Making 768 servers look like 1
planetscale.com142 points by hisamafahri 18 hours ago
142 points by hisamafahri 18 hours ago
What about sequences? The example shows an auto-incrementing user ID. How’s that possible without contention between all shards? Is the proxy responsible for sequences?
What about foreign keys? Do they all have to live on the same shard? How do you do distributed transactions?
On cross-shard reads: how do you do sorting? And cross-shard joins?
I’d love to be proven wrong, but I suspect the 768 servers look like 1 only on the very surface, and you’ll get wildly different characteristics from cross-shard and single-shard queries.
I personally would prefer if they _didn’t_ look like 1 if they can’t behave like 1.
Of course 768 servers NEVER behave as 1. This is physically impossible.
Global services using relational dbs typically severely restrict queries that run against the cluster. So no joins, no intervals, no grouping, etc.
Transactional queries are usually limited to something like "get a single record, preferably from cache". For many typical web services this can go VERY FAR. Only a handful of global services needs more than a few dozen database servers and a caching cluster. In fact, i have seen major businesses running off a pair of very big postgres instances.
Analytical stuff is extracted into dedicated storages optimized for throughput, like Snowflake or Redshift or BigQuery.
> Transactional queries are usually limited to something like "get a single record, preferably from cache".
simply wrong
This seems like the important distinction: making the infrastructure look like one database to the application is different from making it behave like one unrestricted relational database.
At what point does hiding the sharding become counterproductive? I imagine teams still need a fairly deep understanding of shard keys, query routing, and failure modes to avoid accidentally expensive cross-shard operations.
Yes!
Distributed systems introduce severe restrictions on what can be reasonably done at scale. Having a single connection string is one thing, being able to do a massive JOIN is another (and should only be ever done in analytical databases).
The question is not "when sharding becomes counter-productive" but "when it starts making sense".
With sharding something somewhere has to know how to route queries to subsets of data. So it is a complexity price paid for being able to scale. If one can avoid paying this cost then he should.
And USUALLY cross-shard queries are not just expensive but simply impossible in operational clusters. Like, if you do COUNT on a table, you only count within a single db shard table.
Parent commenter was not asking about when sharding becomes counterproductive, it was asking about when hiding sharding from the application becomes counterproductive.
The point is that it’s a leaky abstraction. It should not be at the database proxy level.
Instead, the application should be responsible to route queries to the correct shard. That way it can’t make cross-shard queries or cross-shard transactions accidentally, it must do so explicitly.
A lot of what you're referring to is dictated by the sharding strategy. Vitess and Neki both let you configure this via the VSchema / data topology (That's what I'm getting at here)
https://planetscale.com/blog/making-768-servers-look-like-1#...
This decision matters a ton. I drilled into this specific point on sharding in one of my articles from awhile back:
https://planetscale.com/blog/database-sharding
This is something you face any time you shard data, no matter the system. Single-shard queries are preferred to cross-shard ones. We want 99% of what we do to be single shard, but occasionally cross-shard work is unavoidable.
> We want 99% of what we do to be single shard, but occasionally cross-shard work is unavoidable.
Why optimize for the 1% and put the logic in a database proxy, then? It’s a leaky abstraction.
Require the application to explicitly and loudly make a query for each shard and collate the results.
I had this question too, even their "Database Scaling" course doesn't get to it.
https://vitess.io/docs/faq/sharding/advanced/
Looks like it's not drop-in, you need to heavily modify your data and indexes structurally to make sharding performant. Cross-shard joins are possible but need to be designed for explicitly (to be fair, indexes also need to be designed explicitly in regular SQL). IDK if a drop-in solution is possible or if this is an information-theoretical limitation.
> The example shows an auto-incrementing user ID. How’s that possible without contention between all shards?
Why not divide ID range (63bit?) by the maximum planned number of shards and then set on each shard it's min/max value so ranges will not overlap?
