justify a clustered setup. Single machines can scale up rather quite a lot.
It'll cost money, but you can order a 24x128GB ram, 24x30TB ssd system which will arrive in a few days and give you 3 TB ram, 720 TB (fast) disk. You can go bigger, but it'll be a little exotic and the ordering process might take longer.
If you need more storage/ram than around that, you need clustering. Or if the processing power you get in your single system storage isn't enough, you would need to cluster, but ~ 256 cores of cpu is enough for a lot of things.
It's not about how much data you have, but also the sorts of things you are running on your data. Joins and group by's scale much faster than any aggregation. Additionally, you have a unified platform where large teams can share code in a structured way for all data processing jobs. It's similar in how companies use k8s as a way to manage the human side of software development in that sense.
I can however say that when I had a job at a major cloud provider optimizing spark core for our customers, one of the key areas where we saw rapid improvement was simply through fewer machines with vertically scaled hardware almost always outperformed any sort of distributed system (abet not always from a price performance perspective).
The real value often comes from the ability to do retries, and leverage left over underutilized hardware (i.e. spot instances, or in your own data center at times when scale is lower), handle hardware failures, ect, all with the ability for the full above suite of tools to work.
Disagree, though in practice it depends on the query, cardinality of the various columns across table, indices, and RDBMS implementation (so, everything).
A simple equijoin with high cardinality and indexed columns will usually be extremely fast. The same join in a 1:M might be fast, or it might result in a massive fanout. In the case of the latter, if your RDBMS uses a clustering index, and if you’ve designed your schemata to exploit this fact (e.g. a table called UserPurchase that has a PK of (user_id, purchase_id)) can still be quite fast.
Aggregations often imply large amounts of data being retrieved, though this is not necessarily true.
An aggregation is less work than a join. You are segmenting the data in basically the same way in ideal conditions for a join as you are in an aggregation. Think of an aggregation as an inner join against a table of buckets (plus updating a single value instead of keeping a number of copies around). In practice this holds with aggregation being a linear amount faster than a join over the same data. That delta is the extra work the join needs to do to keep around a list of rows rather than a single value being updated (and in cache) repeatedly. Depending on the data this delta might be quite small. But without a very obtuse aggregation function (maybe ketosis perhaps), the aggregation will be faster. Its updating a single value vs appending to a list with the extra memory overhead this introduces.
That level of database optimization is rare in practice. As soon as a non-database person gets decision making authority there goes your data model and disk layout.
And many important datasets never make it into any kind of database like that. Very few people provide "index columns" in their CSV files. Or they use long variable length strings as their primary key.
OP pertains to that kind of data. Some stuff in text files.
unconvinced. any join needs some kind of seek on the secondary relation index, or a bunch of state if ur stream joining to build temporary index sizes O(n) until end of batch. on the other hand summing N numbers needs O(1) memory and if your data is column shaped it’s like one CPU instruction to process 8 rows. in “big data” context usually there’s no traditional b-tree index to join either. For jobs that process every row in the input set Mr Join is horrible for perf to the point people end up with a dedicated join job/materialized view so downstream jobs don’t have to re do the work
you didn't need to read to rewrite to C# to do that - python should be able to handle streaming that amount/velocity of data fine, at least through a native extension like msgspec or pydantic. additionally, you made it much harder for other data engineers that need to maintain/extend the project in the future to do so.
The C# is probably far more maintainable and less error prone than Python. At least in my experience that's almost always the case.
The amount of Python jobs I've had which run fine for several hours and then break with runtime errors, whereas with C# you can be reliably sure that if it starts running it will finish running.
Not a language problem, it's a dev culture problem. You can hold your devs accountable to the quality of their code. Strong er typing support via static analysis as well as runtime validation with untrusted input/data has really helped python alot.
I'm not necessarily the biggest fan of python, but writing a data engineering tool in a non-data engineering focused language seems like a bad decision. Now when the OP leaves the organization is in a much tougher position.
The bigness of your data has always depended on the what you are doing with it.
Consider the following table of medical surgeries: date,physician_name, surgery_name,success.
"What are the top 10 most common surgeries?" - easy in bash
"Who are the top physicians (% success) in the last year for those surgeries?" - still easy in bash
"Which surgeries are most affected by physician experience?" - very hard in bash, requires calculating for every surgery how many times that physician had performed that surgery on that day, then compare low and high experience outcomes.
A researcher might see a smooth continuum of increasingly complex questions, but there are huge jumps in computational complexity. At 50gb dataset might be 'bigger' than a 2tb one if you are asking tough questions.
It's easier for a business to say "we use Spark for data processing", than "we build bespoke processing engines on a case by case basis".
50GB and 2TB are both sizes that SQLite supports and could handle. You could probably solve all of the problems you mentioned with simple tools on a single server, in the language of your choice.
I think many devs learn the trade with Windows and don't get exposure to these tools.
Plus, they require a bit of reading because they operate on a higher level of abstraction than loops and ifs. You get implicit loops, your fields get cut up automatically, and you can apply regexes simultaneously on all fields. So it's not obvious to the untrained eye.
But you get a lot of power and flexibility on the cli, which enable you to rapidly put together an ad hoc solution which can get the job done or at least serve as a baseline before you reach for the big guns.
> The first thing to do is get a lot of game data. This proved more difficult than I thought it would be, but after some looking around online I found a git repository on GitHub from rozim that had plenty of games. I used this to compile a set of 3.46GB of data, which is about twice what Tom used in his test. The next step is to get all that data into our pipeline.
It would be interesting to redo the benchmark but with a (much) larger database.
Nowadays the biggest open-data for chess must comes from Lichess https://database.lichess.org, with ~7B games and 2.34 TB compressed, ~14TB uncompressed.
