"Overall, we find a Postgres server can handle up to 144K of these writes per second. That’s a lot, equivalent to 12 billion writes per day."
Based on a problem I'm facing with Postgres today, I wonder if this really progresses as linearly as the article wants to make it out.
We're in the middle of evaluating Postgres as a replacement for MySQL, and experience notable slow-down for plain multi-row inserts due to index growth as soon as the table reaches just a couple of dozen million rows. It's an uncomplicated and flat (no constraints or foreign keys etc.) medium width table of about 10-15 columns and a handful of non-composite btree indices - and/or hash indices; we've tried mixing and matching just to see what happens - but ingestion drops to less than half already before 50m rows. At 100m rows the insertion performance is down to a fraction and from there it just gets worse the larger the table and its indices grow. It's as if there's some specific exponential cut-off point where everything goes awry. However, if we simply remove all indices from the table, Postgres will happily insert hundreds of millions rows at a steady and near identical pace from start to end. The exact same table and indices on MySQL, as closely as we can match between MySQL and Postgres, running on the same OS and hardware, maintains more or less linear insertion performance well beyond 500m rows.
Now, there's a lot to say about the whys and why-nots when it comes to keeping tables of this size in an RDBMS and application design relying on it to work out, and probably a fair amount more about tuning Postgres' config, but we're stumped as to why PG's indexing performance falters this early when contrasted against InnoDB/MySQL. 50-100m rows really isn't much. Would greatly appreciate if anyone with insight could shed some light on it and maybe offer a few ideas to test out.
(add.: during these stress tests the hardware is nowhere close to over-encumbered, and there's consistent headroom on both memory, CPU and disk I/O)
problem is table design and write amplification. Every row insert triggers update into every index, so you get classic amplification problem.
Separate your table into Cold (with all indexes and bells and whistles) and Hot (heap table with no indexes except PK).
Insert as many rows as you want into Hot heap, and then move them in the background into cold in batches, so that index recalculation is amortized across many rows, instead of per-row.
Another poster suggested partitioning, thats the same idea: separate Hot and Cold data into partitions and keep hot partition as heap
That’s a really cool idea I had not heard before, thank you for sharing this. It also feels like the type of thing a db ought to be able to do under the hood. I wonder why this is not a config (though there’s a pg extension for everything so maybe it does exist)
As far as I can tell, Postgres is not designed with this inclination towards doing lighter work when clients are waiting and piling up maintenance work to do in background. I think the background work it does is mostly running vacuum on tables now and then.
Contrast that with ClickHouse, for example. It operates in a different niche than Postgres (OLAP instead of OLTP) – with their merge tree engine family [1] that does data deduplication in background.
There is one project of modernizing Postgres' storage engine called OrioleDB [2], but I think the company got acquihired by Supabase [3] and maybe the project has not been progressing very quickly since then.
It wouldn't be atomic, and so would break transaction semantics.
If you committed a row update but didn't update the index, a subsequent query using the not yet updated index would not find the updated row correctly.
It would also only work for certain types of indexes, you couldn't do it for uniqueness constraint for example.
I do agree that in theory you could have some extension to the index declaration that covers all that, but my worry there would be that it would be non obvious and a foot gun. Doing it the way described above makes that break in semantics clear.
> If you committed a row update but didn't update the index, a subsequent query using the not yet updated index would not find the updated row correctly.
I wonder if you could make it so that queries read from both the index and the unindexed changes. It would be slightly slower but as long as the unindexed changes are kept small it might be fine.
My impression is that InnoDB (MySQL's primary storage engine) is doing something like this. We have never seen any slow-downs on adding to the data set I've discussed in this thread, even at hundreds of millions of rows, and per the nature of the system creating this data the majority of these rows are targeted for additional single-row DML within a few seconds of being inserted, with instantaneous effect.
> problem is table design and write amplification. Every row insert triggers update into every index, so you get classic amplification problem.
Yes, this is understood. In particular for b-trees that require some refurnishing when growing. What's less understood is why Postgres hasn't solved this in a way similar to how InnoDB solves it behind the scenes.
You've given us some idea of the volume of your data but there's no mention of what's ingesting it or how.
> during these stress tests the hardware is nowhere close to over-encumbered, and there's consistent headroom on both memory, CPU and disk I/O
This assertion is likely wrong - you're likely skipping over some metrics that has clues to what we need to know. Here are some questions to get the discussion moving.
- Is this PostgreSQL managed or self-hosted?
Your mention of "consistent headroom on both memory, CPU and disk I/O" gives me hope you're self-hosting it but I've heard the same thing in the past from people attempting to use RDS and wondering the same as you are, so no assumptions.
- Are you using COPY or multi-row INSERT statements?
- How much RAM does that server have?
- What is the fillfactor, max_wal_size and checkpoint_timeout?
We can clearly see with iostat/iotop that the server and its storage isn't overworked.
> Is this PostgreSQL managed or self-hosted?
We're evaluating on our own bare metal. It's an 8c/16t Zen 3 with 32 GB of RAM. Storage (where everything incl. the WAL is) is NVMe and the drives' true sustainable/synchronous write speed is about ~750 MiB/second.
The specs are far, far higher than required for something this basic. Total memory usage when we begin observing the problem is a fraction of what's available. The MySQL production and test environments running this without hindrances actually have only half the amount of RAM.
> Are you using COPY or multi-row INSERT statements?
Multi-row INSERTs, one per transaction, with anything from 100 to 500 rows. Evaluation simulates the volumes we can see through our APIs on production, though it omits everything but Postgres itself in order to test Postgres' ingestion capability without other factors getting in the way - it's just Postgres and a light-weight data generator mimicking production data.
> What is the fillfactor, max_wal_size and checkpoint_timeout?
Not sure about the fill factor. Everything is running on default Postgres 17.9 as packaged for Debian 13, so that would be 1GB max_wal_size and checkpoint_timeout is commented out with a default of 5 minutes. Haven't gotten to any performance tuning yet, would be thankful for any suggestions to try out. The only thing we've tried is disabling auto-vacuuming to see if it was too frequent, causing i/o contention or otherwise hogging throughput. Not really any noticable change.
> Are random UUIDs (part of) the index?
No, neither v4 UUIDs nor b-tree-friendlier v7 UUIDs, but a couple of the indexed columns contain random integral numbers that can become quite sprawly inside a b-tree. We do observe somewhat better results indexing these with HASH indices instead, which also makes a lot more sense for that particular data and how we query it. For evaluation our outset was to stick with b-tree indices because that's what's used on the MySQL setup (InnoDB does not support on-disk hash indices).
> What's the iostat or wa during the slowdown?
When we reach the point where ingestion speed has shrunk to about a third we observe iowait peaks having grown to some 15%, which tells us the problem is likely Postgres spending more and more of its time shoveling in the indices rather than storing actual row data. Maximum written data at about 150 MB/second is just a fraction of what the NVMe drives can sustain. None of Postgres' individual processes ever top out anywhere close to 100% of a single core on the machine. Total memory usage is less than 2 GB, and here we suspect we have a lot of tuning to look into. To contrast, the MySQL setup is greedier with both CPU and memory usage, and an educated guess is that the major difference allowing it to keep ingesting hundreds of millions of rows without slowing down is that InnoDB, without us having to jump through any hoops, on one hand defers persisting new index data so that the DML can be finalized as quickly as possible, and on the other hand operates on its indices in a much more efficient way than Postgres.
Everything needed to reproduce the problem is in this paste, which contains a neutral version of the full table and indices, and a simplified version of the data generator: https://paste.debian.net/plainh/ddc819cb
That's a phenomenal script that absolutely belongs on a dba.stackexchange question and even the PostgreSQL mailing list.
I strongly encourage you to post to https://dba.stackexchange.com/, as a HN thread is the wrong place for this discussion (there's a lot of tuning ahead of you and others who are in your situation in the future might skip this nested thread) but be forewarned, the TLDR from dba.stackexchange will either be a quick "you need atleast 64 GB RAM for your PostgreSQL usecase" or there will a better, thorough discussion of increasing max_wal_size and lowering the fillfactor (which is what you're looking for). The ideal answer should even walk you through BRIN (vs btree) indexes. I'm asking you to post there because it will enumerate the WHY of all of these. At that point you can make an informed decision if this all would be worth it.
Now I empathize that MySQL is doing a phenomenal job at only a quarter of that (16 GB you said?) but not for the reasons you might think (and one can certainly argue, nor should an end user care!). MySQL's method of buffering (InnoDB change buffers) and its clustered index gives you the performance you like when suddenly doing bursty writes aka "write sprint". I need you to be aware of that (and dba.stackexchange responses will certainly address that).
I would have written a lengthy post on what to do next but I must first ask:
- Why are you evaluating PostgreSQL in this case when MySQL seems to work well? For example: does it feel like your aggregations are getting slower? As you can see, with PostgreSQL, you will have different set of tradeoffs (RAM, tuning, VACUUM)
- Are there real, limiting business constraints that force you to operate on less than 64 GB of RAM given your volume and throughput expectations (like FF limitations, or these are smaller machines on the edge, etc)
- If you can, as an experiment, while you write your dba.stackexchange question and for the PostgreSQL mailing list, you can tweak multiple parameters and tell me what you see:
- I'm concerned you have a `shared_buffers = 128MB`. Set it to `shared_buffers = 8GB` (give the B-trees room to live in RAM)
- Increase `max_wal_size = 16GB` (stop the checkpoint flooding and let Postgres "breathe" during your batch inserts).