Because if the business grows, then you will need to add servers, and reallocate each existing shard's min/max values. If you left room to grow (divide amongst 2^31 instead of 2^63) you can go a long way before you have to solve that problem. If the business turns out to be wildly successful, you'd probably set up a whole new pool anyway and migrate over (and get a new set of pre-allocated ID ranges).
This might be the perfect case to use an UUIDv7 as a key instead of a number.
Consistent hashing helps solve this. You don't divide into shards based on the id, but on the consistent hash of the id. Then when you have to add shards, it's easier to distribute data to the new shards. If you have 3 shards and expand to 4, only 1/4th of the data from each shard gets moved to the new shard.
low-traffic auto increment like user creation can be relatively performant by reserving batches up front, then using those in each shard.
But the use-case for this is mostly if you do not need, or have very limited use for anything cross-shard.
I disagree with the opening premise:
> A single database server cannot handle such demand, so we must spread the queries and data out across many servers with database sharding
Did you max out the capacity of the best server you can buy?
Such a database can serve millions of customers (the numbers given).
You always want to scale up the other parts first, request handlers, caching, etc. The day you can no longer inspect the essential state of your system is the day your company better be included in NASDAQ and ready to pay a few hundred engineers 300k salaries.
Spreading requests out across hundreds, thousands, and in some cases even more is precisely what is done in the industry for big databases! Good examples:
cashapp: https://code.cash.app/planetscale-metal github: https://github.blog/engineering/infrastructure/partitioning-... etsy: https://www.etsy.com/codeascraft/migrating-etsyas-database-s...
These companies could not realistically operate off a single database server.
The cashapp post is actually working towards GPs point. It explains that networked storage MySQL was unreliable and expensive. They migrated to $fancytool which offers a db instance with attached local nvme.
While cashapp may actually need sharding, there are so many companies who are overpaying for shitty performing network storage databases.
Not to mention you are subject to cloud provider networking and compute allocation code. They change it. Big slow network storage DB suddenly gets even worse and you don’t have leverage to have it fixed on their end.
> These companies could not realistically operate off a single database server.
I want to see it fail first. I suggest their org chart has more to do with their architecture than database performance.
Even so you’re in the category I said. Hundreds of expensive engineers maintain this stuff.
> I want to see it fail first.
your lack of experience at scale doesn't really invalidate this architecture
You’re selling a database technology without understanding ACID or transactions. You are the one who is naive and inexperienced.
Your sense of scale is entirely artificial and cargo cult. And you’re committed to selling that problem.
Have you tried maxing out the server yet? Got a write up about it?
Well, they are selling this thing so they don't want you to buy a big server (with a read replica) as that's much cheaper.
That's not true! We have demand for bigger machines and we also sell them. You can go checkout the PlanetScale website. See: https://x.com/samlambert/status/2077197049129587150
There are truly customers that bigger machine no longer cuts.
Disclaimer: I'm an Engineer at PS.
I did not say there is no market, but I am willing to bet whatever that the majority of your client base would be much cheaper off with a simple setup where they can scale the entirety of their company’s existence with one machine. Not even a large one. This cloud stuff has its place but not for most. For them it’s either themselves overspending or wasting VC money.
We have tons of customers who do exactly this. It's great. Sharding is for customers who out grow this path.
I found the article very helpful from a technical perspective, and didn't focus on the number too much, as it could easily be swapped and the decision-making process for when to shard is kinda out of scope.
But I hadn't considered this, so thanks for pushing back. Good to keep in mind their incentives.
I will say, since their product is a proxy whose interface is a single SQL connection, you should in theory be able to do dev queries through that black box, much the same as application queries? What is so scary here that it would require a hundred engineers?
> you should in theory be able to do dev queries through that black box
Because it’s a leaky abstraction which is trying to make guarantees over network connections which are extremely difficult to make within the same kernel.
A few questions I would start with:
- is the system even ACID compliant?
In my reading of this article, no.
- is my sql feature set limited? Will it enforce all constraints? Or are their cross-shard limitations?
For example the article doesn’t discuss transactions or how they would roll back, or how they guarantee a consistent view of data.