Almost certainly not. You can go on AWS or GCP and spin up a VM with 2.2 TB RAM and 288 vCPUs. Worst case, if streaming the data sequentially isn't fast enough, you can use something like GNU Parallel to launch processes in parallel to use all those 288 cpus. (It's also extremely easy to set up - 'apt install parallel' is about all you need.) That starts to resemble Hadoop, if you squint, except that it's all running on the same machine. As a result, it's going to outperform Hadoop significantly.
The only reason not to do that is if for some reason the workload won't support that kind of out-of-the-box parallelism. But in that case, you'd be writing custom code for Hadoop or Spark anyway, so there's an argument for doing the same to run on a single VM. These days it's pretty easy to essentially vibe code a custom script to do what you need.
At the company I'm with, we use Spark and Apache Beam for many of our large workloads, but that's typically involving data at the petabyte scale. If you're just dealing with a few dozen terabytes, it's often faster and easier to spin up a single large VM. I just ran a process on Friday like that, on a 96-core VM with 350 GB RAM.
If you get all the data on fast SSDs in a single chassis, you probably still beat EMR over S3. But then you have a whole dedicated server to manage your 14TB of chess games.
The "EMR over S3" paradigm is based on the assumption that the data isn't read all that frequently, 1-10x a day typically, so you want your cheap S3 storage but once in a while you'll want to crank up the parallelism to run a big report over longer time periods.
It’s easy to overlook how often straightforward approaches are the best fit when the data and problem are well understood. Large expensive tools can become problems in their own right creating complexity that then requires even more tooling to manage. (Maybe that's the intent?) The issue is that teams and companies often adopt optimization frameworks earlier than necessary. Starting with simpler tools can get you most of the way there and in many cases they turn out to be all that’s needed.
I've contributed to PrestoDB, but the availability of DuckDB and fast multi core machines with even faster SSDs makes the need for distribution all the more niche, or even cargo-culting Google or Meta.
The same thing is true with Sqlite vs Postgres. Most startups need Sqlite, not Postgres. Many queries run an order of magnitude faster. Not only is it better for your users, it's life changing to see the test suites (which would take minutes to run) complete in mere seconds
Feels like quibbling over the differences between two databases that are going to act the same for 90% of projects out there doesn't really matter.
If you want speed, just have your database stored in the same place as your application, locally, rather than hopping across the world to retrieve data that can be located next to the code.
That would probably be the easiest thing to do to get a real measured performance gains.
As other commentators pointed out, computers are extremely powerful. This isn't 1995, you can easily host everything in the same local area and get a very responsive application with very minimal needs to worry about resource constraints.
Given how primitive SQLite's optimizer is and how similar the storage and execution engines between the two are in terms of architecture, this seems unlikely to be the norm unless you did something wrong on the Postgres side. (Of course, no RDBMS optimizer will always give the best answer, so there's bound to be such cases.)
I’m curious about the memory usage of the cat | grep part of the pipeline. I think the author is processing many small files?
In which case it makes the analysis a bit less practical, since the main use case I have for fancy data processing tools is when I can’t load a whole big file into memory.
Unix shell pipelines are task-parallel. Every tool gets spun up as its own unix process — think "program" (fork-exec). Standard input and standard output (stdin, stdout) get hooked up to pipes. Pipes are like temporary files managed by the kernel (hand-wave). Pipe buffer size is a few KB. Grep does a blocking read on stdin. Cat writes to stdout. Both on a kernel I/O boundary. Here the kernel can context-switch the process when waiting for I/O.
In the past there was time-slicing. Now with multiple cores and hardware threads they actually run concurrently.
This is very similar to old-school approach to something like multiple threads, but processes don’t share virtual address spaces in the CPU's memory management unit (MMU).
Further details: look up McIlroy's pipeline design.
The comments here smell of "real engineers use command line". But I am not sure they ever actually worked with analysing data more than using it as a log parser.
Yes Hadoop is 2014.
These days you obviously don't set up a Hadoop cluster. You use the cloud provider service provided (BigQuery or AWS Athena for example).
Or map your data into DuckDB or use polars if it is small.
It depends. I’ve done plenty of data processing, including at large fortune 10s. Most of the big data could be shrunk to small data if you understood the use case— pre-aggregating, filtering to smaller datasets based on known analysis patterns, etc.
Now, you could argue that that’s cheating a bit and introduces preprocessing that is as complex as running Hadoop in the first place, but I think it depends.
In my experience, though, most companies really don’t have big data, and many that do don’t really need to.
Most companies aren’t fortune 500s.
I used to work at Elastic, and I noticed that most (not all!) of the customers who walked up to me at the conferences were there to ask about datasets that easily fit into memory on a cheap VPS.
> But I am not sure they ever actually worked with analysing data more than using it as a log parser.
It really feels that way. Real data analysis involves a lot more than just grepping logs. And the reason to be wary of starting out unprepared for that kind of analysis is that migrating to a better solution later is a nightmare.
In many ways HN is Reddit in denial at this point :) Comments and upvotes that are based mostly on vibes, with depth and discussion usually happening somewhere towards the middle of the comment tree.
MapReduce is from a world with slow HDDs, expensive ram, expensive enterprise class servers, fast network.
In that case to get best performance, you’d have to shard your data across a cluster and use mapreduce.
Even in the authors 2014 SSDs multi-core consumer PC world, their aggregate pipeline would be around 2x faster if the work was split across two equivalent machines.
The limit of how much faster distributed computing is comes down to latency more than throughput. I’d not be surprised if this aggregate query could run in 10ms on pre sharded data in a distributed cluster.
Somebody has to go back to first principles. I wrote pig scripts in 2014 in Palo Alto. Yes, it was shit. IYKYK. But the author, and near everybody in this thread, are wrong to generalize.
PCIe would have to be millions of times faster than Ethernet before command line tools are actually faster than distributed computing and I don't see that happening any time soon.