- Increase `checkpoint_timeout = 30min` (set 30min to the actual window it takes; also, this is temporary but this should push checkpoints out so they don't interrupt your "write sprint").
- Set `maintenance_work_mem = 2GB` (should speed up index creation and vacuuming).
- Lower `fillfactor` on those specific 8 indexes from the default 100 down to 70. The B-trees should now have the ability to absorb those inserts better.
This should get you in the right direction, googling for the right documentation, but there's even more ahead of you, including a separate discussion about BRIN (vs btree) indexes.
> defers persisting new index data so that the DML can be finalized as quickly as possible
"defers persisting" might get misread as if it doesn't write to disk - it does but you're close and you will uncover more :)
MySQL defaults are specifically tuned for your "write sprint" usecase, infact, to actually mask the IO latency of secondary index updates but if your real usecase is not just large sudden bursts of writes to a table that has a btree index, you certainly will appreciate this effort. Happy weekend!
Thanks a bunch for all of the config tuning advice. Can see huge improvements in both flattening the creeping slow-down and really pushing back the point where we eventually hit the wall. Just lowering the fill factor on the indices made a substantial difference even before permitting Postgres to use more RAM. Using BRIN indices for certain columns, as suggested by several others in the discussion, also resulted in some smaller but still valuable speed-ups. It's obvious by now that a careful combination of curated choices greatly matters for the full scope of this data. On the whole, in fact, it looks like all of the combined advice clear this hurdle up so much that we can no longer say we observe Postgres having an actual problem dealing with our current volumes and intensity of ingestion.
We run quite a few of these setups, so cost absolutely plays a big part in the size and selection of hardware. We're evaluating Postgres because of a mergers and acquisitions event prompting us to consolidate portions of the platforms for the sake of savings and reducing the overall "sprawl". Conversely we're also testing how the other party would fare if some of their stuff were to run under MySQL instead of Postgres. There's a chance we end up leaving things just as they are.
> On the whole, in fact, it looks like all of the combined advice clear this hurdle up so much that we can no longer say we observe Postgres having an actual problem dealing with our current volumes and intensity of ingestion.
I am extremely happy for you!
This is fantastic news. I would really appreciate a followup post on DB SE and/or PostgreSQL ML because I assure you, there are others in your position. A deeply nested HN thread is not something that will be visible - you would be saving a lot of people heartburn.
Also remember, we just got started! The tips I gave you earlier is just a drop i the bucket. In the days of AWS RDS, knowledge about these are a "lost art" but running PostgreSQL yourself is extremely empowering. If you could make the posts, please do and I also urge you to test out https://github.com/pgsty/pigsty
With some extra admin work, you can greatly increase your insert throughput, as long as the table load is comprised mostly of inserts:
1. Partition your table by range of a monotonic ID or timestamp. Notice the primary key will have to contain this column. A BIGINT id column should work fine;
2. Remove all the other indexes from the partitioned table. Add them to all the partitions, except the latest one. This way, the latest one can endure a tough write load, while the other ones work fine for reads;
3. Create an admin routine (perhaps with pg_cron) to create a new partition whenever the newest one is getting close to the limit. When the load moves to the newer partition, add indexes concurrently to the old one;
4. You'll notice the newest partition will the optimized for writes but not reads. You can offset some of that by replacing BTREE secondary indexes with BRIN [1], particularly the one with bloom operator (not to be confused with Postgres Bloom regular indexes [2]). BRIN is a family of indexes more optimized for writes than reads. If the partition is not too large, it shouldn't be too bad to read from it.
5. Later you can merge partitions to avoid having too many of them. Postgres has commands for that, but I think they lock the whole table, so a safer bet is to copy small partitions into a new larger one and swap them manually.
I like this strategy a lot, but the performance of read queries suffer if they span partitions, correct?
The issue I'm facing is a very large table, that is both write and read heavy, and the reads do not fall into a specific range of values for any particular column, so I don't think partitioning is an option.
Yes, partitioning will decrease a bit the read performance of queries not correlated with the partition key. That's why you need to periodically merge smaller partitions, so that you can keep the overall partition count bounded.
It is a lot of admin work, but if you really need to scale up Postgres write throughput, I don't see many other options without increasing hardware costs.
I assume you have already picked the low-hanging fruit discussed in the neighboring comments - batch writes, make sure you are using COPY instead of INSERT, tune Postgres parameters adequately and use the fastest disk you can grab for the WAL.
Partitioning is not all that expensive. It is definitely worth testing for your specific workload. We use TimescaleDB, which relies heavily on postgres partitions, have a bit under 100 million rows in our active set (last 90 days), across 120 partitions (device*time), and it works nicely. Over 100 partitions is probably a bit many for this workload, but since it works OK we have not changed it.
These are good suggestions but I'm apprehensive they might come back and say they have 64 GB (or less) of RAM or they are using PostgreSQL RDS on AWS or something.
I don't think it really matters in terms of their question though, given MySql on the same specs doesn't have the problem and postgres does. Quite clearly it has something to do with indexes and what is the wall postgres is running into that causes the drop off on quite low amounts of rows. If the answer is just get more RAM, it kind of implies postgres is not really that scalable. Especially if the drop off is proportional to the number of rows.
Last time I almost used a hash index in Postgres, I learned it was an incomplete feature and not crash-safe yet. This was v9.3? At that same time, MySQL had them and they were ok to use.
Later that got fixed, but I haven't tried again since, just been using btree because it seemed like Postgres favored that and it has theoretical advantages too.
They are fully stable and perform very well in Postgres today. There are some caveats, but they don't result in any sort of hiccups or unpredictable behavior.
Yeah that's what I've read too. Just haven't gotten around to trying them.
You've probably already read the Postgres docs on hash indexes, but just in case, it says "hash indexes may not be suitable for tables with rapidly increasing number of rows." I agree with the other commenter that it's worth at least trying without them if you haven't already, even though you're already VACUUMing.
On some data, for certain ways of using and querying that data, it makes a lot more sense than a b-tree. When we use hash indices for some of the columns in our test just to see how Postgres will perform, and run our test towards a quarter of a billion rows where Postgres' 32-bit hash indices have a relatively high chance of colliding, querying these hash indices and returning the single requested row is still instantaneous.
The problem is row locks when using interactive transactions over the network and contention. That can absolutely kill your performance with postgres, there's not really anything you can do to get around it (other than avoid interactive transactions). [1]
We had an interesting architecture situation at work. Puppet Enterprise uses a single Postgres server. The company had moved from a recommendation of using a single PuppetDB API node (which fell over at high load) to running a PuppetDB API server on each compiler node.
That, however, came with its own set of problems. Of course you have to tune for concurrent connections as you scale wider, but there were much more serious contention issues than you'd expect, and the compilation times were terrible too. It turned out to be because those transactions locked the DB during their (synchronous) operations, and we had a globally distributed set of compilers in order to serve globally distributed traffic.
The solution ended up being to run a separate cluster of API servers in the same region as the DB. The expensive calls from the compilers to the API servers were largely async https so they didn't have to wait on the API nodes, and the API nodes could talk to the DB synchronously with low latency.
The need to change two separate tables atomically means that you probably have info duplicated. Also it can be non-ideal in a web backend cause it means keeping DB connections open longer.
Ext4 with journaling. I've played around with XFS as well just to see what would happen. No major difference. I'm certain it's caused by the way Postgres builds its indices.
While I really like ZFS personally for 99% of things, for "need maximum performance with PostgreSQL and we're already pushing high end hardware to the maximum" ZFS is reportedly the wrong choice.
So you open transaction, insert multiple rows, commit? And you're not using any special xact settings like SERIALIZABLE mode, right? Normally you use COPY FROM if it's a huge number of rows in some batch process, but not like inserting a few rows handling a backend request or something.
I would try filling up the table, forcing a vacuum, then timing the inserts afterwards. Not that you should need to vacuum in real usage, but it might uncover a problem.
> So you open transaction, insert multiple rows, commit?
Pretty much, yes. High volume rapid multi-row INSERTs. Nothing special involved in the DDL or how the data goes into the table. Vacuuming the table (whether automatically or between every 10-20m rows) makes no difference. The indices still penalize the performance substantially the larger they grow.
if the index updates are your bottleneck you can often get away with using a much lighter index (for example BRIN) at a cost of slightly slower queries. This is very often a great and much overlooked tradeoff.
I've run mysql for years in production and have spent probably 30 minutes thinking about managing it. Unless there's some psql feature you need (unlikely), it'll just become a severe pain in your ass down the road because you set something up "wrong."
Just vacuuming can barf completely, leaving you dead in the water.
TL;DR: if you aren't a DBA and don't want to play one on TV don't bother with psql.
Agreed. And, any time I ask for help with something DB-related, people just chastise me for not running postgres and blindly tell me that's my problem. That sort of community is not what I want to look forward to.
It scales beyond the needs that most people have in most situations.
