Allow auto-scale for IOPS.
Allow auto-scale for IOPS.
Babak
shared this idea
-
Ashutosh Sharma
commented
We also need this feature to be implemented badly. At this time, the way auto-scaling is configured for Atlas it only takes into account the CPU and memory utilization. My autoscaling is from M30 to M60 but my main parameter is I/O wait as we have some queries which are more like aggregate and they result into sever I/O threads.
Would like Auto Scale for IOPs.
-
William
commented
CSM opening a feature request for additional autoscale configuration on behalf of a customer.
The customer requests that they would like to have is the ability to autoscale Provisioned IOPS (once it is enabled) by setting a minimum, target and maximum Provisioned IOPS value for a given Atlas cluster.
-
Hear you there: we are working on designing a very different higher level of abstraction offering that aims to deliver to what you want here. This is going to take time but definitely appreciate the feedback.
-Andrew
-
Andrew
commented
Agreed.
The main issue with choosing a fixed value for Provisioned IOPS is you end up paying for IOPS you don't need 90% of the time.
For example, if our service typically requires 500 IOPS and occasionally spikes to 2000 IOPS, we have no choice but to pay for 2000 IOPS all month, even though 90% of the time we only need 500 IOPS.
I have to choose between paying for IOPS I don't need, or downtime. Both of these options suck.
Instead, you should not limit IOPS at all (never cause downtime), and count how many I use per month, and charge me per million. This is a much fairer pay-by-usage model.
-
Hi Babak,
One of the challenges is that it's significantly more expensive to opt into Provisioned IOPS even when not using too many of them -- Can you confirm whether you would be comfortable using Provisioned IOPS all the time (there is a significant premium cost to doing so) in order to leverage auto-scaling for IOPS?
Further, what kinds of indicators for your workload would you see as canonical drivers of needing to scale IOPS? We want to learn from you since this is nuanced and difficult to get right.
-Andrew