Well, at my old company we had some datasets in the 6-8 PB range, so tell me how we would run analytics on that dataset on an Intel NUC.
Just because you don't have experience of these situations, it doesn't mean they don't exist. There's a reason Hadoop and Spark became synonymous with "big data."
Well yeah, but that's a _very_ different engineering decision with different constraints, it's not fully apples to apples.
Having materialised views increases insert load for every view, so if you want to slice your data in a way that wasn't predicted, or that would have increased ingress load beyond what you've got to spare, say, find all devices with a specific model and year+month because there's a dodgy lot, you'll really wish you were on a DB that can actually run that query instead of only being able to return your _precalculated_ results.
Not only is this a contrived non-comparison, but the statement itself is readily disproven by the limitations basically _everyone_ using single instance ClickHouse often run into if they actually have a large dataset.
Spark and Hadoop have their place, maybe not in rinky dink startup land, but definitely in the world of petabyte and exabyte data processing.
And now with things like DuckDB and clickhouse-local you won't have to worry about data processing performance ever again. Just kidding, but especially with ClickHouse it's so much better to handle the large data volume compared to the past, and even a single beefy server is often enough to satisfy all data analytics needs for a moderate-to-large company.
Something to note here is that the result of xargs -P is unlikely to be satisfactory, since all of the subprocesses are simply connected to the terminal and stomp over each other's outputs. A better choice would be something like rush or, for the Perl fans, parallel.
It's worth remembering however that even though it's less efficient per-CPU or whatever to split a large task into many smaller tasks, it may be more efficient overall alongside other workloads as you can bin-pack tasks more efficiently on a cluster, not to mention if tasks fail you are retrying less of the overall work.
All this is to say, the article makes a very good point, but doing it all on one machine also has problems. Just don't cargo cult engineering decisions.
We now have even more layers of abstraction (Airflow, dbt, Snowflake) applied to datasets that often fit entirely in RAM.
I've seen startups burning $5k/mo on distributed compute clusters to process <10GB of daily logs, purely because setting up a 'Modern Data Stack' is what gets you promoted, while writing a robust bash script is seen as 'unscalable' or 'hacky'. The incentives are misaligned with efficiency.
I think a lot of people don't realize machines come with TBs of RAM and hundreds of physical cores. One machine is fucking huge these days.
The problem we have is fucked up piles of shit not that we don’t have kubernetes and don’t have containers.
So I totally understand why people preemptively choose Kubernetes before they are scaling to the point where having a distributed scheduler is strictly necessary. Hadoop, on the other hand, you're definitely paying a large upfront cost for scalability you very much might not need.
It only makes sense if you’re managing large amounts of large siloed bits of kit. I’ve not seen this other than at unnamed big tech companies.
99.9% of people are just burning money for a fashion show where everyone is wearing clown suits because someone said clown suits are good.
The project is a Ruby on Rails app that talks to PostreSQL and a handful of third party services. It just seems unnecessary to include the complexity of containers.
I just don't agree. I don't find Docker too complicated to get started with at all. A lot of my projects have very simple Dockerfiles. For example, here is a Dockerfile I have for a project that has a Node.JS frontend and a Go backend:
It is a glorified shell script that produces an OCI image with just a single binary. There's a bit of boilerplate but it's nothing out of the ordinary in my opinion. It gives you something you can push to an OCI registry and deploy basically anywhere that can run Docker or Podman, whether it's a Kubernetes cluster in GCP, a bare metal machine with systemd and podman, a NAS running Synology DSM or TrueNAS or similar, or even a Raspberry Pi if you build for aarch64. All of the configuration can be passed via environment variables or if you want, additional command line arguments, since starting a container very much is just like starting a process (because it is.)But of course, for development you want to be able to iterate rapidly, and don't want to be dealing with a bunch of Docker build BS for that. I agree with this. However, the utility of Docker doesn't really stop at building for production either. Thanks to the utility of OCI images, it's also pretty good for setting up dev environment boilerplate. Here's a docker-compose file for the same project:
And if your application is built from the ground up to handle these environments well, which doesn't take a whole lot (basically, just needs to be able to handle configuration from the environment, and to make things a little neater it can have defaults that work well for development), this provides a one-command, auto-reloading development environment whose only dependency is having Docker or Podman installed. `docker compose up` gives you a full local development environment.I'm omitting a bit of more advanced topics but these are lightly modified real Docker manifests mainly just reformatted to fewer lines for HN.
I adopted Kubernetes pretty early on. I felt like it was a much better abstraction to use for scheduling compute resources than cloud VMs, and it was how I introduced infrastructure-as-code to one of the first places I ever worked.
I'm less than thrilled about how complex Kubernetes can be, once you start digging into stuff like Helm and ArgoCD and even more, but in general it's an incredible asset that can take a lot of grunt work out of deployment while providing quite a bit of utility on top.
I can maybe make a case for running in containers if you need some specific security properties but .. mostly I think the proliferation of 'fucked up piles of shit' is the problem.
Different processes can need different environments.
I advocate for something lightweight like FreeBSD jails.
Naturally, that detaches all your containers. And theres no seamless reattach for control plane restart.
(Large EKS cluster)
It is not about what you are doing, it is always about how you do it.
This was the same with doing OCR analysis of assembly and production manuals. Quick and dirty, it would of took over 24 hours of processing time, after moving to semaphores with parallelization it took less than two hours to process all the information.
It saddens me to see how the LinkedIn slop style is expanding to other platforms
Unless, of course, you have multiple options and you don’t want to work for a company that’s looking for dumb stuff in interviews.
I optimize my answers for the companies I want to work for, and get rejected by the ones I don't. The hardest part of that strategy is coming to terms with the idea that I constantly get rejected by people that I think are mostly <derogatory_words_here>, but I've developed thick skin over the years.