The constant problem is that "big scale" always means "larger than I've seen", so on any project larger than a person has encountered, they assume they need to pull out the big guns. Also, people worry about things like what happens if they really *do* scale 10 years from now.
Neither is a practical concern for nearly anyone who will ever face this decision.
And then yes, of course, some people have problems that actually can't be solved by Postgres. But verify this first, don't assume.
What gets me is that some people seem to ignore the very real cliff of complexity that ramps up the moment you move to eventual consistency. If you need it you need it, but you have to bake in those assumptions everywhere - and they commonly break the default assumptions of those who don't have a bunch of experience with it or haven't architected their approach to work around those.
And in many cases it's those architectures that force more complexity and make it appear like they have much bigger challenges then they do. Great for resume driven development, but often you can get away with far less.
When you scale horizontally from day one, it usually gives fault tolerance for online service, and this story is not very friendly in case of vertically scaled PG.
My point is that in PG, fault tolerance requires untrivial investments.
If you pick horizontally scalable DB (foundationdb, cocroachdb, scylladb, tidb, etc) they all give you fault tolerance for free without much involvement from your side, because it is part of the nature of those DBs.
DBOS is amazing when it comes to Durable Workflows. There are others in the space - the most popular one being Temporal but I argue, Temporal is also the most complicated one. I often say Temporal is like Kubernetes while DBOS is like `docker compose`. (and for those taking me literally, you can use DBOS in Kubernetes!)
I don't realize why DBOS is not nearly as popular as Temporal but it has made a world of difference building Durable Queues and Long Running, Durable Workflows in Python (it supports other languages too).
As they show in this article, Postgres scales impressively well (4 billion workflows per day, on a db.m7i.24xlarge, enough for most applications), which is why, if you have your PostgreSQL backup/restore strategy knocked out and dialed in, you should really take a close look at DBOS to handle your cloud agnostic or self hosted Durable Queues and Durable Workflows. It's an amazing piece of software founded by the original author of Ingres (precusor to Postgres - the story of DBOS itself is captivating. I believe it started from being unable to scale Spark job scheduling)
The reason that DBOS isn't as popular is because it's younger. DBOS launched in the form we know it in 2024. Temporal is much older; Temporal is technically a fork of Cadence and Cadence released originally in 2017, with Temporal forking and releasing back in 2020. When all three are trying to be "the same sort of thing" and that thing is new, it's hard to show up 7-8 years after the trailblazers and say "oh yeah, we're clearly better" when the existing thing works and is used by tons of folks.
Temporal is a dumpster fire, they've gotten so much VC funding (recently had D, 300M at a 5bn valuation) with ... nothing to build except ways to trap customers into their SAAS.
I give them about a year or two before the wheels fall off, then it's off to Broadcom and friends.
But I could be wrong as now they're not in the 'durable execution' space at all, it's 'durable execution for ai' according to their latest conference.
Got to spend that VC dosh somewhere I suppose, they're certainly not spending it on making a good product.
Temporal employee here. I'm very surprised by your comment.
It's true that we recently had a Series D and that VC firms recognize the value of what we do. The Temporal Server software is 100% open source (MIT license: https://github.com/temporalio/temporal/blob/main/LICENSE). It's totally free and you don't even need to fill out a registration form, just download precompiled binaries from GitHub or clone the repo and build it yourself. You can self-host it anywhere you like, no restrictions on scale or commercial usage. We offer SaaS (Temporal Cloud), which customers can choose as an alternative self-hosting, based on their needs. The migration path is bi-directional, so not a trap by any definition.
Regarding AI, Temporal is widely used in that space, but that does not negate the thousands of other companies that use Temporal for other things (e.g., order management systems, customer onboarding, loan origination, money movement, cloud infrastructure management, and so on). In fact, our growth in the AI market came about because companies who were already using Temporal for other use cases realized that it also solved the problems they encountered in their AI projects.
Maybe. But as someone who happily self hosting pretty big Temporal workloads for my day job (I inherited it from early adopters circa 2022), it definitely does not feel like a dumpster fire. It chugs along unglamorously and I enjoy working on it.
DBOS looks simple (good), but from the docs below, executor elasticity appears to be locked behind license purchase. So it truly is like docker compose, good and bad parts?
>When self-hosting in a distributed setting without Conductor, it is important to manage workflow recovery so that when an executor crashes, restarts, or is shut down, its workflows are recovered. You should assign each executor running a DBOS application an executor ID through DBOS configuration. Each workflow is tagged with the ID of the executor that started it. When an application with an executor ID restarts, it only recovers pending workflows assigned to that executor ID.
This is a good question! No, it's not like docker compose (I imagine you implied the swarm and hub pull limits?)
DBOS Conductor is an out of band management service that IIRC helps you mainly observe your DBOS and recover failures in a seamless way. As far as I could see, it's not necessary, for you to use DBOS workflows and queues. Don't quote me though and reach out to their forum and verify in case I'm missing certain usecases.
Personally, I do not use DBOS Conductor - I have my own observable setup using Grafana/VictioriaMetrics as my workflows are instrumented with OTel. I had initially set Conductor up for development (and it looked to be free for development although I recall some major limitations on how many workflows etc - which is why I put my own alternate monitoring setup).
They also have a very reasonably priced cloud hosted DBOS Conductor. I think my first 30 days were completely free and then they moved me to a "hobby" tier. It's a fantastic way to help decide whether it's for you.
I believe DBOS Conductor is how DBOS pays the bills but you can use DBOS workflows and queues unlimited without DBOS Conductor. If you don't want to pay for Conductor - their out of band management service, you can put together your own just fine, like I did. My own Grafana/VictioriaMetrics setup answers my questions but I would imagine Claude/Codex/Cursor should be able to put something fairly useful if you didn't want to go down my route.
> executor elasticity appears to be locked behind license purchase
DBOS has designed their system to be extremely flexible and extensible. While yes, Conductor can absolutely manage your executors for you, it's not the only thing that can. You're not limited to using Conductor. As I said, I manage my own - everything you need to know to do so is in the code and documentation. They even have a document for LLMs and agents. I have had to interact with the DBOS team 0 times to set everything up.
I prefer this business model (an optional tool - Conductor, is paid) vs. DBOS offering just everything across the stack on a "free tier" but with caps on DBOS workflows and queues. In their current business model DBOS workflows and queues are completely uncapped (atleast from what I can make out).
If you do reach out to them, I would appreciate if you let me know anything to the contrary.
We have a simple worker setup and temporal is pretty easy to setup
Out only issue is really needing an intermediary data store for task result storage
We are using DBOS in new projects as it's even simpler and the downside (task log interface behind saas) is easily remedied with a copilot generated task viewer
There is a steep learning curve sure, and there is lot of manual steps to do some critical actions perhaps.
Documentation however is the last thing I would criticize PostgreSQL for. PostgreSQL documentation is by far best amongst all databases open or closed. It is exhaustive, well maintained, well written, and quite accurate and kept up for decades - The oldest supported version (in documentation) is 6.3 released in 1998.
Only played around with it but you can use patroni, etcd and HAproxy to achieve this. It’s a pain, but I believe there was some kind of coolify-style open source application to do this for you but I can’t for the life of me remember its name
Atleast I hope you are! Nothing else has been as well battletested. Unfortunately, perhaps because if its name, it gets no facetime on HN. Its last few mentions here barely received attention it deserved.
Yep, this is what I think about when “scaling” is mentioned. Maybe I’m too distributed-compute brained, but throwing CPU at a db isn’t what I was hoping would be the answer.
So the point of distributed compute is to reduce the compute needed? I’ve generally found that distributed compute requires more compute than vertical scaling while getting clobbered by network bandwidth / latency.
Theoretically with 2 to 10x compute required and in practice 100 to 500x
I think for databases horizontal scaling for writes only makes sense once vertical scaling stops working. It comes with high complexity, annoying limitations, and often higher cost.
Horizontal scaling for reads on the other hand is much easier. If you have multiple replicas for high availability, you might as well put them to work. It can also reduce the risk for read heavy tasks interfering with transaction processing. You can even go a step further and replicate to a database that's optimized for analytical tasks.
Horizontal scaling for stateless applications (e.g. web servers or job processors) is often easier and more robust than vertical scaling, with little to no downsides.
The advice I’ve gotten is that you want to move computation to data that is already distributed. The cost of moving large amounts of data usually dwarfs compute costs (usually, not always), and so the performance win comes from distributing the computation and then (depending on the problem) centralizing aggregate results.
Practically trivial to do in 2026 even by hand, and there are a couple of ready to use solutions that even make it automated.
If you're running it in kubernetes with cloudnativepg it's even easier.
The only thing it doesn't do well is master master replication which is why most of these does it scale posts mostly talk about how slow writes are. And they are pretty slow.
This post conflates scalability and performance. PostgreSQL is fast on smallish systems, but try adding more CPU cores and you'll see performance gains will not be linear at all. Modern server can ship with 256 or more cores and a single instance of PostgreSQL will struggle to take advantage of these.
What kind of cases were you measuring? I would think that, e.g. 256 separate long-lived connections in a setup like that would scale less-than-linearly but not dramatically so?
> we find a Postgres server can handle up to 144K of these writes per second. That’s a lot, equivalent to 12 billion writes per day.
