Creating Your Own EC2 Spot Market
techblog.netflix.com
techblog.netflix.com
It would be a very nice feature if EC2 could incorporate each user's pool of reserved instances into their view of the spot market, but maybe that would create too much complexity on Amazon's side.
The spot pool is amazons unused capacity vs their allotment of RIs.
RIs always get priority capacity over spot, and spot instances always have highest bidder priority.
If there are a ton of spot capacity one hour, then Netflix scales up in that zone to launch jobs on their RIs - the RIs get launched and spot instances get whacked with the message "capacity over subscribed".
The RI is a guarantee that when you choose to launch that instance, capacity exists. Period. (Unless your instance limit for that type/zone/region is hit)
So, no Amazon could never do that.
However, AFAIK, Netflix can do what it wants with the RIs...
Think of Heroku: millions of sub-leases on their RIs to user apps.... But Amazon cannot meddle with the availability of the RIs that Heroku has purchased
The diff is that AWS spot instance pull from capacity that is unused, irrespective of RIs that you don't own... But when i come to launch my RI and capacity is full, your spot instance gets whacked.
What netflix is doing is an internal meta-spot-market on THEIR idle RIs. This compute capacity will never be available to anyone but them unless they create a service to sublet the capacity via the same tools they do to their internal teams... From amazons persective that capacity is already paid for regardless of consumption...
Forgive me if I am articulating this poorly...
In the end, I just find the numbers disclosed in here to be awesome and I suspect they are just but a fraction of total capacity Netflix requires.
Given that they disclose 12k idle trough RIs is very telling in efficiencies - but since I don't have the whole picture of their footprint, or instance demographic, I can't tell how much of this completes the picture of their landscape.
I'll assume 12k represents perhaps 80% of the machines they use... They have control-plan, orchestration, DR, private peering nodes (can't recall the name) but like backblaze pods, and any number of varied stage test debut envs...
So from a pure "stream or encode my shit" maybe 80% works, hopefully it's better - but could be worse...
Netflix's original mandate to push all DC ops and what not out of the building is not only super awesome, but is one of the linchpins in AWS business model. There should be whole career paths in college designed around this model...
Aws has resources. They are priced at X.
If you buy on demand, you pay X
If you buy an RI, you pay X-[Ri discount]
If you launch an instance, and you have an RI that applies, you get the discount.
Spot instances are a diff game. Spot and RI are mutually exclusive when costing.
Spot is only affected by capacity when it comes to RIs.
If you have RIs, you can't swap a running instance between spot or RI or OD pricing...
Aws calls this "lifecycle" - spot or normal. RI discount "coupons" only apply to an instance in the same zone on normal lifecycle that can receive the available RI pricing discount.
They do not overlap...
First - I've been seeking insight to how Netflix handles their footprint and what instance types they use. I dealt with these problems on a MUCH smaller scale (1K machines) but the challenges are the same...
Here is what I can derive from this info:
So given just 1500 idle Personalization R3.4xlarge instance for a given month -- if they did not use them, they were losing $265,140 per month on these instances (assuming a 36-month RI on these instances...) If they had these as no-upfront RIs -- then they are wasting $475,200 per month on these instances.
Obviously Netflix gets significant discounts on their RIs -- as far as I am aware the common customer will only get a discount of up to 5% off RIs after your initial $500K RI buy... im sure netflix has even better...
Anyway - as we dont know what the other instance types are that they use, if we were only to assume that the other 12K instances were the same and their utilization was only 50% of the day -- they would be wasting as little as $2MM/mo to as much as $3.8MM per month...
Finally - its great to see them mention specifically that the granularity of AWS is only hourly for instances - but their internal granularity is to the minute. The hourly granularity of AWS has been a thorn for some time. I certainly hope that Netflix will open source their spot manager at some point (Recall that AWS employees went out and started ClusterK with their spot manager, the Balancer, and then AWS came along and bought that little company up)....
There is significant value in being able to manage the spot market (your meta-version, like netflix, and AWS)....
What I would REALLY find interesting is if the Netflix Spot Manager were able to split and combine RIs and move them from zone to zone.
*Disclaimer: the napkin math above obviously does not take into account that the RIs and instances are in all regions - so the pricing will be much different and by region... but it does illustrate, even by those 1500 RIs how much money can be wasted on under-utilized instances.
Netflix could start a service offering batch jobs to other companies to utilize capacity if they wanted. 12K instances unused for even an hour is going to be say, $6,000/hour that could potentially be wasted...
I looked into it and they do billing by the minute with a minimum of 10 minutes billed.
Darn!