I'd much rather spend a year unemployed (and do a ton of painful interviews) and find a company who's values align with mine, than work for a year on a team I disagree with constantly and quit out of frustration.
I also believe that running a broken interview process actively selects for qualities you actually don't want, so it's much more likely that teams conducting those interviews aren't teams I want to work on.
Edit: As credence for my claims, the best team I've ever worked on was a team I did 90%+ of the hiring for, and we didn't do any of the 'typical' interview bullshit most companies do.
What we did instead was sit people down and have deep technical conversations about systems they'd worked on in the past. The candidate would explain, in as much detail as they could muster, a system they'd worked on in the past, down to the lowest level details. Usually, they would talk to us for at least 20-30 minutes, then, we (the interviewers) would pose questions, usually starting with the form 'if we changed X, what effect would it have'. Doing interviews in this style make a few things immediately obvious:
1. Did the candidate have a deep, systemic understanding of the system they worked on?
2. Does the candidate have a good mental model for evaluating change in the system?
That's how I conduct interviews, and unsurprisingly, when I get interviewed like that, my success rate is 100%. I don't think I've ever done an interview like that which did not result in an offer.
Anyways, there's some rambling and unsolicited opinions for you :)
Demonstrating competency is always good.
You could have learned this if you were better about collecting requirements. You can tell the interviewer "I'd do it like this for this size data, but I'd do it like this for 100x data. Which size should I design this for?" If they're looking for one direction and you ask which one, interviewers will tell you.
Said another way, how do you have a meaningful conversation about scaling with a person who thinks their application is huge, but in reality only requires a tiny fraction of a single machine? Sometimes, there's such a massive gulf between perception and reality that the only thing to do is chuckle and move on.
Edit: I did try JSON.parse() first, which I expected to fail and it did fail BUT it's important that you try anyway.
Probably a better outcome than being hired onto a team where everyone know you're technically correct but they ignore your suggestions for some mysterious (to you) reason.
Fuck off ..you're 10 person startup with an MVP and no revenue stream needs customers first..
It may or may not be related that the places that this happened were always very ethnically monotone with narrow age ranges (nothing against any particular ethnic group, they were all different ethnic monotones)
let’s see how they think and turn this into a paid interview
Moore's law was supposed to make it simpler and cheaper to do more computationally expensive tasks. But in the meantime, everyone kept inflating the difficulty of a task faster than Moore could keep up.
I think some of this is because of the incredible amounts of capital that startups seem to be able to acquire. If startups had to demonstrate profitability before they were given any money to scale, the story would be very different I think.
You're looking for your first DevOps person, so you want someone who has experience doing DevOps. They'll tell you about all the fancy frameworks and tooling they've used to do Serious Business™, and you'll be impressed and hire them. They'll then proceed to do exactly that for your company, and you'll feel good because you feel it sets you up for the future.
Nobody's against it. So you end up in that situation, which even a basic home desktop would be more than capable of handling.
Cost is usually not a huge problem beyond seed stage. Series A-B the biggest problem is growing the customer base so the fixed infra costs become a rounding error. We've built the product and we're usually focused on customer enablement and technical wins - proving that the product works 100% of the time to large enterprises so we can close deals. We can't afford weird flakiness in the middle of a POC.
Another factor I rarely see discussed is bus factor. I've been in the industry for over a decade, and I like to be able to go on vacation. It's nice to hand off the pager sometimes. Using established technologies makes it possible to delegate responsibility to the rest of the team, instead of me owning a little rats nest fiefdom of my own design.
The fact is that if 5k/month infra cost for a core part of the service sinks your VC backed startup, you've got bigger problems. Investors gave you a big pile of money to go and get customers _now_. An extra month of runway isn't going to save you.
I once interviewed with a company that did some machine learning stuff, this was a while back when that typically meant "1 layer of weights from a regression we run overnight every night". The company asked how I had solved the complex problem of getting the weights to inference servers. I said we had a 30 line shell script that ssh'd them over and then mv'd them into place. Meanwhile the application reopened the file every so often. Zero problems with it ever. They thought I was a caveman.
I have recently started using terraform/tofu and ansible to automate nearly all of the devops operations. We are at a point where Claude Code can use these tools and our existing configs to make configuration changes, debug issues by reviewing logs etc. It is much faster at debugging an issue than I am and I know our stuff inside and out.
I am beginning to think that AI will soon force people to rethink their cloud hosting strategy.
Basically discoverability is where shell script fail
Or python. The python3 standard library is pretty capable, and it's ubiquitous. You can do a lot in 50-100 lines (counting documentation) with no dependencies. In turn it's easy to plug into the other stuff.
No, it's lack of documentation and no amount of $$$$/m enterprise AI solutions (R)(TM) would help you if there is no documentation.
And then I got laid off. Now, I've got very few modern frameworks on my resume and I've been jobless for over a year.
I'm feeling a right fool now.
There is something wrong with the industry in chasing fads and group think. It has always been this way. Businesses chased Java in the late 90s, early 00s. They chased CORBA, WSDL, ESB, ERP and a host of other acronyms back in the day.
More recently, Data Lake, Big Data, Cloud Compute, AI.
Most of the executives I have met really have no clue. They just go with what is being promoted in the space because it offers a safety net. Look, we are "not behind the curve!". We are innovating along with the rest of the industry.
Interviews do not really test much for ability to think and reason. If you ran an entire ISP, if you figured out, on your own, without any help, how to shard databases, put in multiple layers of redundancy, caching... well, nobody cares now. You had to do it in AWS or Azure or whatever stack they have currently.
Sadly, I do not think it will ever be fixed. It is something intrinsic to human nature.
Need training and something to show? Contribute to some FOSS project.
If you're willing and able to promote yourself internally, you can make people give a shit, or at least publicly claim they do. That's 260k+ per year, and even big businesses are going to care about that at some level, especially if it's something that can be replicated. Find 10 systems you can do that with, and it's 2.6m+ per year.