Based on the shown graph, this is misleading at best, essentially false. After 120K writes/s p50 spikes from 10ms to 1s (1 second for a write!!!!). That's two orders of magnitude latency spike, and an unacceptable one for an OLTP workload. It clearly shows the server is completely saturated, which is clearly a non operational regime. Quoting 144K is equivalent to quoting the throughput of a highway at the moment traffic comes to a standstill.
Based on this graph the highest number I'd quote is 120K. And probably you want to keep operating the server within a safe margin below peak, but since this is a benchmark, let's call 120K the peak. Because actually p50 is not even the clear-cut. It should be a higher percentile (say p95) at which latency is within reasonable bounds. But for the shake of not over complicating, it could be taken as a reference.
> We found that the bottleneck was in flushing the Postgres write-ahead log (WAL) to disk
Therefore, you are not measuring Postgres peak performance, but rather Postgres performance under the IO constraints of this particular system. Certainly, 120K IOPS is the maximum that this particular instance can have. But it doesn't show if Postgres could do better under a more performant IO disk. Actually, a good test would have been to try the next instance (db.m7i.48xlarge) with 240K IOPS and see if performance doubles (within the same envelop of p50 latency) or not. And afterwards to test on an instance with local NVMe (you won't find this in RDS).
> From [1]:
> Postgres insert throughput
> uv run python benchmarks/postgres_insert.py --rps 1000 --duration 300
300 seconds test duration?? This is not operational. You are not accounting for checkpoints, background writer, and especially autovacuum. Given that workflow pattern includes UPDATEs, you must validate bloat generation (or, equivalently, bloat removal) by a) observing much longer periods of time (e.g. 1h) and b) making sure the autovacuum configuration (and/or individual table vacuum configuration if required) makes bloat contained in a stable way. Otherwise, shown performance numbers will degrade over time, making them not realistic.
Only after proper autovacuum tuning and under the effects of bgwriter (also tuned!) and checkpoints (all required, especially for write intensive workloads), over much larger periods of time, the benchmark could be considered meaningful.
> We next measure the scalability of Postgres-backed queues.
I'd recommend benchmarking the recently announced PgQue project, that operates bloat-free (one of the largest operational hurdles in queue-like Postgres workloads). See [2] for a previous discussion in HN.
> Based on the shown graph, this is misleading at best, essentially false. After 120K writes/s p50 spikes from 10ms to 1s (1 second for a write!!!!). That's two orders of magnitude latency spike, and an unacceptable one for an OLTP workload. It clearly shows the server is completely saturated, which is clearly a non operational regime. Quoting 144K is equivalent to quoting the throughput of a highway at the moment traffic comes to a standstill.
> Based on this graph the highest number I'd quote is 120K. And probably you want to keep operating the server within a safe margin below peak, but since this is a benchmark, let's call 120K the peak. Because actually p50 is not even the clear-cut. It should be a higher percentile (say p95) at which latency is within reasonable bounds. But for the shake of not over complicating, it could be taken as a reference.
You definitely don't want to run a production system at saturation! But it's worthwhile to measure a complex system like Postgres at saturation, see when it gets there and how it behaves there, and then run at a slightly lower throughput.
> Therefore, you are not measuring Postgres peak performance, but rather Postgres performance under the IO constraints of this particular system. Certainly, 120K IOPS is the maximum that this particular instance can have. But it doesn't show if Postgres could do better under a more performant IO disk. Actually, a good test would have been to try the next instance (db.m7i.48xlarge) with 240K IOPS and see if performance doubles (within the same envelop of p50 latency) or not. And afterwards to test on an instance with local NVMe (you won't find this in RDS).
I've done some testing (not in the blog post)--doubling instance size/IOPS doesn't improve performance significantly because it doesn't affect the WAL bottleneck. Local NVMe should have a significant impact in theory, but I haven't tested this myself.
> 300 seconds test duration?? This is not operational. You are not accounting for checkpoints, background writer, and especially autovacuum. Given that workflow pattern includes UPDATEs, you must validate bloat generation (or, equivalently, bloat removal) by a) observing much longer periods of time (e.g. 1h) and b) making sure the autovacuum configuration (and/or individual table vacuum configuration if required) makes bloat contained in a stable way. Otherwise, shown performance numbers will degrade over time, making them not realistic.
Those are usage examples (notice the 1000 rps)--actual benchmarks were run at and were stable at much longer duration.
> You definitely don't want to run a production system at saturation! But it's worthwhile to measure a complex system like Postgres at saturation, see when it gets there and how it behaves there, and then run at a slightly lower throughput.
I disagree. It's worthless a number at saturation. Because "a slightly lower throughput" is at best an unqualified hand-waving. Real numbers can be quite far from that saturation point.
Quote instead real production numbers. You can define them clearly, it's not that hard. E.g.: p95 below 10ms latency. That's it. Measure and report that number.
> I've done some testing (not in the blog post)--doubling instance size/IOPS doesn't improve performance significantly because it doesn't affect the WAL bottleneck. Local NVMe should have a significant impact in theory, but I haven't tested this myself.
But those would be interesting numbers to share! "Doesn't improve performance significantly" --sorry, I'm not big friend of unqualified data points. Is it 10%, 20%, 50%? And definitely, when measured at saturation, surely you don't see improvements. But if measured at an operational regime, you should probably see notable improvements (unless other scaling factors start to dominate, in which case your benchmark becomes much more meaningful because then you are finding Postgres scaling limits and not just the limits of the disk on which it's running). Changes the picture dramatically.
> Those are usage examples (notice the 1000 rps)--actual benchmarks were run at and were stable at much longer duration.
Sorry, but if you use that as an example, gives me little confidence about the real intent. But glad to hear you run at longer duration --add that information to the post! But again, that's not enough. Show the bloat and demonstrate how stable it is, given the tuning required to keep it contained, of course. Also show the tps over time --I'm sure it drops notably in the presence of checkpoints-- and then the "under 10ms latency at p95" will become dominated by write performance during checkpoints.
Because when you determine your SLOs, it's not at the happy path, but the opposite. And saying "Postgres can do 144K writes/sec on this machine" is beyond the happy path, so it's not meaningful for me.
People scale postgres by horizontally sharding it at the app-tier, and end up reducing its query capabilities to a glorified K/V store... They do this routinely and think it's normal. Then they use other things like Cassandra or Cosmos DB to augment the scale gaps that this leaves out (see what OpenAI is doing). Then they typically run a search engine and a cache as well. People think all of this complexity is normal/unavoidable because their cloud providers will happily see them consume all these different services.
Write QPS isn't what you want to measure, you should look at index fragmentation, dead tuple usage etc etc. All of those generally go in directions that you don't want.
"Overall, we find a Postgres server can handle up to 144K of these writes per second. That’s a lot, equivalent to 12 billion writes per day."
Based on a problem I'm facing with Postgres today, I wonder if this really progresses as linearly as the article wants to make it out.
We're in the middle of evaluating Postgres as a replacement for MySQL, and experience notable slow-down for plain multi-row inserts due to index growth as soon as the table reaches just a couple of dozen million rows. It's an uncomplicated and flat (no constraints or foreign keys etc.) medium width table of about 10-15 columns and a handful of non-composite btree indices - and/or hash indices; we've tried mixing and matching just to see what happens - but ingestion drops to less than half already before 50m rows. At 100m rows the insertion performance is down to a fraction and from there it just gets worse the larger the table and its indices grow. It's as if there's some specific exponential cut-off point where everything goes awry. However, if we simply remove all indices from the table, Postgres will happily insert hundreds of millions rows at a steady and near identical pace from start to end. The exact same table and indices on MySQL, as closely as we can match between MySQL and Postgres, running on the same OS and hardware, maintains more or less linear insertion performance well beyond 500m rows.
Now, there's a lot to say about the whys and why-nots when it comes to keeping tables of this size in an RDBMS and application design relying on it to work out, and probably a fair amount more about tuning Postgres' config, but we're stumped as to why PG's indexing performance falters this early when contrasted against InnoDB/MySQL. 50-100m rows really isn't much. Would greatly appreciate if anyone with insight could shed some light on it and maybe offer a few ideas to test out.
(add.: during these stress tests the hardware is nowhere close to over-encumbered, and there's consistent headroom on both memory, CPU and disk I/O)
problem is table design and write amplification. Every row insert triggers update into every index, so you get classic amplification problem.
Separate your table into Cold (with all indexes and bells and whistles) and Hot (heap table with no indexes except PK).
Insert as many rows as you want into Hot heap, and then move them in the background into cold in batches, so that index recalculation is amortized across many rows, instead of per-row.
Another poster suggested partitioning, thats the same idea: separate Hot and Cold data into partitions and keep hot partition as heap
That’s a really cool idea I had not heard before, thank you for sharing this. It also feels like the type of thing a db ought to be able to do under the hood. I wonder why this is not a config (though there’s a pg extension for everything so maybe it does exist)
As far as I can tell, Postgres is not designed with this inclination towards doing lighter work when clients are waiting and piling up maintenance work to do in background. I think the background work it does is mostly running vacuum on tables now and then.
Contrast that with ClickHouse, for example. It operates in a different niche than Postgres (OLAP instead of OLTP) – with their merge tree engine family [1] that does data deduplication in background.
