> While I agree that Google's pricing model is superior, the author's position on reserved instances accounts for ~40% of the cost difference.
Author here :D
Gotta take a stance to draw graphs, right? Comparing no reserved instance with the automatic discount is simple and realistic in my experience.
I could make pricing graphs accounting for 25% of AWS reservations and 75% of automatic google discount. In my experience, that's the proportions we _may_ have in practise. It would still be the same conclusion. Google is massively cheaper.
I could make TCO pricing graphs, accounting for the thing from the last paragraph, plus the human time (500$/day) required to understand, pick and manage reservations. I didn't do all the computations but so far it points towards a negative ROI (compared to not reserving), unless the infrastructure is big enough to have the economy of scale (hundreds thousands dollar per year). (Note: I really don't want to go against all the HN/reddit crowd who drank the reservations marketing koolaid, so purposefully avoid to talk about the TCO pricing model publicly ^^).
Thus, in all scenarios, google instances are massively cheaper. YMMV.
As a real life user of AWS I agree with this decision. RIs are not only a headache from a time perspective but we pay a non-trivial amount for a separate tool to manage purchasing them since AWS's built in tools aren't great.
Even with this tool we frequently make planning mistakes due to shifting business priorities or other reasons. Like - oh hay we have to upgrade our database instance but that needs a newer version because only the newer versions support something we need (i.e more than X GB ram). So engineering needs to modify the app to support the newer version.(Oh by the way AWS has a boatload of subtle restrictions like that.)
In this case we were sitting on a RI till EOY and for some reason you can't resell RDS RIs like normal ones (another subtle restriction) so we couldn't even partially recoup costs. Sometimes if you haggle with AWS Enterprise Support you can get them to help out if you're spending the money elsewhere but it's a pain. Good luck if you don't pay for that though (which itself is 15k a year or more).
No one doing over $100k per month is matching utilization perfectly with RIs. Either you are not utilizing some percent of RIs in which case you are overpaying through breakage, or else you are paying on-demand rates for the instances which are not under RI.
The calculus involved with matching RIs to forecSted demand in a dynamic growing workload is NP complete. And any inefficiency delta pays AWS. There's a small part of the curve where your savings is minimal if you don't match perfectly, it it easily goes to negative on the RI change.
Wait, $500/day to manage RIs? As someone who manages RIs for my org, that's way overblown. At most it'd be four figures per year. And once you get big enough, you leverage tools to help you figure out what RIs to get, which more than pay for themselves.
Author here :D
Gotta take a stance to draw graphs, right? Comparing no reserved instance with the automatic discount is simple and realistic in my experience.
I could make pricing graphs accounting for 25% of AWS reservations and 75% of automatic google discount. In my experience, that's the proportions we _may_ have in practise. It would still be the same conclusion. Google is massively cheaper.
I could make TCO pricing graphs, accounting for the thing from the last paragraph, plus the human time (500$/day) required to understand, pick and manage reservations. I didn't do all the computations but so far it points towards a negative ROI (compared to not reserving), unless the infrastructure is big enough to have the economy of scale (hundreds thousands dollar per year). (Note: I really don't want to go against all the HN/reddit crowd who drank the reservations marketing koolaid, so purposefully avoid to talk about the TCO pricing model publicly ^^).
Thus, in all scenarios, google instances are massively cheaper. YMMV.