But, if you don't want to play the self-promotion game, yeah someone else is going to benefit from your work.
Yep, and a lot more datasets fit entirely into RAM now. Ignoring the recent price spikes for a moment, 128GB of RAM in a laptop is entirely achievable and not even the limit of what is possible. That was a pipe dream in 2014 when computers with only 4GB were still common. And of course for servers the max RAM is much higher, and in a lot of scenarios streaming data off a fast local SSD may be almost as good.
https://yourdatafitsinram.net/
This kind of practice is insidious because early on, they charge $20/month to get started on the first 100mb of log ingestion, and you can have it up and running in 30 seconds with a credit card. Who would turn that down?
Revisit that set up 2 years later and it’s turned into a 60k/y behemoth that no one can unwind
It's just like the systemd people talking about sysvinit. "Eww, shell scripts! What a terrible hack!" says the guy with no clue and no skills.
It's like the whole ship is being steered by noobs.
That's funny. I used to have to clean up the messes caused by systemd's design limitations and flaws, until I built my own distro with a sane init system installed.
Many of the noobs groaning about the indignity of shell scripts don't even realize that they could write init 'scripts' in whatever language they want, including Python (the language these types usually love so much, if they do any programming at all.)
For example, I’ve been dealing with SysV since the early 90s and while it’s gotten better since we no longer have to support the really bizarre Unix variants, my problem with init scripts wasn’t “indignity” but the lack of consistency across distributions and versions, which affects anyone shipping software professionally (“can’t do this easily until $distro upgrades coreutils”), and from an operator’s perspective using Python doesn’t make that better because instead of supporting one consistent thing you’d end up with the subset of features each application team felt like implementing, consistent only to the extent that they care to follow other projects. One virtue of systemd is that having a single common way to specify dependencies, restarts, customization, etc. avoids the ops people having to learn dozens of different variations of the same ideas and especially how to deal with their gaps. A few years back, a data center power outage at one place I worked really highlighted that: the systemd-based servers recovered quickly because they actually had working retries; all of the older stuff using SysV had to be manually reviewed because there were all kinds of problems like races on dependencies like DNS or NFS, retry logic which failed hard after a short period of time, failures because a stale PID file wasn’t removed, or cases where a vendor had simply never implemented retries in their init scripts. While in theory you can handle all of those in SysV most people never did.
After a couple decades of that, a lot of us don’t want to spend time on problems Microsoft solved in Bill Clinton’s first term.
Nothing insurmountable but it meant init files were inevitably much longer than the corresponding Upstart or systemd files despite doing less, and anytime we shipped a new version you had more testing since you had to implement a lot of functionality which is built in to other things.
It's the same thing any corporation should be doing if they were smart, instead of outsourcing everything to RedHat, Microsoft, Google, etc.
Systemd unified and simplified administration across a lot of distributions. Before, it was a hodge podge, and there was a lot of knowledge lost going from rhel to Debian.
I honestly do not like systemd, either. It is okay for managing processes but I wish it didn't spread into everything else in the machine.
Or if it must, could it actually work cohesively across their concepts? Would be nice to have an obvious and easy way to run Quadlet as its own user to isolate further, would be nice to have systemd-sysusers present in /etc/subuid so they can run containers.
I like what they are doing with atomic distros. It would be great to have a single file declarative setup for something like running a containerized reverse HTTP proxy with an isolated user. Instead of "atomic" but you manually edit files in /etc after install.
I've been using this pattern (scripts or code that execute commands against DuckDB) to process data more recently, and the ability to do deep investigations on the data as you're designing the pipeline (or when things go wrong) is very useful. Doing it with a code-based solution (read data into objects in memory) is much more challenging to view the data. Debugging tools to inspect the objects on the heap is painful compared to being able to JOIN/WHERE/GROUP BY your data.
The bottleneck in the example was maxing out disk IO, which I don't think duckdb can help with.
On the other hand, unix sockets combined with socat can perform some real wizardry, but I never quite got the hang of that style.
If the tool of interest works with files (like the UNIX tools do) it fits very well.
If the tool doesn't work with single files I have had some success in using Makefiles for generic processing tasks by creating a marker file that a given task was complete as part of the target.
It’s the same story as always, just it used to be Oracle certified tech, now it’s the AWS tech certified to ensure you pay Amazon.
It looked good on someone’s resume and that was it. They are long gone.
The issue is you can run sub tib jobs on a few small/standard instances with better tooling. Spark and Hadoop are for when you need multiple machines.
Dbt and airflow let you represent your data as a DAG and operate on that, which is critical if you want to actually maintain and correct data issues and keep your data transforms timely.
edit: a little surprised at multiple downvotes. My point is, you can run airflow and dbt on small instances, and you can do all your data processing on small instances with tools like duckdb or polars.
But it is very useful to use a tool like dbt that allows you to re-build and manage your data in a clear way, or a tool like airflow which lets you specify dependencies for runs.
After say 30 jobs or so, you'll find that being able to re-run all downstreams of a model starts to payoff.
These hardly exist in practice.
But I get what you mean.
PS None of the companies you linked seem to be using a datapath architecture which is the key to the highest level of performance
I guarantee those rust projects have spent more time playing with rust and library design than the domain problem they are trying to solve.
Many of the projects I mentioned you could see as a response to OP and the 2015 “Scalability, but at what COST?” paper which benchmarked distributed systems to see how many cores they need to beat a single thread. (https://news.ycombinator.com/item?id=26925449)
So Hadoop was doing distributed compute wrong but now they have it figured out?
The point is that there is enormous overhead and complexity in going it in any kind of system. And your computer has a lot of power you probably aren’t maxing out.
> which is a very speedy CLI SQL thingy that reads and writes data in all sorts of formats.
Do you know about SQLite?
