💾 Archived View for dioskouroi.xyz › thread › 29397235 captured on 2021-12-04 at 18:04:22. Gemini links have been rewritten to link to archived content
⬅️ Previous capture (2021-12-03)
➡️ Next capture (2021-12-05)
-=-=-=-=-=-=-
________________________________________________________________________________
This a) is not new, and b) is not unique. The other clouds have their equivalents: Azure Stack, Google Anthos/Distributed Cloud.
For most part, with any of these you get the steep pricing of cloud with the maintenance overhead, lack of flexibility and lengthy commitment periods of on-prem, meaning they're unlikely to be a sensible option unless you have regulatory requirements that force you into it. One use case is wanting to run the same cloud stack globally, but having a market where there is no local region and local law requires that data stay in country.
>This a) is not new
This _specific_ announcement is new, because this specific announcement is about Outposts in a new, smaller form factor that just went GA today.
>One use case is wanting to run the same cloud stack globally, but having a market where there is no local region and local law requires that data stay in country.
The use case mentioned in the announcement is more about running EC2 instances in small branch offices or retail stores where you 1) still want to run AWS, 2) need the servers to be in very close proximity, and 3) don't have the room or infrastructure for a full rack.
Why do you "need the servers to be in very close proximity" in a retail store? It's not high frequency trading
Poor connectivity to the internet?
Given that AWS Outposts behave badly without connection to mothership, well...
Why do you want it close then?
Purely local latency?
https://news.ycombinator.com/item?id=29398880
summarizes it best when it comes to my answer :)
For some customers, I imagine this will start off as an emotionally-driven request...until they see the price.
My understanding is that Anthos is more like AWS K8s Everywhere (and Azure Arc), and that Google doesn't really have an Outposts style offering. Is that not the case?
They announced Google Distributed Cloud recently:
> _Google Distributed Cloud is a portfolio of fully managed hardware and software solutions which extends Google Cloud’s infrastructure and services to the edge and into your data centers._
https://cloud.google.com/distributed-cloud
Outposts is certainly not a cheap solution, a dev rack with one m5.24xl is $124K, a similarly configured Dell server (but without the rack or AWS management) is less than half that price.
https://aws.amazon.com/outposts/rack/pricing/
I didn't calculate to see if the bigger racks are more cost effective, but I'm guessing not if 5 m5.24xlarge, 4 c5.24xlarge, 3 r5.24xlarge costs $690K.
The ability to lift and shift existing AWS automation tools to on-prem workloads is more the worth the money. At a lot of organizations on-prem servers are heavily controlled by teams that have no interest in automation or quick provisioning.
Being able to stick one of these into a data center and get full access to a significant number of AWS services while being able to tell security and others that certain data isn’t in the cloud is priceless. People overlook the cost savings of being able to turn your Devs and DevOps teams loose on-prem and no longer be blocked by system admins who have no desire to help you increase your development velocity.
I have had the pleasure of building on an outpost for several months. It’s an interesting experience.
I'd be interested in knowing what drove you to Outposts in the first place -- low latency or regulatory reasons? (Or something else?)
Now you have us interested..
Please tell us more
I read an article that claims sales for Outposts have not met expectations due to software issues and their very high price.
Not exactly the same, but the anywhere stuff (ECS etc) - I've found it pretty enjoyable to use. You basically can throw some beefy local servers into the pools and away you go with your standard jobs. When you need to maintain things locally, you can (usually) migrate loads pretty directly to an AWS region.
Pricing locally is really good in many cases vs AWS and the ECS Anywhere fees themselves for example are pretty modest. $8/month range per machine (which can of course host a fair bit).
So, it’s the fog instead of the cloud.
The Gavin Belson Signature Box
So I must be missing something here. They drive to your workplace with a truck and drop off a bunch of servers and wires and stuff and get them all set up for you. Then you log into them and use the same APIs that you'd use if you connected to a normal AWS bucket? Why not just do the cloud server instead? Is this for people who have to move/store very large amounts of data such that transfer time is an issue? Or is it just a thing to let you not worry about sysadmin-type stuff because the Amazon dude does all of it for you?
This is for businesses which have regulatory requirements that they store data in-state. State governments don't like it when certain data is stored out of state. They like the idea that they can just dispatch their own police to seize the required records from the server rather than needing to ask for the cooperation of another government and its own police force.