There is one project of modernizing Postgres' storage engine called OrioleDB [2], but I think the company got acquihired by Supabase [3] and maybe the project has not been progressing very quickly since then.
[1] https://clickhouse.com/docs/engines/table-engines/mergetree-... [2] https://www.orioledb.com/ [3] https://supabase.com/blog/supabase-acquires-oriole
Oriole is under very heavy development since we acquired them. We cut a release just 2 weeks ago:
https://github.com/orioledb/orioledb/releases/tag/beta15
We expect it to be production ready this year
It wouldn't be atomic, and so would break transaction semantics.
If you committed a row update but didn't update the index, a subsequent query using the not yet updated index would not find the updated row correctly.
It would also only work for certain types of indexes, you couldn't do it for uniqueness constraint for example.
I do agree that in theory you could have some extension to the index declaration that covers all that, but my worry there would be that it would be non obvious and a foot gun. Doing it the way described above makes that break in semantics clear.
> If you committed a row update but didn't update the index, a subsequent query using the not yet updated index would not find the updated row correctly.
I wonder if you could make it so that queries read from both the index and the unindexed changes. It would be slightly slower but as long as the unindexed changes are kept small it might be fine.
My impression is that InnoDB (MySQL's primary storage engine) is doing something like this. We have never seen any slow-downs on adding to the data set I've discussed in this thread, even at hundreds of millions of rows, and per the nature of the system creating this data the majority of these rows are targeted for additional single-row DML within a few seconds of being inserted, with instantaneous effect.
> problem is table design and write amplification. Every row insert triggers update into every index, so you get classic amplification problem.
Yes, this is understood. In particular for b-trees that require some refurnishing when growing. What's less understood is why Postgres hasn't solved this in a way similar to how InnoDB solves it behind the scenes.
You've given us some idea of the volume of your data but there's no mention of what's ingesting it or how.
> during these stress tests the hardware is nowhere close to over-encumbered, and there's consistent headroom on both memory, CPU and disk I/O
This assertion is likely wrong - you're likely skipping over some metrics that has clues to what we need to know. Here are some questions to get the discussion moving.
- Is this PostgreSQL managed or self-hosted?
Your mention of "consistent headroom on both memory, CPU and disk I/O" gives me hope you're self-hosting it but I've heard the same thing in the past from people attempting to use RDS and wondering the same as you are, so no assumptions.
- Are you using COPY or multi-row INSERT statements?
- How much RAM does that server have?
- What is the fillfactor, max_wal_size and checkpoint_timeout?
- Is the WAL on NVMe?
- What's the iostat or wa during the slowdown?
- Are random UUIDs (part of) the index?
Have you posted to https://dba.stackexchange.com/
If I were you, I would create a GitHub repo that has scripts that synthesize the data and reproduce the issues you're seeing.
> This assertion is likely wrong
We can clearly see with iostat/iotop that the server and its storage isn't overworked.
> Is this PostgreSQL managed or self-hosted?
We're evaluating on our own bare metal. It's an 8c/16t Zen 3 with 32 GB of RAM. Storage (where everything incl. the WAL is) is NVMe and the drives' true sustainable/synchronous write speed is about ~750 MiB/second.
The specs are far, far higher than required for something this basic. Total memory usage when we begin observing the problem is a fraction of what's available. The MySQL production and test environments running this without hindrances actually have only half the amount of RAM.
> Are you using COPY or multi-row INSERT statements?
Multi-row INSERTs, one per transaction, with anything from 100 to 500 rows. Evaluation simulates the volumes we can see through our APIs on production, though it omits everything but Postgres itself in order to test Postgres' ingestion capability without other factors getting in the way - it's just Postgres and a light-weight data generator mimicking production data.
> What is the fillfactor, max_wal_size and checkpoint_timeout?
Not sure about the fill factor. Everything is running on default Postgres 17.9 as packaged for Debian 13, so that would be 1GB max_wal_size and checkpoint_timeout is commented out with a default of 5 minutes. Haven't gotten to any performance tuning yet, would be thankful for any suggestions to try out. The only thing we've tried is disabling auto-vacuuming to see if it was too frequent, causing i/o contention or otherwise hogging throughput. Not really any noticable change.
> Are random UUIDs (part of) the index?
No, neither v4 UUIDs nor b-tree-friendlier v7 UUIDs, but a couple of the indexed columns contain random integral numbers that can become quite sprawly inside a b-tree. We do observe somewhat better results indexing these with HASH indices instead, which also makes a lot more sense for that particular data and how we query it. For evaluation our outset was to stick with b-tree indices because that's what's used on the MySQL setup (InnoDB does not support on-disk hash indices).
> What's the iostat or wa during the slowdown?
When we reach the point where ingestion speed has shrunk to about a third we observe iowait peaks having grown to some 15%, which tells us the problem is likely Postgres spending more and more of its time shoveling in the indices rather than storing actual row data. Maximum written data at about 150 MB/second is just a fraction of what the NVMe drives can sustain. None of Postgres' individual processes ever top out anywhere close to 100% of a single core on the machine. Total memory usage is less than 2 GB, and here we suspect we have a lot of tuning to look into. To contrast, the MySQL setup is greedier with both CPU and memory usage, and an educated guess is that the major difference allowing it to keep ingesting hundreds of millions of rows without slowing down is that InnoDB, without us having to jump through any hoops, on one hand defers persisting new index data so that the DML can be finalized as quickly as possible, and on the other hand operates on its indices in a much more efficient way than Postgres.
Everything needed to reproduce the problem is in this paste, which contains a neutral version of the full table and indices, and a simplified version of the data generator: https://paste.debian.net/plainh/ddc819cb
That's a phenomenal script that absolutely belongs on a dba.stackexchange question and even the PostgreSQL mailing list.
I strongly encourage you to post to https://dba.stackexchange.com/, as a HN thread is the wrong place for this discussion (there's a lot of tuning ahead of you and others who are in your situation in the future might skip this nested thread) but be forewarned, the TLDR from dba.stackexchange will either be a quick "you need atleast 64 GB RAM for your PostgreSQL usecase" or there will a better, thorough discussion of increasing max_wal_size and lowering the fillfactor (which is what you're looking for). The ideal answer should even walk you through BRIN (vs btree) indexes. I'm asking you to post there because it will enumerate the WHY of all of these. At that point you can make an informed decision if this all would be worth it.
Now I empathize that MySQL is doing a phenomenal job at only a quarter of that (16 GB you said?) but not for the reasons you might think (and one can certainly argue, nor should an end user care!). MySQL's method of buffering (InnoDB change buffers) and its clustered index gives you the performance you like when suddenly doing bursty writes aka "write sprint". I need you to be aware of that (and dba.stackexchange responses will certainly address that).
I would have written a lengthy post on what to do next but I must first ask:
- Why are you evaluating PostgreSQL in this case when MySQL seems to work well? For example: does it feel like your aggregations are getting slower? As you can see, with PostgreSQL, you will have different set of tradeoffs (RAM, tuning, VACUUM)
- Are there real, limiting business constraints that force you to operate on less than 64 GB of RAM given your volume and throughput expectations (like FF limitations, or these are smaller machines on the edge, etc)
- If you can, as an experiment, while you write your dba.stackexchange question and for the PostgreSQL mailing list, you can tweak multiple parameters and tell me what you see:
- I'm concerned you have a `shared_buffers = 128MB`. Set it to `shared_buffers = 8GB` (give the B-trees room to live in RAM)
- Increase `max_wal_size = 16GB` (stop the checkpoint flooding and let Postgres "breathe" during your batch inserts).
- Increase `checkpoint_timeout = 30min` (set 30min to the actual window it takes; also, this is temporary but this should push checkpoints out so they don't interrupt your "write sprint").
- Set `maintenance_work_mem = 2GB` (should speed up index creation and vacuuming).
- Lower `fillfactor` on those specific 8 indexes from the default 100 down to 70. The B-trees should now have the ability to absorb those inserts better.
This should get you in the right direction, googling for the right documentation, but there's even more ahead of you, including a separate discussion about BRIN (vs btree) indexes.
> defers persisting new index data so that the DML can be finalized as quickly as possible
"defers persisting" might get misread as if it doesn't write to disk - it does but you're close and you will uncover more :)
MySQL defaults are specifically tuned for your "write sprint" usecase, infact, to actually mask the IO latency of secondary index updates but if your real usecase is not just large sudden bursts of writes to a table that has a btree index, you certainly will appreciate this effort. Happy weekend!
Thanks a bunch for all of the config tuning advice. Can see huge improvements in both flattening the creeping slow-down and really pushing back the point where we eventually hit the wall. Just lowering the fill factor on the indices made a substantial difference even before permitting Postgres to use more RAM. Using BRIN indices for certain columns, as suggested by several others in the discussion, also resulted in some smaller but still valuable speed-ups. It's obvious by now that a careful combination of curated choices greatly matters for the full scope of this data. On the whole, in fact, it looks like all of the combined advice clear this hurdle up so much that we can no longer say we observe Postgres having an actual problem dealing with our current volumes and intensity of ingestion.