> they were doing distributed compute wrong but now they have it figured out?
Like anything the future is here but it’s unevenly distributed. Frank McSherry, the first author of “Scalability but at what COST” wrote Timely Dataflow as his answer to that question. ByteWax is based on Timely as is Materialize. Stuff is still complex but these more modern systems with performance as their goal are orders of magnitude better than the Hadoop era Java stuff.
Also seen strange responses from HN commenters when it's mentioned that bash is large and slow compared to ash and bash is better suited for use as an interactive shell whereas ash is better suited for use as a non-interactive shell, i.e., a scripting shell
I also use ash (with tabcomplete) as an interactive shell for several reasons
datalake (DuckLake), pipelines (hubspot, stripe, postgres), and dashboards in a single app for $250/mo.
marketing/finance get dashboards, everyone else gets SQL + AI access. one abstraction instead of five, for a fraction of your Snowflake bill.
It’s not just that, it’s that you better know their specific tech stack to even get hired. It’s a lot of dumb engineering leaders pretending that AWS, Azure and Snowflake are such wildly different ecosystems that not having direct experience in theirs is disqualifying (for pure DE roles, not talking broader sysadmin).
The entire data world is rife with people who don’t have the faintest clue what they’re doing, who really like buzzwords, and who have never thought about their problem space critically.
Yes it is an additional layer, but if your orchestration starts concerning itself with what it is doing then something is wrong. It is not a layer on top of other logic, it is a single layer where you define how to start your tasks, how to tell when something is wrong, and when to run them.
If you don't insist on doing heavy compitations within the airflow worker it is dirt cheap. If it's something that can easily be done in bash or python you can do it within the worker as long as you're willing to throw a minimal amount of hardware at it.
It's great to see this post I wrote years ago still being useful for people.
I agree with many here that the situation is arguably worse in many ways. However, along similar lines, I've been pleased to see a move away from cargo culting microservices (another topic I addressed in a separate post on that site).
To all those helping companies and teams improve performance, keep it up! There is hope!
Thank you very much!
Been re-reading your post multiple times.
You inspired me to port Waters-Series (kind-of streams) to JavaScript to get pipelining for stream processing.
I think its a similar pattern to web dev influencers have convinced everyone to build huge hydrated-spa-framework-craziness where a static site would do.
My advice to get out of this mess:
- Managers, don't ask for specific solutions (spark, react). Ask for clever engineers to solve problems and optimise / track what you vare about (cost, performance etc). You hired them to know best, and they probably do.
- Technical leads, if your manager is saying "what about hyperscale?" You don't have to say "our existing solution will scale forever". It's fine to say, "our pipelines handle dataset up to 20GB, we don't expect to see anything larger soon, and if we do we'll do x/y/z to meet that scale". Your manager probably just wants to know scaling isn't going to crash everything, not that you've optimised the hell out of everything for your excel spreadsheet processing pipeline.
And I immediately asked, "in what capacity?" And the answer was don't-know/doesn't-matter, it's just important that we can say we're using it. I really wish I understood where that was coming from (his manager resume-building? somebody getting a kickback?)
mrjob, the tool mentioned in the article, has a local mode that does not use Hadoop, but just runs on the local computer. That mode is primarily for developing jobs you'll later run on a Hadoop cluster over more data. But, for smaller datasets, that local mode can be significantly faster than running on a cluster with Hadoop. That's especially true for transient AWS EMR clusters — for smaller jobs, local mode often finishes before the cluster is up and ready to start working.
Even so, I bet the author's approach is still significantly faster than mrjob's local mode for that dataset. What MapReduce brought was a constrained computation model that made it easy to scale way up. That has trade-offs that typically aren't worth it if you don't need that scale. Scaling up here refers to data that wouldn't easily fit on disks of the day — the ability to seamlessly stream input/output data from/to S3 was powerful.
I used mrjob a lot in the early 2010s — jobs that I worked on cumulatively processed many petabytes of data. What it enabled you to do, and how easy it was to do it, was pretty amazing when it was first released in 2010. But it hasn't been very relevant for a while now.
By applying some trivial optimizations, like streaming the parsing, I essentially managed to get it to run at almost disk speed (1GB/s on an SSD back then).
Just how much data do you need when these sort of clustered approaches really start to make sense?
Hah, incredibly funny, I remember doing the complete opposite about 15 years ago, some beginner developer had setup a whole interconnected system with multiple processes and what not in order to process a bunch of JSON and it took forever. Got replaced with a bash script + Python!
> Just how much data do you need when these sort of clustered approaches really start to make sense?
I dunno exactly what thresholds others use, but I usually say if it'd take longer than a day to process (efficiently), then you probably want to figure out a better way than just running a program on a single machine to do it.
Quick Python/bash to cleanup data is fine too I suppose and with LLMs, it's easier than ever to write the quick throwaway script.
I think most people used R. Free and great graphing. Though the interactivity of Excel is great for what ifs. I never got R till I took that class. Though RStudio makes R seem like scriptable excel.
R/Python are fast enough for most things though a lot of genomic stuff (Blast alignments etc..) are in compiled languages.
In practice most AWS instances are 10Gbps capped. I have seen ~5Gbps consistently read from GCS and S3. Nitro based images are in theory 100Gbps capable, in practice I've never seen that.
This has, at multiple companies for me, been the cause of surprise incidents, where people were unaware of this fact and were then surprised when the bandwidth suddenly plummeted by 50% or more after a sustained load.
I did not see your comment earlier, but to stay with Chess see https://news.ycombinator.com/item?id=46667287, with ~14Tb uncompressed.
It's not humongous and it can certainly fit on disk(s), but not on a typical laptop.
The bulk of the data was in big JSON arrays, so you basically consumed the array start token, then used the parser to consume an entire objects which could be turned into a C# object by the deserializer, then you consumed a comma or end array token until you ran out of tokens.