I know specifically its popular in the regulated gambling and insurance industries.
Yes - take for example gene sequencing - that requires a huge amount of data shuffling and going cloud for that is infeasible simply for the data transfer costs.
With Outpost this becomes feasible. This "cloud in a box" idea has been around for a long time, but I guess Amazon has enough brainshare for the tooling that it may make sense for companies to have this for data-intensive compute and keep the same skillset requirements as the rest of their fully-clouded operations.
I do genome sequencing and this is very true. You are either all-in on cloud or do the processing completely in house. The data egress fees are more than enough to offset any benefit to a hybrid approach.
I’m more curious if there is an equivalent to S3 for on-prem. I bet not due to reliability constraints, but something like this would be nice. I can’t say there would be much benefit to using AWS Outposts for genome sequencing without an S3 local cache.
If you are asking if Outputs includes S3, yes,
https://aws.amazon.com/s3/outposts/
If you are asking if there are stand-alone S3 API offerings, you might want to look at MinIO. Reliability mostly requires extra hardware and configuration -
https://docs.min.io/docs/minio-erasure-code-quickstart-guide...
I didn't know they also had S3 for outposts... that's interesting. That also requires an Outposts rack as opposed to the servers the original article was talking about.
_> 380 TB S3 tier is priced at $0.1000/GB-mo ($38,912.00/month)_
But it's also extremely expensive. $38K per month for 380 TB of storage. Even if we assume 3X redundancy, you can get 500TB of raw capacity (12-16TB HDDs) for ~ $25k. After 3 months of usage, you'd have already covered that initial expense. I know there are other benefits to having managed services, but again, cost is definitely prohibitive for many use-cases.
I use minio locally to support S3-workloads on our local cluster. We can then migrate workloads to AWS/S3 proper with minimal changes, if needed.
How do you even get the petabytes of data into S3 without incurring massive charges? Previous estimations of getting the initial dataset into cloud was.. prohibitive (note this was 3-4 years ago).
What charges in particular are you referring to? The cloud services only charge data egress fees. Data ingress is free.
The ENCODE Project data were a hundred terabytes or so at the point I migrated it into S3 (it's now over a petabyte.) We were on a fast academic network with a 100GB interconnect so could just `aws s3 sync` rather than worry about using AWS Snowballs or anything.
Uploading data into AWS is free.
Regulated industries of any kind (whether legally or pragmatically), latency, security (not necessarily for what runs on Outposts, but for what Outposts _connect to_), etc. etc. while treating it essentially as another amazon region.
In the early days of the "cloud" it was common to talk about "shifting loads to cloud" as temporary scaling measure, where cloud services were not default and instead an extension of your on-premise setup - this is the opposite way around, extending cloud to on-premise.
> Why not just do the cloud server instead?
Having explored this previously in the context of this (which was mostly vapourware when I looked at it a couple of years back) and the Azure equivalent:
1. Regulatory requirements.
2. There are no network - and hence egress - charges to the rest of your infra.
3. Limited ability to operate disconnected.
On the last point, the Azure offering is vastly, vastly superior, since the AWS one used to phone home and stop after a few hours, which undercuts the value on the third point. There are quite a few places with no local cloud presence, and unreliable connectivity to the nearest one; for example, last time I was in New Caledonia, there were multi-day outages and degradation to their sole fibre to the world. They do significant research workloads (thanks to the SPC) and something that could endure that while allowing you to keep using cloud APIs would be very convenient.
The attraction is _"works like AWS but with high bandwidth and no ingress/egress charges between the Outpost and your on-prem equipment"_. Which doesn't help with every use case, but helps with many.
This seems like it's available in surprisingly small increments, does anyone have pricing info? The public page easy to find has things for very different scales (at over 5k/mo to start).
You're probably on the Outposts rack section; you want to be on the Outposts servers section:
https://aws.amazon.com/outposts/servers/pricing/?nc=sn&loc=3
I was, yes, thank you!
I'm interested in this as well.
c6gd.16xlarge instance for example.
Even going through the order flow - nothing actually shows up for pricing which is a bit unusual for an AWS service. Must be somewhat manual on the backend still?
They price by the rack:
https://aws.amazon.com/outposts/rack/pricing/
Is my understanding correct that you may not overprovision an Outpost host like you would with a VMware host?