We run quite a few of these setups, so cost absolutely plays a big part in the size and selection of hardware. We're evaluating Postgres because of a mergers and acquisitions event prompting us to consolidate portions of the platforms for the sake of savings and reducing the overall "sprawl". Conversely we're also testing how the other party would fare if some of their stuff were to run under MySQL instead of Postgres. There's a chance we end up leaving things just as they are.
> On the whole, in fact, it looks like all of the combined advice clear this hurdle up so much that we can no longer say we observe Postgres having an actual problem dealing with our current volumes and intensity of ingestion.
I am extremely happy for you!
This is fantastic news. I would really appreciate a followup post on DB SE and/or PostgreSQL ML because I assure you, there are others in your position. A deeply nested HN thread is not something that will be visible - you would be saving a lot of people heartburn.
Also remember, we just got started! The tips I gave you earlier is just a drop i the bucket. In the days of AWS RDS, knowledge about these are a "lost art" but running PostgreSQL yourself is extremely empowering. If you could make the posts, please do and I also urge you to test out https://github.com/pgsty/pigsty
With some extra admin work, you can greatly increase your insert throughput, as long as the table load is comprised mostly of inserts: 1. Partition your table by range of a monotonic ID or timestamp. Notice the primary key will have to contain this column. A BIGINT id column should work fine; 2. Remove all the other indexes from the partitioned table. Add them to all the partitions, except the latest one. This way, the latest one can endure a tough write load, while the other ones work fine for reads; 3. Create an admin routine (perhaps with pg_cron) to create a new partition whenever the newest one is getting close to the limit. When the load moves to the newer partition, add indexes concurrently to the old one; 4. You'll notice the newest partition will the optimized for writes but not reads. You can offset some of that by replacing BTREE secondary indexes with BRIN [1], particularly the one with bloom operator (not to be confused with Postgres Bloom regular indexes [2]). BRIN is a family of indexes more optimized for writes than reads. If the partition is not too large, it shouldn't be too bad to read from it. 5. Later you can merge partitions to avoid having too many of them. Postgres has commands for that, but I think they lock the whole table, so a safer bet is to copy small partitions into a new larger one and swap them manually.
[1] https://www.postgresql.org/docs/current/brin.html [2] https://www.postgresql.org/docs/current/bloom.html
I like this strategy a lot, but the performance of read queries suffer if they span partitions, correct?
The issue I'm facing is a very large table, that is both write and read heavy, and the reads do not fall into a specific range of values for any particular column, so I don't think partitioning is an option.
Yes, partitioning will decrease a bit the read performance of queries not correlated with the partition key. That's why you need to periodically merge smaller partitions, so that you can keep the overall partition count bounded.
It is a lot of admin work, but if you really need to scale up Postgres write throughput, I don't see many other options without increasing hardware costs.
I assume you have already picked the low-hanging fruit discussed in the neighboring comments - batch writes, make sure you are using COPY instead of INSERT, tune Postgres parameters adequately and use the fastest disk you can grab for the WAL.
Partitioning is not all that expensive. It is definitely worth testing for your specific workload. We use TimescaleDB, which relies heavily on postgres partitions, have a bit under 100 million rows in our active set (last 90 days), across 120 partitions (device*time), and it works nicely. Over 100 partitions is probably a bit many for this workload, but since it works OK we have not changed it.
These are good suggestions but I'm apprehensive they might come back and say they have 64 GB (or less) of RAM or they are using PostgreSQL RDS on AWS or something.
I asked them for specifics.
I don't think it really matters in terms of their question though, given MySql on the same specs doesn't have the problem and postgres does. Quite clearly it has something to do with indexes and what is the wall postgres is running into that causes the drop off on quite low amounts of rows. If the answer is just get more RAM, it kind of implies postgres is not really that scalable. Especially if the drop off is proportional to the number of rows.
Why are you using hash indexes? They're much less widely used than standard B-Tree indexes. The bucket split code likely isn't very scalable [1].
I suggest testing the same workload with your existing hash indexes replaced with equivalent B-Trees.
[1] https://github.com/postgres/postgres/blob/master/src/backend...
Last time I almost used a hash index in Postgres, I learned it was an incomplete feature and not crash-safe yet. This was v9.3? At that same time, MySQL had them and they were ok to use.
Later that got fixed, but I haven't tried again since, just been using btree because it seemed like Postgres favored that and it has theoretical advantages too.
They are fully stable and perform very well in Postgres today. There are some caveats, but they don't result in any sort of hiccups or unpredictable behavior.
Yeah that's what I've read too. Just haven't gotten around to trying them.
You've probably already read the Postgres docs on hash indexes, but just in case, it says "hash indexes may not be suitable for tables with rapidly increasing number of rows." I agree with the other commenter that it's worth at least trying without them if you haven't already, even though you're already VACUUMing.
> Why are you using hash indexes?
On some data, for certain ways of using and querying that data, it makes a lot more sense than a b-tree. When we use hash indices for some of the columns in our test just to see how Postgres will perform, and run our test towards a quarter of a billion rows where Postgres' 32-bit hash indices have a relatively high chance of colliding, querying these hash indices and returning the single requested row is still instantaneous.
Thanks for sharing this "tactic". Will look into it.
The problem is row locks when using interactive transactions over the network and contention. That can absolutely kill your performance with postgres, there's not really anything you can do to get around it (other than avoid interactive transactions). [1]
[1] - https://andersmurphy.com/2025/12/02/100000-tps-over-a-billio...
We had an interesting architecture situation at work. Puppet Enterprise uses a single Postgres server. The company had moved from a recommendation of using a single PuppetDB API node (which fell over at high load) to running a PuppetDB API server on each compiler node.
That, however, came with its own set of problems. Of course you have to tune for concurrent connections as you scale wider, but there were much more serious contention issues than you'd expect, and the compilation times were terrible too. It turned out to be because those transactions locked the DB during their (synchronous) operations, and we had a globally distributed set of compilers in order to serve globally distributed traffic.
The solution ended up being to run a separate cluster of API servers in the same region as the DB. The expensive calls from the compilers to the API servers were largely async https so they didn't have to wait on the API nodes, and the API nodes could talk to the DB synchronously with low latency.
If you find yourself doing a lot of explicit transactions, it can be a sign that your schema isn't as normalized as it should be.
How does better normalization reduce the need for explicit transactions?
The need to change two separate tables atomically means that you probably have info duplicated. Also it can be non-ideal in a web backend cause it means keeping DB connections open longer.
Sorry ignore the second sentence, I meant keeps the connection checked out of the pool longer, but also it has nothing to do with normalization.
In this case that is not the problem.
What's the underlying filesystem(s) you're using for the data storage?
Ext4 with journaling. I've played around with XFS as well just to see what would happen. No major difference. I'm certain it's caused by the way Postgres builds its indices.
No worries. Just checking if it was ZFS.
While I really like ZFS personally for 99% of things, for "need maximum performance with PostgreSQL and we're already pushing high end hardware to the maximum" ZFS is reportedly the wrong choice.
We've reached the same conclusion in other situations. Great for reliable and scalable storage, not so great for focused intense i/o.
So you open transaction, insert multiple rows, commit? And you're not using any special xact settings like SERIALIZABLE mode, right? Normally you use COPY FROM if it's a huge number of rows in some batch process, but not like inserting a few rows handling a backend request or something.
I would try filling up the table, forcing a vacuum, then timing the inserts afterwards. Not that you should need to vacuum in real usage, but it might uncover a problem.
> So you open transaction, insert multiple rows, commit?
Pretty much, yes. High volume rapid multi-row INSERTs. Nothing special involved in the DDL or how the data goes into the table. Vacuuming the table (whether automatically or between every 10-20m rows) makes no difference. The indices still penalize the performance substantially the larger they grow.
if the index updates are your bottleneck you can often get away with using a much lighter index (for example BRIN) at a cost of slightly slower queries. This is very often a great and much overlooked tradeoff.
Using BRIN for certain columns resulted in an improvement. Thanks for the suggestion.
Just don't bother with Postgres.
I've run mysql for years in production and have spent probably 30 minutes thinking about managing it. Unless there's some psql feature you need (unlikely), it'll just become a severe pain in your ass down the road because you set something up "wrong."
Just vacuuming can barf completely, leaving you dead in the water.
TL;DR: if you aren't a DBA and don't want to play one on TV don't bother with psql.
Dunno, I'm not a DBA but find Postgres to be fine.
Agreed. And, any time I ask for help with something DB-related, people just chastise me for not running postgres and blindly tell me that's my problem. That sort of community is not what I want to look forward to.
For what it's worth, I would never say that, even though Postgres is my typical choice. MySQL is fine too.
Deadlocks?
It scales beyond the needs that most people have in most situations.
The constant problem is that "big scale" always means "larger than I've seen", so on any project larger than a person has encountered, they assume they need to pull out the big guns. Also, people worry about things like what happens if they really *do* scale 10 years from now.
Neither is a practical concern for nearly anyone who will ever face this decision.
And then yes, of course, some people have problems that actually can't be solved by Postgres. But verify this first, don't assume.
What gets me is that some people seem to ignore the very real cliff of complexity that ramps up the moment you move to eventual consistency. If you need it you need it, but you have to bake in those assumptions everywhere - and they commonly break the default assumptions of those who don't have a bunch of experience with it or haven't architected their approach to work around those.
