SSDs have become fast, except in the cloud

At my job at a telco, I had a 13 billion record file to scan and index for duplicates and bad addresses.

Consultants brought in to move our apps (some of which were Excel macros, others SAS scripts running on old desktop) to Azure. The Azure architects identified Postgres as the best tool. Consultants attempted to create a Postgres index in a small Azure instance but their tests would fail without completion (they were string concatenation rather than the native indexing function).

Consultants' conclusion: file too big for Postgres.

I disputed this. Plenty of literature out there on Pg handling bigger files. The Postgres (for Windows!) instance on my Core I7 laptop with an nVME drive could index the file about an hour. As an experiment I spun up a bare metal nVME instance on a Ryzen 7600 (lowest power, 6 core) Zen 4 CPU pc with a 1TB Samsung PCIe 4 nVME drive.

Got my index in 10 minutes.

I then tried to replicate this in Azure, upping the CPUs, memory, and to the nVME Azure CPU family (Ebsv5). Even at a $2000/mo level, I could not get the Azure instance any faster than one fifth (about an hour) of the speed of my bare metal experiment. I probably could have matched it eventually with more cores, but did not want to get called on the carpet for a ten grand Azure bill.

All this happened while I was working from home (one can't spin up an experimental bare metal system at a drop-in spot in the communal workroom).

What happened next I don't know, because I left in the midst of RTO fever. I was given the option of moving 1000 miles to commute to a hub office, or retire "voluntarily with severance." I chose the latter.

This matches my experience. Something about cloud systems makes them incredibly slow compared to real hardware. Not just the disk, the CPU is more limited too.

I fire up vCPU or dedicated or bare metal in the cloud, doesn't matter, I simply cannot match the equivalent compute of real hardware and it's not even close.

As someone who works with Azure daily, I am amazed not just at the initial consultant's conclusion (that is, alas, typical of folk who do not understand database engines), but also to your struggle with NVMe storage (I have some pretty large SQLite databases on my personal projects).

You should not have needed an Ebsv5 (memory-optimised) instance. For that kind of thing, you should only have needed a D-series VM with a premium storage data disk (or, if you wanted a hypervisor-adjacent, very low latency volume, a temp volume in another SKU).

Anyway, many people fail to understand that Azure Storage works more like a SAN than a directly attached disk--when you attach a disk volume to the VM, you are actually attaching a _replica set_ of that storage that is at least three-way replicated and distributed across the datacenter to avoid data loss. You get RAID for free, if you will.

That is inherently slower than a hypervisor-adjacent (i.e., on-board) volume.

Isn't that expected? I would assume cloud stuff is slower because it's essentially an emulation of the real thing.

> Anyway, many people fail to understand that Azure Storage works more like a SAN than a directly attached disk--when you attach a disk volume to the VM, you are actually attaching a _replica set_ of that storage that is at least three-way replicated and distributed across the datacenter to avoid data loss. You get RAID for free, if you will.

I've said this a bit more sarcastically elsewhere in this thread, but basically, why would you expect people to understand this? Cloud is sold as abstracting away hardware details and giving performance SLAs billed by the hour (or minute, second, whatever). If you need to know significant details of their implementation, then you're getting to the point where you might as well buy your own hardware and save a bunch of money (which seems to be gaining some steam in a minor but noticeable cloud repatriation movement).

I may have the "Ebsv5" series code incorrect. I'd look it up, but I don't have access to the subscription any longer.

What I chose ultimately was definitely "nVME attached" and definitely pricey. The "hypervisor-adjacent, very low latency volume" was not an obvious choice.

The best performing configuration did come from me--the db admin learning Azure on the fly--and not the four Azure architects nor the half dozen consultants with Azure credentials brought onto the project.

Well, in short, people need to understand that cloud is not their computer. It is resource allocation with underlying assumptions around availability, redundancy and performance at a scale well beyond what they would experience in their own datacenter.

And they absolutely must understand this to avoid mis-designing things. Failure to do so is just bad engineering, and a LOT of time is spent educating customers on these differences.

A case in point that aligns with is that I used to work with Hadoop clusters, where you would use data replication for both redundancy and distributed processing. Moving Hadoop to Azure and maintaining conventional design rules (i.e., tripling the amount of disks) is the wrong way do do things, because it isn't required neither for redundancy nor for performance (they are both catered for by the storage resources).

(Of course there are better solutions than Hadoop these days - Spark being one that is very nice from a cloud resource perspective - but many people have nine times the storage they need allocated in their cloud Hadoop clusters because of lack of understanding...)

I would think that lifting and shifting a Hadoop setup into the cloud would be considered an anti-pattern anyway; typically you would be told to find a managed, cloud-native solution.

Ebsv5 does not have a local nvme. To get local drive you need to pick the version with a “d” in the name.

The cloud is also being sold as “don’t worry about data loss”.

To actually deliver on that promise while maintaining abstraction of just “dump your data on C:/ as you are used to”, there are compromises in performance that need to be taken. This is one of the biggest pitfalls of the cloud if you care more about performance than resiliency. Finding disks that don’t have such guarantees is still possible, just be aware of it.

You would be surprised at what corporate thinking and procurement departments actually think is best.

Why should it be expected? I'm being sold compute with quoted GHz CPU speeds, RAM and types of SSDs.

I would vaguely expect it to not match my workstation, sure, but all throughout this thread (and others) people have cited outrageous disparities i.e. 5x less performance that you'd expect even if you managed your expectations to e.g. 2x less due to the cloud compute not being a bare metal machine.

In other words, and to illustrate this with a bad example: I'd be fine paying for an i7 CPU and ending up at i5 speeds... but I'm absolutely not fine with ending up at Celeron speeds.

When I spin up a vm on my hardware and run an application. The performance is generally about 70% of what I can get in a container running on an OS on that very same bare metal.

But that isn't the delta I'm seeing, it's 5-10x performance delta not a 30-50% delta.

Ebsv5 and Ebdsv5 somewhat uniquely provide the highest possible storage performance right now in Azure, partly because they support NVMe controllers instead of SCSI.

However, the disks are still remote replicas sets as someone else mentioned. They’re not flash drives plugged into the host, despite appearances.

Something to try is (specifically) the Ebdsv5 series with the ‘d’ meaning it has local SSD cache and temp disks. Configure Postgres to use the temp disk for its scratch space and turn on read/write caching for the data disks.

You should see better performance, but still not as good as a laptop… that will have to wait for the v6 generation of VMs.

Was this on the L-series or managed disk?

Lsv3 series

Makes sense.

I tried a variety of configurations. The E-series was one of them, as it's advertised as "Great for relational database servers, medium to large caches, and in-memory analytics." Premium and Premium V2, tried those at larger capacities I didn't need just to get higher IOPS.

None came within an order of magnitude of a Ryzen 7600/nVME mobo sitting in my son's old gaming case.

An option I did not try was Ultra disk, which I recall being significantly more expensive and was not part of the standard corporate offering. I wasn't itching to get dragged in front of the architecture review board again, anyway.