Next:
- how does it respond to error?
- how does it respond to load?
This is a complex system and complexity breeds bugs. But now you don’t have the tools or procedures to investigate those bugs because you can’t poke the system at desk with tools. You can’t run experiments; you can’t even see all the data.
Thank you, good points. I'm just learning about this in real time.
It looks like it does support transactions, but they basically destroy the performance benefits: https://vitess.io/docs/faq/sharding/advanced/can-i-use-vites...
The more I read the more I'm struggling to understand the benefit of a router like this that sits on top of a monolithic SQL, vs a truly distributed DB like cockroach.
Like you I'd love to learn more about the internals of their actual SQL engine, which is just barely touched on in the article. The idea of ripping out the layer of a SQL interpreter that does just enough to route it to a real server tickles my brain in the same way as when I learned how node.js ripped the js interpreter out of a browser.
What problem are you trying to solve? What did you try?
Unfortunately the SQL language is the simplest part of the database. The concurrency and consistency guarantees are the key technology,
Most non FAANG orgs could probably serve all their customers from postgres on a laptop.
Let's hypothesize this is true.
I worked at a startup that had _money_ - a lot of it - flowing in. Stakeholders wanted to see growth, which meant features. They did not want to see us making platform improvements; those didn't show well to clients. Best throw money at the problem, rather than sacrifice the all-important features.
We had a number of architectural quirks that reflected this. One was that we ran our databases in AWS on ephemeral instance store. Yes, the kind that doesn't persist when you stop or lose the instance. The kind that you should never run a database on. We did it so that we could join them eight of them together in RAID 0 on the biggest instance types, and pull 400,000 IOPS to the disk that way. We paired this with hot spares of each shard, as well as hourly backups.
We did this to support a Rails / ActiveRecord data model, for data that should never have been part of a Rails data model. We had exactly one person working part-time on query optimization.
The maintenance burden was immense. The AWS bill, even more so. But again, we are hypothesizing that we could host all of this data on a laptop.
To an extent, maybe up until our thousandth customer or so, I actually think we could have. But we wouldn't have spent most of our time working on _features_ - we'd have instead spent most of our time trying to performance optimize. That would've meant the investors wouldn't have seen growth, which in turn would've highly impacted our success. We had a well defined domain, and optimizing the queries for performance was not it.
What if there was someone involved who had built web applications than run on laptops before and made different architectural decisions than the path your company went down.
And what if they cost more than your entire AWS bill?
You seem to have misunderstanding on what scales some things run at, and expense curves of getting them to run at those levels.
A poorly built app running on hundreds of servers can be massively expensive. A well built app running on a single server can also be massively expensive because people that can build apps like that tend to command whatever price they want under the banner of server cost savings.
Can you build software this efficiently? If you can't, then being an armchair expert about it helps noone. If you can, you're undercharging your clients.
It’s not deep magic or expensive. It’s simply asking hmm, how does my operating systems textbook solve this? Before asking how AWS solves this.
So, no, you don't actually work anywhere with a global team running tons of projects.
If my grandma had wheels she would have been a bike. Most of these decisions are made by someone very early on in the process, either a cofounder or an underpaid junior engineer, when the company has zero revenue and no PMF.
It's unrealistic to expect a startup to build a product (or at least it was prior to LLMs generating a lot of boilerplate) with best practices in place.
Surely the cost of running sql server on premium hardware with replication would be more than running on commodity hardware with sharding?
It depends on who you are paying for it, but generally a distributed system is harder to reason about, harder to fix, has weirder edge cases, and much more easily get into situations where it requires even MORE expertise to fix than just having a big honking server.
When you start calculating things that are not just the server, the single server looks cheaper and cheaper. How do you get a consistent backup? How do you do DR? How do you tune queries when it could go to this node or that node? Now writes are going to be significantly slower if you need multi-node commit because no matter what you are racing the speed of light on the network.
