Introducing Hourly Billing
blog.linode.com
blog.linode.com
1) Would hourly billing attract a different kind of usage pattern than what Linodes have so far been used for?
2) Will this hurt performance for other customers?
The reason I'm curious about this is that even the cheapest Linode comes with a lot of CPU power, especially compared to EC2. But the CPU is not dedicated, in fact it's massively oversold. This is great for people who occasionally need to use 8 cores but stay relatively idle most of the time. Guess what, if you keep a server online for months, it's going to be idle a lof of the time.
Hourly billing, on the other hand, encourages people to fire up an instance, compute whatever they need to compute as quickly as possible, and shut down the instance as soon as they're done. In fact, that's exactly how I use EC2's high-CPU instances. Everything is scripted to ensure the most efficient use of resources throughout the lifetime of the instance, because every hour costs money. Why use 50% CPU for two hours when you can use 100% CPU for one hour for half the cost?
This kind of usage pattern, if widespread among customers, can lead to higher overall CPU usage on the host node (as well as of other scarce resources such as I/O), because fewer instances will sit idle. With EC2 this is not a problem AFAIK because the CPUs are dedicated to my instance. But Linode's CPUs are shared.
For a "traditional" Linode customer, the implication is that there will be busier neighbors. Officially, of course, you would never get less resources than a fair share of the host node. But it has usually been the case that, in Linode, you get a lot more resources most of the time. If people can no longer rely on this unofficial assumption, there could be trouble ahead for Linode.
This is doubly good news because Linode instances are well-geared towards CPU work (having 8 cores by default), so if you use a queue-based architecture for background processing, you can now spin up 10 instances to process a bunch of work, and take them down a few hours later when the rush is over.
I especially love the monthly cap, so people who just want a server don't pay unpredictable amounts every month.
Thanks, Linode.
I have a Postgres-based service on a Linode — it works fine in my (mainly read-based) use case, but I do pay for quite a lot of RAM in order to get enough caching.
Though, FWIW, still sticking with DO for the 5$/mo.
DO claims that the $10/month plan is their most popular... I express doubt, as most people I have heard from are on the $5 plan. From looking at their pricing, the $5/month plan looks like a significantly better deal, in terms of resources per dollar, assuming there is no overselling of disk involved, which is my current assumption.
(My motives for collecting this data are self-serving, but they are also transparent.)
Now, SSD is still slower than ram... way slower than ram. But it's way faster than spinning disk, especially on reads. I imagine there are some applications where SSD is fast enough that hitting disk is okay; you don't need to worry about having enough ram to cache all data you might read, and ram is still like eight or nine bucks a gig (vs. $0.6 to $2 for mlc ssd) - and if we're talking about reads here? MLC is fine for reads.
Of course, part of that ram difference might be that ram prices haven't fallen much over the last year or two, (in fact, most of the ram I bought last year, I bought at a lower price than the ram I'm looking at now.) while SSD prices have really fallen fast, so I don't know if this is the long-term price-equilibrium; still, I would expect ssd to continue to be significantly less expensive than ram.
Because of this, having enough ram to cache all the data you read often is extremely important to system performance when you are on spinning disk. (and having a pagecache that doesn't dump all the stuff you use often when you do a one-time read of a big file.)
I was suggesting that SSD might have fast enough read-access that for some applications, the penalty for reading from ssd rather than from pagecache might be small enough to not be a big deal, meaning you could get acceptable performance out of a system with less total ram.
The best answer sounds like it would be to decouple disk and ram allocations, use network storage and oversubscribe disk, as I imagine utilization on the disk of the $5 plans is low, but I would prefer to avoid doing that for the same reason that I believe Linode and DO avoids doing that. Network storage: reliable, fast, cheap. Pick two, if you choose really well.
From what I understand of my industry, and if nothing else, I've been in it for a while, new multi-user hardware, usually, is expected to be covered by subscription fees usually within 4-6 months, for this sort of thing. SSD is going to cost you between $0.6 (for cheap consumer-grade in raid5) and $2 (for intel S3500 level 'enterprise' mlc stuff in raid10) per gigabyte, so... it makes a difference.
ObLinode:
Of course, if you are using real live 'enterprise' slc, the price goes way up from there... but I doubt anyone in my sector is using those for anything but cache; in fact, I think Linode posted a while back that they were doing just that, using really good SSD for write-cache.
It's a good idea, really; once your server warms up, for most use cases, most of your reads are going to be cached in ram (assuming you have enough, which at today's prices, you should.) - If good ssd write-cache with something like bcache or dm-cache or your hw raid driver or what have you can get you good write speeds, that might be just as good as SSD in a real-world environment, assuming your ram cache is larger than the data you access often.
Of course, you are still going to get killed by the all-ssd host on disk benchmarks and on metrics like 'how long does it take to spin up a new guest?' - there are other ways to cheat on those metrics, but all that I know of have pretty big downsides for real-world performance.
I still think of Linode as more of a "real" type hosting environment, in the time I used them I didn't didn't notice any difference from actual bare metal servers. If I didn't know better that's what I would think they were.
DO, to me is much more of a "toy" and experimenting type of service. I have tried hosting things on "droplets" and had strange issues, dropping, outages etc. Not frequent of course, but over say a week or two of uptime. I'm ok with it because it's $5 and what else would you expect? But I wouldn't think of hosting any real sites on droplets personally.
It's just my opinion and many people I've talked to have had the same experiences. Both are great services for different applications.
They have a long history of hiding information from customers during security incidents and outages.
No one running a proper site would be crazy enough to host on Linode.
For example:
I took the time to search for five seconds because you wanted references, but were somehow incapable of making your own research, or making up your own mind.
If you search you will find repeated instances, and that means a long history, with no changes after even a single security problem.
My expectations are that they take that extra money and give me more assurances. Pricing is screwy and I fully admit this just may be pricing psychology.
For the extra amount I expect:
* Better support
* More stable network and disk I/O
* Feature decisions that favor stability over bleeding edge.
Every quarter I am looking at digital ocean and wondering if I need to move our production boxes over. So far though Linode has me hooked. The extra amount I save is not worth it. Digital ocean wins hard on price but they need more than that to buy a customer. Just like the competition between Azure and Amazon, Azure needs to do MORE than price match. Azure needs to go above and beyond what I expect from Amazon.
On the other hand, if there were good alternatives to OpenSSL the community probably would be talking about considering them right now.
As a Linode customer this makes me very happy, especially knowing it should hopefully garner the company a few more customers. A great company that deserves the success they get.
AWS does not do this. Stop an EC2 instance and you stop paying hourly instance fees. Even if it hasn't been terminated.
There are small fees for maintaining EBS volumes if you are using them for inactive instances, however.
You can improve your luck by going for a more expensive plan.
--(updated with links)--
On one of my 4096 boxes there:
> You will experience downtime while your Linode is migrated. We estimate 144 minutes to migrate your Linode, but that may vary based on host and network load.
It's amazing that the industry has come so far that it now seems normal to be able to spin up a server, have it available in a minute or so, then bring it down when done with it and only pay by the minute or hour.