I had to do it like this because DS-es were running into the problem that some of the files didn't fit into memory. The previous approach took 1 hour, involved reading the whole file into memory and parsing it as JSON (when some of the files got over 10GB, even 64GB memory wasnt enough and the system started swapping).
It wasn't fast even before swapping (I learned just how slow Python can be), but then basically it took a day to run a single experiment. Then the data got turned into a dataframe.
I replaced that part of the Python code processing and outputted a CSV which Pandas could read without having to trip through Python code (I guess it has an internal optimized C implementation).
The preprocessor was able to run on the build machines and DSes consumed the CSV directly.
Any large XML document will clobber a program using the in-memory representations, and the solution is to move to XmlReader. System.Text.Json (.NET built-in parsing) has a similar token-based reader in addition to the standard (de)serialization to objects approach.
I've seen so many times that data processing quickly became a bottleneck and source of frustration with Python that stuff needed to be rewritten, that I came to not bother writing stuff in Python in the first place.
You can make Python fast by relying on NumPy and pandas with array programming, but doing so can be quite challenging to format and massage the data so that the things that you want can be expressed as array programming ops, that usually it became too much of a burden for me.
I wish Python was at least as fast as Node (which also can have its own share of performance cliffs)
It's possible that nowadays Python has JITs that improve performance to Java levels while keeping compatibility with most existing code - I haven't used Python professionally in quite a few years.
> native code parsing speedups for most common platforms
Which is to say, roughly analogous to "relying on NumPy". (A well-designed system avoids repeatedly calling from Python to C and prefers to let loops live within the C code; that applies at least as much to tree-like data as array-like data.)
> I wish Python was at least as fast as Node (which also can have its own share of performance cliffs) It's possible that nowadays Python has JITs that improve performance to Java levels while keeping compatibility with most existing code - I haven't used Python professionally in quite a few years.
No guarantees, but have you tried PyPy? It's existed since 2007 and definitely improved over time.
I would say that "performance cliffs" are just endemic to programming. Even in C you find people writing bad algorithms because better ones seem (at least superficially) much harder to write — especially if the good algorithm requires, say, a hash table. (C++ standard library containers definitely ameliorate this effect, but you pay in code complexity, especially where templates are needed.) And on the other hand you sometimes see big improvements from dropping to assembly (cf. ffmpeg).
Anyway, you write a state machine that processes the string in chunks – as you would do with a regular parser – but the difference is that the parser is eager to spit out a stream of data that matches the query as soon as you find it.
The objective is to reduce the memory consumption as much as possible, so that your program can handle an unbounded JSON string and only keep track of where in the structure it currently is – like a jQuery selector.
You can buffer data, or yield as it becomes available before discarding, or use the visitor pattern, and others.
One Python library that handles pretty much all of them, as a place to start learning, would be: https://github.com/daggaz/json-stream
https://learn.microsoft.com/en-us/dotnet/standard/serializat...
You really need an enormous amount of data (or data processing) to justify a clustered setup. Single machines can scale up rather quite a lot.
It'll cost money, but you can order a 24x128GB ram, 24x30TB ssd system which will arrive in a few days and give you 3 TB ram, 720 TB (fast) disk. You can go bigger, but it'll be a little exotic and the ordering process might take longer.
If you need more storage/ram than around that, you need clustering. Or if the processing power you get in your single system storage isn't enough, you would need to cluster, but ~ 256 cores of cpu is enough for a lot of things.
I have no experience with these, but lots of good experiences with last decade supermicro systems.
I can however say that when I had a job at a major cloud provider optimizing spark core for our customers, one of the key areas where we saw rapid improvement was simply through fewer machines with vertically scaled hardware almost always outperformed any sort of distributed system (abet not always from a price performance perspective).
The real value often comes from the ability to do retries, and leverage left over underutilized hardware (i.e. spot instances, or in your own data center at times when scale is lower), handle hardware failures, ect, all with the ability for the full above suite of tools to work.
A simple equijoin with high cardinality and indexed columns will usually be extremely fast. The same join in a 1:M might be fast, or it might result in a massive fanout. In the case of the latter, if your RDBMS uses a clustering index, and if you’ve designed your schemata to exploit this fact (e.g. a table called UserPurchase that has a PK of (user_id, purchase_id)) can still be quite fast.
Aggregations often imply large amounts of data being retrieved, though this is not necessarily true.
And many important datasets never make it into any kind of database like that. Very few people provide "index columns" in their CSV files. Or they use long variable length strings as their primary key.
OP pertains to that kind of data. Some stuff in text files.
The amount of Python jobs I've had which run fine for several hours and then break with runtime errors, whereas with C# you can be reliably sure that if it starts running it will finish running.
I'm not necessarily the biggest fan of python, but writing a data engineering tool in a non-data engineering focused language seems like a bad decision. Now when the OP leaves the organization is in a much tougher position.
(2018, 222 comments) https://news.ycombinator.com/item?id=17135841
(2022, 166 comments) https://news.ycombinator.com/item?id=30595026
(2024, 139 comments) https://news.ycombinator.com/item?id=39136472 - by the same submitter as this post.
Consider the following table of medical surgeries: date,physician_name, surgery_name,success.
"What are the top 10 most common surgeries?" - easy in bash
"Who are the top physicians (% success) in the last year for those surgeries?" - still easy in bash
"Which surgeries are most affected by physician experience?" - very hard in bash, requires calculating for every surgery how many times that physician had performed that surgery on that day, then compare low and high experience outcomes.
A researcher might see a smooth continuum of increasingly complex questions, but there are huge jumps in computational complexity. At 50gb dataset might be 'bigger' than a 2tb one if you are asking tough questions.
It's easier for a business to say "we use Spark for data processing", than "we build bespoke processing engines on a case by case basis".