Another thing worth thinking about when running anything in parallel is the rather massive increase in error rate. When you're essentially rolling 768x the number of dice, failure outcomes will happen a lot more often than if you just roll the one or two.
It's a scale that transforms a 10 year MTBF to a ~5 day MTBF.
Only if your engineering resources and free and unbounded.
Even then I would probably use those resources to optimize software instead.
In the physical world of trucks and cranes no company would make that mistake to try to save 30-80k.
As I understand the author: They describe the journey of scale from a single database server over read replicas to sharding. And its not only about a single server, its also about network (backup, restore).
Hetzner has this for 100usd a month, 256gb ram can handle quite a lot of traffic already
€84.70 max. per month €0.1357 per hour CPU Intel Xeon E5-1650V3 RAM 256 GB Drives 2 × 6.0 TB Enterprise HDD Location #FSN1-DC1 Information IPv4 ECC iNIC
Lol that isn't a database server. Come on, if it has hard drives rather than (good) SSD in 2026, it's not a server you should be using for databases unless it's for something with next to no load, and you are also running your application server on the same machine.
I'm surprised no one has mentioned Multigres yet, they are the competitor of Neki. I'm a big fan of both and has been following them since they were announced last year. I think this blog post is the first time they are talking about the internals of Neki. In contrast, Multigres is being built in public since day 1, you can see their high-level architecture here [1], though I'm still waiting for more details on their sharding model.
Sibling post has author answering questions in comments: https://news.ycombinator.com/item?id=48925420
Even with a large database servers (10s of CPU cores, 100s of gigabytes of RAM) bottlenecks arise pretty quickly.
Err, do they? For what percent of real world use cases?
The database can scale to handle more traffic by adding replicas. An extreme example of this is OpenAI's use of 50 replicas on a single Primary.
So an extreme example is OpenAI needing 50 replicas, but we're doing five blades ... err, we're doing 768 servers because the need arose "pretty quickly"?
When we needed to store a petabyte of data (one million gigabytes), we'd need many more shards
For who? The United States government? How many end-users are running 1PB Postgres database on DBaaS?
> So an extreme example is OpenAI needing 50 replicas, but we're doing five blades ... err, we're doing 768 servers because the need arose "pretty quickly"?
If you read the OpenAI article, you'll see that they actually used sharding to offload a bunch of work from their "1 primary 50 replicas setup"
>>> "To mitigate these limitations and reduce write pressure, we’ve migrated, and continue to migrate, shardable (i.e. workloads that can be horizontally partitioned), write-heavy workloads to sharded systems such as Azure Cosmos DB..."
The 768 servers and 1PB example is just one of many configurations. A business with 10TB may choose to go from a monolithic database to a 8-shard setup to improve backup times, eliminate single-point-of-failure, have more breathing room for scaling.
Personally, I know several databases where single tables have +500GB and the database has +100TB. With this huge databases the restore and backup process over network become indeed a bottleneck. So I can agree with the author. Also, the author does not say that they can't start with a single database server and just read replicas and max hardware out. Real world use cases with +100TB I know about are Stock Market, Traffic Data, Warehouses, Analytics and Monitoring.
Took me an embarrassing amount of time to realize this is an ad.
How do you deal with master failures for specific shard? Does switch happen automatically once server becomes unresponsive?
> Most applications you've ever used function in this way, or at least did early in their existence.
I'm old enough that this is not true.
Looks like the GIF is fully built out in code. It's really nice to look at, well made, and easy to understand too. I wonder what program or code they used. I'd love to know.
p.sI thought it was a GIF, but it's an iframe. That was a nice little surprise.
Author here, thank you.
Technically, they are using js + gsap + svg embedded i the article with iframes.
Process-wise, I drafted most of them as static images in excalidraw, passed the images along to cursor for a first draft, applied styling rules, and then did a bunch of fine-tuning.
Thank you so much.!!!! It can't be easy to put this much care into helping others understand, with that hand drawn feel and everything. I really learn a lot from it. The animations were great and made everything really easy to understand. Have a nice day.!