And in many cases it's those architectures that force more complexity and make it appear like they have much bigger challenges then they do. Great for resume driven development, but often you can get away with far less.
The other thing is that now a days you scale way way further vertically before you scale horizontally (assuming you are not using a cloud provider)
Everyone is hung up making their shit "scalable" like its a systems design interview at google in 2010.
Now a days you get a box with 600+ cores and 4TB of RAM. That is going to cover a very very large percentage of most enterprises.
When you scale horizontally from day one, it usually gives fault tolerance for online service, and this story is not very friendly in case of vertically scaled PG.
Scaling vertically does not mean you cant have fault tolerance and I feel like your comment just makes my point.
My point is that in PG, fault tolerance requires untrivial investments.
If you pick horizontally scalable DB (foundationdb, cocroachdb, scylladb, tidb, etc) they all give you fault tolerance for free without much involvement from your side, because it is part of the nature of those DBs.
DBOS is amazing when it comes to Durable Workflows. There are others in the space - the most popular one being Temporal but I argue, Temporal is also the most complicated one. I often say Temporal is like Kubernetes while DBOS is like `docker compose`. (and for those taking me literally, you can use DBOS in Kubernetes!)
I don't realize why DBOS is not nearly as popular as Temporal but it has made a world of difference building Durable Queues and Long Running, Durable Workflows in Python (it supports other languages too).
As they show in this article, Postgres scales impressively well (4 billion workflows per day, on a db.m7i.24xlarge, enough for most applications), which is why, if you have your PostgreSQL backup/restore strategy knocked out and dialed in, you should really take a close look at DBOS to handle your cloud agnostic or self hosted Durable Queues and Durable Workflows. It's an amazing piece of software founded by the original author of Ingres (precusor to Postgres - the story of DBOS itself is captivating. I believe it started from being unable to scale Spark job scheduling)
The reason that DBOS isn't as popular is because it's younger. DBOS launched in the form we know it in 2024. Temporal is much older; Temporal is technically a fork of Cadence and Cadence released originally in 2017, with Temporal forking and releasing back in 2020. When all three are trying to be "the same sort of thing" and that thing is new, it's hard to show up 7-8 years after the trailblazers and say "oh yeah, we're clearly better" when the existing thing works and is used by tons of folks.
Temporal is a dumpster fire, they've gotten so much VC funding (recently had D, 300M at a 5bn valuation) with ... nothing to build except ways to trap customers into their SAAS.
I give them about a year or two before the wheels fall off, then it's off to Broadcom and friends.
But I could be wrong as now they're not in the 'durable execution' space at all, it's 'durable execution for ai' according to their latest conference.
Got to spend that VC dosh somewhere I suppose, they're certainly not spending it on making a good product.
Temporal employee here. I'm very surprised by your comment.
It's true that we recently had a Series D and that VC firms recognize the value of what we do. The Temporal Server software is 100% open source (MIT license: https://github.com/temporalio/temporal/blob/main/LICENSE). It's totally free and you don't even need to fill out a registration form, just download precompiled binaries from GitHub or clone the repo and build it yourself. You can self-host it anywhere you like, no restrictions on scale or commercial usage. We offer SaaS (Temporal Cloud), which customers can choose as an alternative self-hosting, based on their needs. The migration path is bi-directional, so not a trap by any definition.
Regarding AI, Temporal is widely used in that space, but that does not negate the thousands of other companies that use Temporal for other things (e.g., order management systems, customer onboarding, loan origination, money movement, cloud infrastructure management, and so on). In fact, our growth in the AI market came about because companies who were already using Temporal for other use cases realized that it also solved the problems they encountered in their AI projects.
And to your last point, we've made dozens of enhancements to the product (here's a small sample: https://temporal.io/blog/categories/product-news). I'd encourage you to follow the news from next week's Replay conference (https://replay.temporal.io/) because we'll be announcing many more.
Classic temporalio, reply to criticism with an advert. <3
Maybe. But as someone who happily self hosting pretty big Temporal workloads for my day job (I inherited it from early adopters circa 2022), it definitely does not feel like a dumpster fire. It chugs along unglamorously and I enjoy working on it.
I think we have wildly different definitions of pretty big.
[dead]
DBOS looks simple (good), but from the docs below, executor elasticity appears to be locked behind license purchase. So it truly is like docker compose, good and bad parts?
https://docs.dbos.dev/production/workflow-recovery#recovery-...
>When self-hosting in a distributed setting without Conductor, it is important to manage workflow recovery so that when an executor crashes, restarts, or is shut down, its workflows are recovered. You should assign each executor running a DBOS application an executor ID through DBOS configuration. Each workflow is tagged with the ID of the executor that started it. When an application with an executor ID restarts, it only recovers pending workflows assigned to that executor ID.
https://docs.dbos.dev/production/hosting-conductor
> Self-hosted Conductor is released under a proprietary license. Self-hosting Conductor for commercial or production use requires a paid license key.
This is a good question! No, it's not like docker compose (I imagine you implied the swarm and hub pull limits?)
DBOS Conductor is an out of band management service that IIRC helps you mainly observe your DBOS and recover failures in a seamless way. As far as I could see, it's not necessary, for you to use DBOS workflows and queues. Don't quote me though and reach out to their forum and verify in case I'm missing certain usecases.
Personally, I do not use DBOS Conductor - I have my own observable setup using Grafana/VictioriaMetrics as my workflows are instrumented with OTel. I had initially set Conductor up for development (and it looked to be free for development although I recall some major limitations on how many workflows etc - which is why I put my own alternate monitoring setup).
They also have a very reasonably priced cloud hosted DBOS Conductor. I think my first 30 days were completely free and then they moved me to a "hobby" tier. It's a fantastic way to help decide whether it's for you.
I believe DBOS Conductor is how DBOS pays the bills but you can use DBOS workflows and queues unlimited without DBOS Conductor. If you don't want to pay for Conductor - their out of band management service, you can put together your own just fine, like I did. My own Grafana/VictioriaMetrics setup answers my questions but I would imagine Claude/Codex/Cursor should be able to put something fairly useful if you didn't want to go down my route.
> executor elasticity appears to be locked behind license purchase
DBOS has designed their system to be extremely flexible and extensible. While yes, Conductor can absolutely manage your executors for you, it's not the only thing that can. You're not limited to using Conductor. As I said, I manage my own - everything you need to know to do so is in the code and documentation. They even have a document for LLMs and agents. I have had to interact with the DBOS team 0 times to set everything up.
I prefer this business model (an optional tool - Conductor, is paid) vs. DBOS offering just everything across the stack on a "free tier" but with caps on DBOS workflows and queues. In their current business model DBOS workflows and queues are completely uncapped (atleast from what I can make out).
If you do reach out to them, I would appreciate if you let me know anything to the contrary.
Are you using Temporal with distributed workers?
We have a simple worker setup and temporal is pretty easy to setup
Out only issue is really needing an intermediary data store for task result storage
We are using DBOS in new projects as it's even simpler and the downside (task log interface behind saas) is easily remedied with a copilot generated task viewer
Postgres can scale, just like any database can scale.
The issue is there's a lot of lore and esoterica required to get it to scale.
Every time there's an issue with Postgres there's someone that's all "just do xyzzy, it's super obvious that this undocumented setting fixes it."
Just look at the slowdown/big pages problem.
If it works for you, use it.
> undocumented.
There is a steep learning curve sure, and there is lot of manual steps to do some critical actions perhaps.
Documentation however is the last thing I would criticize PostgreSQL for. PostgreSQL documentation is by far best amongst all databases open or closed. It is exhaustive, well maintained, well written, and quite accurate and kept up for decades - The oldest supported version (in documentation) is 6.3 released in 1998.
I thought this was a fun article from a couple months back: https://openai.com/index/scaling-postgresql/
Yes, you can scale it quite well vertically.
But how about horizontally? It would be nice to have high availability, or even to be able to upgrade the OS and postgres itself without downtime.
Shameless plug[0].
[0] https://pgdog.dev
AFAIK that's what Multigres[0] and Neki[1] are trying to solve.
[0] https://multigres.com/ [1] https://neki.dev/
Only played around with it but you can use patroni, etcd and HAproxy to achieve this. It’s a pain, but I believe there was some kind of coolify-style open source application to do this for you but I can’t for the life of me remember its name
You might be thinking of Pigsty?
Atleast I hope you are! Nothing else has been as well battletested. Unfortunately, perhaps because if its name, it gets no facetime on HN. Its last few mentions here barely received attention it deserved.
https://github.com/pgsty/pigsty
https://pigsty.io/
autobase[1] is the one I can think of
[1] https://github.com/autobase-tech/autobase
Yep, this is what I think about when “scaling” is mentioned. Maybe I’m too distributed-compute brained, but throwing CPU at a db isn’t what I was hoping would be the answer.
So the point of distributed compute is to reduce the compute needed? I’ve generally found that distributed compute requires more compute than vertical scaling while getting clobbered by network bandwidth / latency.
Theoretically with 2 to 10x compute required and in practice 100 to 500x
I think for databases horizontal scaling for writes only makes sense once vertical scaling stops working. It comes with high complexity, annoying limitations, and often higher cost.