1K rows: use excel
1M rows: use pandas/polars
1B rows: use shakti
1T rows: only shakti
Source: https://web.archive.org/web/20230331180931/https://shakti.co...
Plus, they require a bit of reading because they operate on a higher level of abstraction than loops and ifs. You get implicit loops, your fields get cut up automatically, and you can apply regexes simultaneously on all fields. So it's not obvious to the untrained eye.
But you get a lot of power and flexibility on the cli, which enable you to rapidly put together an ad hoc solution which can get the job done or at least serve as a baseline before you reach for the big guns.
It would be interesting to redo the benchmark but with a (much) larger database.
Nowadays the biggest open-data for chess must comes from Lichess https://database.lichess.org, with ~7B games and 2.34 TB compressed, ~14TB uncompressed.
Would Hadoop win here?
The only reason not to do that is if for some reason the workload won't support that kind of out-of-the-box parallelism. But in that case, you'd be writing custom code for Hadoop or Spark anyway, so there's an argument for doing the same to run on a single VM. These days it's pretty easy to essentially vibe code a custom script to do what you need.
At the company I'm with, we use Spark and Apache Beam for many of our large workloads, but that's typically involving data at the petabyte scale. If you're just dealing with a few dozen terabytes, it's often faster and easier to spin up a single large VM. I just ran a process on Friday like that, on a 96-core VM with 350 GB RAM.
The "EMR over S3" paradigm is based on the assumption that the data isn't read all that frequently, 1-10x a day typically, so you want your cheap S3 storage but once in a while you'll want to crank up the parallelism to run a big report over longer time periods.
Hadoop never wins. Its the worst of all possible worlds.
The compressed data can fit onto a local SSD. Decompression can definitely be streamed.
I could be tempted to do some fun on an NVL72 ;-)
If you want speed, just have your database stored in the same place as your application, locally, rather than hopping across the world to retrieve data that can be located next to the code.
That would probably be the easiest thing to do to get a real measured performance gains.
As other commentators pointed out, computers are extremely powerful. This isn't 1995, you can easily host everything in the same local area and get a very responsive application with very minimal needs to worry about resource constraints.
Given how primitive SQLite's optimizer is and how similar the storage and execution engines between the two are in terms of architecture, this seems unlikely to be the norm unless you did something wrong on the Postgres side. (Of course, no RDBMS optimizer will always give the best answer, so there's bound to be such cases.)
In which case it makes the analysis a bit less practical, since the main use case I have for fancy data processing tools is when I can’t load a whole big file into memory.
Unix shell pipelines are task-parallel. Every tool gets spun up as its own unix process — think "program" (fork-exec). Standard input and standard output (stdin, stdout) get hooked up to pipes. Pipes are like temporary files managed by the kernel (hand-wave). Pipe buffer size is a few KB. Grep does a blocking read on stdin. Cat writes to stdout. Both on a kernel I/O boundary. Here the kernel can context-switch the process when waiting for I/O.
In the past there was time-slicing. Now with multiple cores and hardware threads they actually run concurrently.
This is very similar to old-school approach to something like multiple threads, but processes don’t share virtual address spaces in the CPU's memory management unit (MMU).
Further details: look up McIlroy's pipeline design.
The comments here smell of "real engineers use command line". But I am not sure they ever actually worked with analysing data more than using it as a log parser.
Yes Hadoop is 2014.
These days you obviously don't set up a Hadoop cluster. You use the cloud provider service provided (BigQuery or AWS Athena for example).
Or map your data into DuckDB or use polars if it is small.
Now, you could argue that that’s cheating a bit and introduces preprocessing that is as complex as running Hadoop in the first place, but I think it depends.
In my experience, though, most companies really don’t have big data, and many that do don’t really need to.
Most companies aren’t fortune 500s.
I used to work at Elastic, and I noticed that most (not all!) of the customers who walked up to me at the conferences were there to ask about datasets that easily fit into memory on a cheap VPS.
It really feels that way. Real data analysis involves a lot more than just grepping logs. And the reason to be wary of starting out unprepared for that kind of analysis is that migrating to a better solution later is a nightmare.
In that case to get best performance, you’d have to shard your data across a cluster and use mapreduce.
Even in the authors 2014 SSDs multi-core consumer PC world, their aggregate pipeline would be around 2x faster if the work was split across two equivalent machines.
The limit of how much faster distributed computing is comes down to latency more than throughput. I’d not be surprised if this aggregate query could run in 10ms on pre sharded data in a distributed cluster.
PCIe would have to be millions of times faster than Ethernet before command line tools are actually faster than distributed computing and I don't see that happening any time soon.
I like this one where they put a dataset on 80 machines only then for someone to put the same dataset on 1 Intel NUC and outperform in query time.
https://altinity.com/blog/2020-1-1-clickhouse-cost-efficienc...
Datasets never become big enough…
Just because you don't have experience of these situations, it doesn't mean they don't exist. There's a reason Hadoop and Spark became synonymous with "big data."
The solutions are well known even to many non-programmers who actually have that problem:
There are also sensor arrays that write 100,000 data points per millisecond. But again, that is a hardware problem not a software problem.
Having materialised views increases insert load for every view, so if you want to slice your data in a way that wasn't predicted, or that would have increased ingress load beyond what you've got to spare, say, find all devices with a specific model and year+month because there's a dodgy lot, you'll really wish you were on a DB that can actually run that query instead of only being able to return your _precalculated_ results.
Not only is this a contrived non-comparison, but the statement itself is readily disproven by the limitations basically _everyone_ using single instance ClickHouse often run into if they actually have a large dataset.
Spark and Hadoop have their place, maybe not in rinky dink startup land, but definitely in the world of petabyte and exabyte data processing.
Bane's rule, you don't understand a distributed computing problem until you can get it to fit on a single machine first.
[1] https://news.ycombinator.com/item?id=8902739