Specifically, it's an JS-controlled/animated SVG embedded in an iframe.
Yeah, I'm looking at it in developer mode. It's using a GSAP timeline approach to update SVG properties. I'm curious how they handle security and caching for something like this. It looks like they're using Tailwind, at least. but this approach is really clean and nice.
It really feels like the best way to learn is by studying other people's code.
What kind of security and caching concerns do they need to handle to animate an SVG?
I didn't explain myself very well. With iframes, there can be security concerns in general. Though in practice, there's usually no real issue. It just surfaces the surface level risk that JavaScript could be tampered with, but it's almost never something to actually worry about.
As for caching strategy, cache invalidation issues do pop up from time to time. Since the filenames usually include a content hash, I'm guessing they're using a Vite style caching strategy. Cache invalidation might almost never be a problem, but you still plan for it just in case.
I sometimes feel that when the industry moved from pets to cattle, what really happened is that the cattle turned out to need an exotic farm to live on, negating the savings. You can have a few honking servers or you can hand massage exotic k8s setups on your fleet. Pick your poison, but dont delude yourself that the TCO of the latter is lower than the former.
Load balancers, microservices and horizontal scaling?
i do wonder how something like this can be generally implemented. i presume this must only support a subset of SQL/plpgsql, as some things would be.. utterly insane to manage manually. e.g., if i have a table with a btree-gist overlap constraint, or some inclusion-exclusion check-constraint (or literally any constraint that requires multiple rows to be fully determined - there are quite a lot of them), how on earth does this work?
there's a reason why postgres writing is (mostly) serialised (asterisk) to a single writer (asterisk asterisk). something something ACID, but in short by having multiple writers improves availability, but weakens integrity.
what did they use to make those diagrams/animations?
Hey, author here.
Technically, they're powered by js + gsap + svg.
Process wise (for most of them) I sketched them out in advance in excalidraw for figure out layout, then passed these along to cursor to have it build out an initial draft from the image, then used some styling rules to get all the styles inline, then did a bunch of fine-tuning.
How does Sharding works when You do complex joins ? Seems tricky , runs on each server and gets data back and aggregate it again?
I should start by disclosing that I work part time in the Oracle Database group, but - of course - my HN account is entirely my own, despite occasional mild shilling. The article itself is shilling for PlanetScale so that seems OK.
The author - certainly not deliberately! - says some untrue things about relational databases. The most important one is this:
> To understand why sharding is a necessary part of scaling relational databases...
But sharding isn't a necessary part of scaling relational databases. It's really just a requirement of simple databases like Postgres and MySQL.
If you want a relational database cluster that really does make 768 servers look like one, then you want what Oracle calls RAC ("real application clusters"). Your cloud will probably rent you access to one under a name like Autonomous Database. Self hosted it may be called ExaData, which is a unified hardware/software "rent a rack" style offering. You may be surprised to discover that it's not much more expensive than many managed Postgres offerings.
RAC can scale a non-sharded relational database horizontally. That means all queries can access all data, all SQL features like sequences and joins work, any server can take part in transactions with any other and in general it looks exactly like a really big single machine would. In other words, it's a synchronous multi-write-master system.
RAC scales very well. 768 servers is well within reach as long as your query patterns scale too i.e. don't all contend on writing to one row. Behind the scenes it uses a dedicated high speed RDMA network with lock coordination to transfer data blocks directly between nodes, never hitting disk for memory that's already in the buffer cache.
Additionally RAC is fully HA and supports rolling upgrades of the cluster whilst live. You don't need any NLBs or routers either. The client drivers automatically discover and load balance between nodes without needing intermediaries, transactions can start on one node and fail over to another without applications noticing, and so on. There's plenty of opportunity for caching and replication. You can run asynchronously replicated failover clusters, run multiple clusters in a Raft-driven globally coordinated super-cluster and can deploy coherent read-through caches anywhere; the main clusters will inform them the moment data in them becomes stale.
In other words, it can do a lot.
