Ksplice launches rebootless Linux update service - $4/server/mo
ksplice.com
ksplice.com
I'm happy to see someone making money off free software and I hope there'll be more of it. There just isn't much software one can spend money on besides the basic like RHEL. If I want to plunk down $5000 on high-performance something it seems like commercial lisp vendors are the only ones that can give that to you. I'd like to give $5000 for a faster Python.
The pessimist in me says to wait until a botched patch causes systems worldwide to crash and burn, though. It would also be interesting to find out more about Ksplice's own security as they are now in a position to create the mother of all DoS attacks.
I'm not following you here. Why does reboot-less updating of the kernel mean that the patches won't be tested. Also what DoS attack were you thinking of?
Lastly, clean up memory? The top uptimes on Netcraft are >1000 days, is this necessary, are there specific memory leaks you're considering?
I'm sure the people at Ksplice are much smarter than I am, but every time a security update for the kernel is released they will be stressing to release a Ksplice update to be able to provide value for their customers (the whole point of a subscription is so that you get the security update faster than planning a reboot takes).
The way Ksplice works is that it, for each update, injects all changed functions into the kernel and replaces the old functions with a jump to the new version of the same function. Even though it might not cause what you would consider a production-threatening memory leak, I would still like to clear out the old code from memory every once in a while.
So, the reboot is required because the patches cause the old code to remain but simply be long-jumped past to the injected new version of that block of instructions. Got it.
I'm not sure I'm in with you on Ksplice having to be really fast. The benefit is virtually no downtime but people are still going to test the patches on their testing servers and thus will still be testing in the wild before implementing fully.
I don't see a huge value to IaaS customers as their infrastructures are mostly ephemeral anyway, and are serviced by the likes of RightScale and Opscode who both provide tools that help you get your cloud image booted and configured.
As for the larger own-their-datacenter customers, they'll likely have some solution based on Chef (Opscode) or Puppet, and be more tolerant to reboots.
Smaller data center users might get a kick out of it though.
If you're large enough for the downtime associated with a reboot to be a big issue then you've presumably already got a minimum of two machines performing the same function. That being the case, it's no problem to remove each node from the cluster in turn, upgrade the kernel, reboot, test and add it back in.
That rather makes the service more of a convenience. For example, for active/passive setups where a failover produces an undesirable amount of downtime (say 30 seconds whilst services migrate), or single server setups where uptime isn't critical, but downtime is still a pain.
Cool technology, but perhaps an example of a solution looking for a problem. I think at $4/month the convenience aspect can probably justify the cost in many cases.