Horizontal scaling for reads on the other hand is much easier. If you have multiple replicas for high availability, you might as well put them to work. It can also reduce the risk for read heavy tasks interfering with transaction processing. You can even go a step further and replicate to a database that's optimized for analytical tasks.
Horizontal scaling for stateless applications (e.g. web servers or job processors) is often easier and more robust than vertical scaling, with little to no downsides.
The point of distributed computing is to do computing that you can't do on a vertically scaled system or to increase availability.
If you're doing it for other reasons it's usually a mistake.
The advice I’ve gotten is that you want to move computation to data that is already distributed. The cost of moving large amounts of data usually dwarfs compute costs (usually, not always), and so the performance win comes from distributing the computation and then (depending on the problem) centralizing aggregate results.
Another pretty good reason to do distributed computing is to move the computation closer to where the data is or where the data will be consumed.
Practically trivial to do in 2026 even by hand, and there are a couple of ready to use solutions that even make it automated.
If you're running it in kubernetes with cloudnativepg it's even easier.
The only thing it doesn't do well is master master replication which is why most of these does it scale posts mostly talk about how slow writes are. And they are pretty slow.
It all depends on the storage underneath. If you have got good storage, with CNPG you have comparable results you get on bare metal PostgreSQL.
See for example: https://infohub.delltechnologies.com/en-us/t/edb-postgres-ai...
Only reads scale. You get (much) worse writes for sure.
This post conflates scalability and performance. PostgreSQL is fast on smallish systems, but try adding more CPU cores and you'll see performance gains will not be linear at all. Modern server can ship with 256 or more cores and a single instance of PostgreSQL will struggle to take advantage of these.
4-8 cores is no problem at all, though.
What kind of cases were you measuring? I would think that, e.g. 256 separate long-lived connections in a setup like that would scale less-than-linearly but not dramatically so?
> we find a Postgres server can handle up to 144K of these writes per second. That’s a lot, equivalent to 12 billion writes per day.
Based on the shown graph, this is misleading at best, essentially false. After 120K writes/s p50 spikes from 10ms to 1s (1 second for a write!!!!). That's two orders of magnitude latency spike, and an unacceptable one for an OLTP workload. It clearly shows the server is completely saturated, which is clearly a non operational regime. Quoting 144K is equivalent to quoting the throughput of a highway at the moment traffic comes to a standstill.
Based on this graph the highest number I'd quote is 120K. And probably you want to keep operating the server within a safe margin below peak, but since this is a benchmark, let's call 120K the peak. Because actually p50 is not even the clear-cut. It should be a higher percentile (say p95) at which latency is within reasonable bounds. But for the shake of not over complicating, it could be taken as a reference.
> We found that the bottleneck was in flushing the Postgres write-ahead log (WAL) to disk
Therefore, you are not measuring Postgres peak performance, but rather Postgres performance under the IO constraints of this particular system. Certainly, 120K IOPS is the maximum that this particular instance can have. But it doesn't show if Postgres could do better under a more performant IO disk. Actually, a good test would have been to try the next instance (db.m7i.48xlarge) with 240K IOPS and see if performance doubles (within the same envelop of p50 latency) or not. And afterwards to test on an instance with local NVMe (you won't find this in RDS).
> From [1]: > Postgres insert throughput > uv run python benchmarks/postgres_insert.py --rps 1000 --duration 300
300 seconds test duration?? This is not operational. You are not accounting for checkpoints, background writer, and especially autovacuum. Given that workflow pattern includes UPDATEs, you must validate bloat generation (or, equivalently, bloat removal) by a) observing much longer periods of time (e.g. 1h) and b) making sure the autovacuum configuration (and/or individual table vacuum configuration if required) makes bloat contained in a stable way. Otherwise, shown performance numbers will degrade over time, making them not realistic.
Only after proper autovacuum tuning and under the effects of bgwriter (also tuned!) and checkpoints (all required, especially for write intensive workloads), over much larger periods of time, the benchmark could be considered meaningful.
> We next measure the scalability of Postgres-backed queues.
I'd recommend benchmarking the recently announced PgQue project, that operates bloat-free (one of the largest operational hurdles in queue-like Postgres workloads). See [2] for a previous discussion in HN.
[1]: https://github.com/dbos-inc/dbos-postgres-benchmark [2]: https://news.ycombinator.com/item?id=47817349
> Based on the shown graph, this is misleading at best, essentially false. After 120K writes/s p50 spikes from 10ms to 1s (1 second for a write!!!!). That's two orders of magnitude latency spike, and an unacceptable one for an OLTP workload. It clearly shows the server is completely saturated, which is clearly a non operational regime. Quoting 144K is equivalent to quoting the throughput of a highway at the moment traffic comes to a standstill.
> Based on this graph the highest number I'd quote is 120K. And probably you want to keep operating the server within a safe margin below peak, but since this is a benchmark, let's call 120K the peak. Because actually p50 is not even the clear-cut. It should be a higher percentile (say p95) at which latency is within reasonable bounds. But for the shake of not over complicating, it could be taken as a reference.
You definitely don't want to run a production system at saturation! But it's worthwhile to measure a complex system like Postgres at saturation, see when it gets there and how it behaves there, and then run at a slightly lower throughput.
> Therefore, you are not measuring Postgres peak performance, but rather Postgres performance under the IO constraints of this particular system. Certainly, 120K IOPS is the maximum that this particular instance can have. But it doesn't show if Postgres could do better under a more performant IO disk. Actually, a good test would have been to try the next instance (db.m7i.48xlarge) with 240K IOPS and see if performance doubles (within the same envelop of p50 latency) or not. And afterwards to test on an instance with local NVMe (you won't find this in RDS).
I've done some testing (not in the blog post)--doubling instance size/IOPS doesn't improve performance significantly because it doesn't affect the WAL bottleneck. Local NVMe should have a significant impact in theory, but I haven't tested this myself.
> 300 seconds test duration?? This is not operational. You are not accounting for checkpoints, background writer, and especially autovacuum. Given that workflow pattern includes UPDATEs, you must validate bloat generation (or, equivalently, bloat removal) by a) observing much longer periods of time (e.g. 1h) and b) making sure the autovacuum configuration (and/or individual table vacuum configuration if required) makes bloat contained in a stable way. Otherwise, shown performance numbers will degrade over time, making them not realistic.
Those are usage examples (notice the 1000 rps)--actual benchmarks were run at and were stable at much longer duration.
> You definitely don't want to run a production system at saturation! But it's worthwhile to measure a complex system like Postgres at saturation, see when it gets there and how it behaves there, and then run at a slightly lower throughput.
I disagree. It's worthless a number at saturation. Because "a slightly lower throughput" is at best an unqualified hand-waving. Real numbers can be quite far from that saturation point.
Quote instead real production numbers. You can define them clearly, it's not that hard. E.g.: p95 below 10ms latency. That's it. Measure and report that number.
> I've done some testing (not in the blog post)--doubling instance size/IOPS doesn't improve performance significantly because it doesn't affect the WAL bottleneck. Local NVMe should have a significant impact in theory, but I haven't tested this myself.
But those would be interesting numbers to share! "Doesn't improve performance significantly" --sorry, I'm not big friend of unqualified data points. Is it 10%, 20%, 50%? And definitely, when measured at saturation, surely you don't see improvements. But if measured at an operational regime, you should probably see notable improvements (unless other scaling factors start to dominate, in which case your benchmark becomes much more meaningful because then you are finding Postgres scaling limits and not just the limits of the disk on which it's running). Changes the picture dramatically.
> Those are usage examples (notice the 1000 rps)--actual benchmarks were run at and were stable at much longer duration.
Sorry, but if you use that as an example, gives me little confidence about the real intent. But glad to hear you run at longer duration --add that information to the post! But again, that's not enough. Show the bloat and demonstrate how stable it is, given the tuning required to keep it contained, of course. Also show the tps over time --I'm sure it drops notably in the presence of checkpoints-- and then the "under 10ms latency at p95" will become dominated by write performance during checkpoints.
Because when you determine your SLOs, it's not at the happy path, but the opposite. And saying "Postgres can do 144K writes/sec on this machine" is beyond the happy path, so it's not meaningful for me.
They can adjust their checkpoint settings to increase throughput further - https://www.postgresql.org/docs/current/wal-configuration.ht...
Yes, this benchmark deliberately uses RDS defaults to make the comparison fairer/more general.
One warning--the setting that would increase throughput the most (synchronous_commit = off) sacrifices durability to do so.
People scale postgres by horizontally sharding it at the app-tier, and end up reducing its query capabilities to a glorified K/V store... They do this routinely and think it's normal. Then they use other things like Cassandra or Cosmos DB to augment the scale gaps that this leaves out (see what OpenAI is doing). Then they typically run a search engine and a cache as well. People think all of this complexity is normal/unavoidable because their cloud providers will happily see them consume all these different services.
The implication is that scaling pg vertically should also obviate the need for a cache or search engine?
Write QPS isn't what you want to measure, you should look at index fragmentation, dead tuple usage etc etc. All of those generally go in directions that you don't want.
Of course, there are many plugins and similar things.
And Does Postrgres Backup scale?
[flagged]
[flagged]
[flagged]
[flagged]
when discussing DB it becomes so so interesting not because db itself but the people trying to ask some infeasible questions