If for some reason you do need sharding then that's also supported with features like automatic sharding key distribution to client drivers that transparently route queries correctly, but most apps don't need this.
In case you're wondering why I say all this, firstly, obviously, I have a financial conflict of interest. But the database hasn't driven Oracle's stock price for a long time, so it's not a big one. These days it's all about cloud and AI.
No, the main reason is that HN fills up every month with blog posts where engineers talk about the incredible pain involved with scaling and running Postgres. And almost always, it's clear that they don't realize there's any alternative to that pain. It's not that they considered the options and then explain why they picked this one, it's that they think - as claimed in this article - that it's almost some fundamental limitation of reality itself, imposed by the laws of computer science.
There are lots of startups that lose time and money due to database problems they simply don't need to have. And that sucks. If they'd prefer to spend that time, pain and money to avoid using a DB from Oracle, fine, so be it. I won't argue with random devs about lawnmower memes. But if it's because they don't realize what's possible.... well, maybe someone will be helped by being aware of this. Database scaling problems are a choice, not an inevitability.
Lol about buying Oracle for performance!
But anyway, do you have benchmarks from anybody not related to Oracle showing that performance? Because Oracle forbids talking about it...
Not forbidden. You can email a specific address to share results/setup and get permission to publish if you want. Other commercial databases also do that, because there's so many ways to misconfigure a database to make it slow and competitors are strongly incentivized to do so.
The question is what you'd want to benchmark. For example, imagine testing Postgres with a write load that goes well beyond what a single machine can do. It would collapse and query latencies would go through the roof. A horizontally scaled DB would keep going and process all those queries. Would you accept this as evidence that Postgres is slow or would you say it's not valid to benchmark at traffic loads Postgres physically cannot handle, given it never claimed to scale horizontally? Stuff like this is where benchmarking gets complicated.
I used to work at a different company that sold a kind of database system. It was much faster than our nearest competitor, so we were surprised when that competitor claimed to an important customer they were just as fast as us. Their benchmark counted transactions that failed and rolled back due to overload (optimistic concurrency) as "successful".
I suppose Oracle will only allow publication if your benchmark results look good. Can you tell us about a time you did a benchmark that made Oracle look bad, and they still let you publish it?
> Other commercial databases also do that
That's why you won't see many serious comments bragging about the performance of SQL Server either.
Even though is has way fewer performance land-mines that you must design your entire architecture around than Oracle.
Anyway, you are correct that RAC is a very impressive piece of software. If Postgres had something like it, it would be a beast. It's not enough to save Oracle, though.
What sort of landmines are you referring to? This is quite interesting.
Still it's not just SQL Server and Oracle. PlanetScale - the author of this blog post - also had a DeWitt clause until literally just days ago. They replaced it with a slightly different clause but I wouldn't be willing to benchmark under their new rules.
https://planetscale.com/blog/transparency-in-benchmarking
It says benchmarking is allowed only if "the Benchmark is conducted in good faith and uses a fair and transparent methodology". What does good faith mean, or fair and transparent? You still effectively need their permission because otherwise they might decide you weren't benchmarking in "good faith" retroactively.
It's no surprise. The stuff that makes database benchmarking useless is the variety of features. You can always design a benchmark that makes a specific vendor look good or bad.
Like, Oracle can add missing indexes on the fly. If someone designed a benchmark in which the core table was huge but missing an index, PostgreSQL would die and Oracle would recover then fly. But how many would accept that as a fair test?
> What sort of landmines are you referring to?
For example the Oracle's incapacity of reordering tests of different kinds (as in constants, functions, rows...) when optimizing a query; or the quadratic run time of triggers that change data in large updates.
Does something similar to RAC exist for other commercial databases like Microsoft SQL Server or IBM DB2?
DB2 yes, SQL Server no.
There are some cloud databases that scale horizontally but they often only support a subset of SQL, or require careful design to avoid creating hot tablets due to using a shared nothing design e.g. due to hot read nodes.
[flagged]
[dead]
[dead]
[dead]